PRoNTo Manual Prt

prt_manual

User Manual:

Open the PDF directly: View PDF .
Page Count: 208

Download
Open PDF In Browser	View PDF

PRoNTo Manual
The PRoNTo Development Group
(and honorary members)

John Ashburner
Carlton Chu
Andre Marquand
Janaina Mourao-Miranda
Joao M. Monteiro
Christophe Phillips
Jonas Richiardi
Jane Rondina
Maria J. Rosa
Jessica Schrouff
Tong Wu

Machine Learning & Neuroimaging Laboratory
Centre for Computational Statistics and Machine Learning
Computer Science department, UCL
Malet Place, London WC1E 6BT, UK
March 26, 2018
http://www.mlnl.cs.ucl.ac.uk/pronto

2

Contents
1 Introduction
1.1
Background . . . . . . . . . . . . . .
1.2
Methods . . . . . . . . . . . . . . . .
1.2.1
Inputs and preprocessing . .
1.2.2
Machine learning algorithms
1.3
Installing & launching the toolbox .
1.3.1
Installation . . . . . . . . . .
1.3.2
Launching and batching . . .
1.3.3
Troubleshooting . . . . . . .
1.4
What’s new? . . . . . . . . . . . . .
1.4.1
Version 2.1 . . . . . . . . . .
1.4.2
Version 2.0 . . . . . . . . . .
1.4.3
Version 1.1 . . . . . . . . . .
1.4.4
Version 1.0 . . . . . . . . . .
1.4.5
How to cite . . . . . . . . . .
1.5
Main contributors . . . . . . . . . . .
1.6
Acknoweldgements . . . . . . . . . .

I

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

Description of PRoNTo tools

2 Data & Design
2.1
Introduction . . . . . . . . . . .
2.2
Methods . . . . . . . . . . . . .
2.2.1
Data and design input .
2.2.2
Data and design output
2.2.3
Review . . . . . . . . .
2.2.4
HRF correction . . . . .
2.3
Graphical User interface . . . .
2.3.1
PRT directory . . . . .
2.3.2
Groups . . . . . . . . .
2.3.3
Subjects . . . . . . . . .
2.3.4
Modalities . . . . . . . .
2.3.5
Masks . . . . . . . . . .
2.3.6
Review . . . . . . . . .
2.3.7
Load, Save and Quit . .
2.4
matlabbatch interface . . . . .
3 Prepare feature set
3.1
Introduction . . . . . . . .
3.2
Methods and resources . .
3.3
Graphical User interfaces .
3.4
matlabbatch interface . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

9
9
10
10
11
11
12
12
12
14
14
14
15
15
15
15
17

19
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
3

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

21
21
22
22
22
22
22
24
24
24
25
25
28
29
29
30

.
.
.
.

33
33
33
35
37

4

CONTENTS

4 Model Specification and Estimation
4.1
Introduction . . . . . . . . . . . . . . . . . .
4.2
Beginning a model specification . . . . . . .
4.3
Feature set . . . . . . . . . . . . . . . . . .
4.4
Model type / pattern recognition algorithm
4.4.1
Classification . . . . . . . . . . . . .
4.4.2
Regression . . . . . . . . . . . . . . .
4.4.3
Hyper-parameter optimization . . .
4.5
Cross-validation . . . . . . . . . . . . . . . .
4.6
Specify / Run model . . . . . . . . . . . . .
4.7
Batch interface . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

39
39
39
40
40
41
42
42
43
46
47

5 Computing Feature and Region Contributions
5.1
Introduction . . . . . . . . . . . . . . . . . . .
5.2
Methods . . . . . . . . . . . . . . . . . . . . .
5.3
Graphical user interface . . . . . . . . . . . .
5.4
matlabbatch interface . . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

49
49
49
50
50

6 Display Model Performance
6.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . .
6.2
Launching results display . . . . . . . . . . . . . . . . .
6.3
The main results display window . . . . . . . . . . . . .
6.4
Analysing a machine’s performance graphically . . . . .
6.4.1
Classification . . . . . . . . . . . . . . . . . . . .
6.4.2
Regression . . . . . . . . . . . . . . . . . . . . . .
6.4.3
Influence of the hyper-parameter on performance
6.5
Statistical analysis of a machine’s performance . . . . .
6.5.1
Classification . . . . . . . . . . . . . . . . . . . .
6.5.2
Regression . . . . . . . . . . . . . . . . . . . . . .
6.5.3
Permutation testing . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

53
53
53
54
54
55
56
57
58
59
59
61

7 Display voxel and region contribution
7.1
Introduction . . . . . . . . . . . . . .
7.2
Displaying weights . . . . . . . . . .
7.2.1
Select image to display . . . .
7.2.2
Weights map . . . . . . . . .
7.2.3
Anatomical image . . . . . .
7.2.4
Additional plots . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

63
63
64
64
65
65
66

II

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

Batch interfaces

8 Data & Design
8.1
Directory . . . . . .
8.2
Groups . . . . . . . .
8.2.1
Group . . . .
8.3
Masks . . . . . . . .
8.3.1
Modality . .
8.4
fMRI Des . . . . . .
8.4.1
HRF overlap
8.4.2
HRF delay .
8.5
Review . . . . . . . .

69
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

71
71
71
71
73
73
74
74
74
74

CONTENTS

9 Feature set/Kernel
9.1
Load PRT.mat . . . . . . .
9.2
Feature/kernel name . . . .
9.3
Modalities . . . . . . . . . .
9.3.1
Modality . . . . . .
9.4
Use one kernel per modality

5

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

10 Specify model
10.1 Load PRT.mat . . . . . . . . . . . . . .
10.2 Model name . . . . . . . . . . . . . . . .
10.3 Use kernels . . . . . . . . . . . . . . . .
10.4 Feature sets . . . . . . . . . . . . . . . .
10.5 Model Type . . . . . . . . . . . . . . .
10.5.1 Classification . . . . . . . . . . .
10.5.2 Regression . . . . . . . . . . . . .
10.6 Cross-validation type . . . . . . . . . . .
10.6.1 Leave one subject out . . . . . .
10.6.2 k-folds CV on subjects . . . . . .
10.6.3 Leave one subject per group out
10.6.4 k-folds CV on subjects per group
10.6.5 Leave one block out . . . . . . .
10.6.6 k-folds CV on blocks . . . . . . .
10.6.7 Leave one run/session out . . . .
10.6.8 Custom . . . . . . . . . . . . . .
10.7 Include all scans . . . . . . . . . . . . .
10.8 Data operations . . . . . . . . . . . . . .
10.8.1 Mean centre features . . . . . . .
10.8.2 Other Operations . . . . . . . . .
11 Run
11.1
11.2
11.3

III

model
Load PRT.mat . . . . . . .
Model name . . . . . . . . .
Do permutation test? . . . .
11.3.1 No permutation test
11.3.2 Permutation test . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

75
75
75
75
75
76

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

77
77
77
77
77
77
77
79
81
81
81
81
81
81
81
82
82
82
82
82
82

.
.
.
.
.

83
83
83
83
83
83

Data processing examples

12 Block design fMRI dataset
12.1 GUI analysis . . . . . . . . . . . . . . .
12.1.1 Data & Design . . . . . . . . . .
12.1.2 Prepare feature set . . . . . . . .
12.1.3 Specify model . . . . . . . . . . .
12.1.4 Display model (optional step) . .
12.1.5 Compute weights (optional step)
12.1.6 Display results . . . . . . . . . .
12.1.7 Display weights . . . . . . . . . .
12.2 Batch analysis . . . . . . . . . . . . . . .
12.2.1 Data & Design . . . . . . . . . .
12.2.2 Feature set / Kernel . . . . . . .
12.2.3 Specify model . . . . . . . . . . .
12.2.4 Run model . . . . . . . . . . . .
12.2.5 Compute weights (optional step)

85
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

87
87
88
90
92
94
95
96
96
97
97
101
103
106
106

6

CONTENTS

13 Regression dataset
13.1 GUI analysis . . . . . . . . . . . . .
13.1.1 Data & Design . . . . . . . .
13.1.2 Prepare feature set . . . . . .
13.1.3 Specify model . . . . . . . . .
13.1.4 Display results . . . . . . . .
13.2 Batch analysis . . . . . . . . . . . . .
13.2.1 Data & Design . . . . . . . .
13.2.2 Feature set/Kernel . . . . . .
13.2.3 Specify model (KRR) . . . .
13.2.4 Run model (KRR) . . . . . .
13.2.5 Specify and Run model (RVR

. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
and GPR)

14 Multiple Kernel Learning Example
14.1 GUI analysis . . . . . . . . . . . . . . .
14.1.1 Data & Design . . . . . . . . . .
14.1.2 Prepare feature set . . . . . . . .
14.1.3 Specify model . . . . . . . . . . .
14.1.4 Display model (optional step) . .
14.1.5 Compute weights (optional step)
14.1.6 Display results . . . . . . . . . .
14.1.7 Display weights . . . . . . . . . .
14.2 Batch analysis . . . . . . . . . . . . . . .
14.2.1 Data & Design . . . . . . . . . .
14.2.2 Feature set / Kernel . . . . . . .
14.2.3 Specify model . . . . . . . . . . .
14.2.4 Run model . . . . . . . . . . . .
14.2.5 Compute weights (optional step)
15 Removing confounds: a classification
15.1 Introduction . . . . . . . . . . . . .
15.2 GUI analysis . . . . . . . . . . . .
15.2.1 Data & Design . . . . . . .
15.2.2 Prepare feature set . . . . .
15.2.3 Specify model . . . . . . . .
15.2.4 Display results . . . . . . .
15.3 Batch analysis . . . . . . . . . . . .
15.3.1 Data & Design . . . . . . .
15.3.2 Feature set / Kernel . . . .
15.3.3 Specify model . . . . . . . .
15.3.4 Run model . . . . . . . . .
15.3.5 Display weights . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

109
109
109
111
111
113
113
113
114
115
116
116

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

119
120
120
121
122
123
124
124
125
125
126
129
131
132
134

example
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

135
135
136
136
138
138
139
140
140
141
142
143
144

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

145
145
145
145
146
147
148
148
148
149
152
153
153

.
.
.
.
.
.
.
.
.
.
.
.
.
.

16 Removing confounds: a regression example
16.1 Introduction . . . . . . . . . . . . . . . . .
16.2 GUI analysis . . . . . . . . . . . . . . . .
16.2.1 Data & Design . . . . . . . . . . .
16.2.2 Prepare feature set . . . . . . . . .
16.2.3 Specify model . . . . . . . . . . . .
16.2.4 Display results . . . . . . . . . . .
16.2.5 Compute weights (optional step) .
16.3 Batch analysis . . . . . . . . . . . . . . . .
16.3.1 Data & Design . . . . . . . . . . .
16.3.2 Feature set / Kernel . . . . . . . .
16.3.3 Specify model . . . . . . . . . . . .
16.3.4 Run model . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

CONTENTS

IV

7

Advanced topics

17 Developer’s guide
17.1 Introduction . . . . . . .
17.2 Code organisation . . . .
17.2.1 User interface . .
17.2.2 Machine learning
17.2.3 Machines . . . .

155
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

18 PRT structure
19 List of PRoNTo functions
19.1 pronto.m . . . . . . . . . . . . . .
19.2 prt.m . . . . . . . . . . . . . . . .
19.3 prt apply operation.m . . . . . .
19.4 prt batch.m . . . . . . . . . . . .
19.5 prt build region weights.m . . . .
19.6 prt check.m . . . . . . . . . . . .
19.7 prt check design.m . . . . . . . .
19.8 prt check flag.m . . . . . . . . . .
19.9 prt comp ranking dist.m . . . . .
19.10 prt compute cv mat.m . . . . . .
19.11 prt compute weights.m . . . . . .
19.12 prt compute weights class.m . . .
19.13 prt compute weights regre.m . .
19.14 prt cv fold.m . . . . . . . . . . .
19.15 prt cv model.m . . . . . . . . . .
19.16 prt cv opt param.m . . . . . . .
19.17 prt data conditions.m . . . . . .
19.18 prt data modality.m . . . . . . .
19.19 prt data review.m . . . . . . . . .
19.20 prt defaults.m . . . . . . . . . . .
19.21 prt fs.m . . . . . . . . . . . . . .
19.22 prt fs modality.m . . . . . . . . .
19.23 prt func2html.m . . . . . . . . .
19.24 prt getKernelModel.m . . . . . .
19.25 prt get defaults.m . . . . . . . . .
19.26 prt get filename.m . . . . . . . .
19.27 prt init fs.m . . . . . . . . . . . .
19.28 prt init model.m . . . . . . . . .
19.29 prt latex.m . . . . . . . . . . . .
19.30 prt load.m . . . . . . . . . . . . .
19.31 prt load blocks.m . . . . . . . . .
19.32 prt model.m . . . . . . . . . . . .
19.33 prt nested cv.m . . . . . . . . . .
19.34 prt permutation.m . . . . . . . .
19.35 prt plot ROC.m . . . . . . . . . .
19.36 prt plot confusion matrix.m . . .
19.37 prt plot histograms.m . . . . . .
19.38 prt plot nested cv.m . . . . . . .
19.39 prt plot prediction.m . . . . . . .
19.40 prt plot prediction reg bar.m . .
19.41 prt plot prediction reg line.m . .
19.42 prt plot prediction reg scatter.m
19.43 prt region histogram.m . . . . . .
19.44 prt remove confounds.m . . . . .
19.45 prt stats.m . . . . . . . . . . . .

157
157
157
158
158
158
161

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

165
167
167
167
168
168
168
169
169
170
170
170
170
171
171
172
172
173
173
173
174
174
175
175
176
176
176
177
178
178
179
179
179
180
180
181
181
181
182
182
183
183
183
183
183
184

8

CONTENTS

19.46
19.47
19.48
19.49
19.50
19.51
19.52
19.53
19.54
19.55
19.56
19.57
19.58
19.59
19.60
19.61
19.62
19.63
19.64
19.65
19.66
19.67
19.68
19.69
19.70

prt struct.m . . . . . . . . . . . . . . . . . . .
prt struct2latex.m . . . . . . . . . . . . . . .
prt text input.m . . . . . . . . . . . . . . . .
prt ui compute weights.m . . . . . . . . . . .
prt ui custom CV.m . . . . . . . . . . . . . .
prt ui cv model.m . . . . . . . . . . . . . . .
prt ui design.m . . . . . . . . . . . . . . . . .
prt ui disp weights.m . . . . . . . . . . . . . .
prt ui kernel construction.m . . . . . . . . . .
prt ui main.m . . . . . . . . . . . . . . . . . .
prt ui model.m . . . . . . . . . . . . . . . . .
prt ui prepare data.m . . . . . . . . . . . . .
prt ui prepare datamod.m . . . . . . . . . . .
prt ui results.m . . . . . . . . . . . . . . . . .
prt ui results ROI.m . . . . . . . . . . . . . .
prt ui results help.m . . . . . . . . . . . . . .
prt ui results stats.m . . . . . . . . . . . . . .
prt ui reviewCV.m . . . . . . . . . . . . . . .
prt ui reviewmodel.m . . . . . . . . . . . . . .
prt ui select class.m . . . . . . . . . . . . . .
prt ui select reg.m . . . . . . . . . . . . . . .
prt ui specify CV basis.m . . . . . . . . . . .
prt ui stats.m . . . . . . . . . . . . . . . . . .
prt ui sure.m . . . . . . . . . . . . . . . . . .
machines . . . . . . . . . . . . . . . . . . . . .
19.70.1 machines\prt KRR.m . . . . . . . . .
19.70.2 machines\prt machine.m . . . . . . .
19.70.3 machines\prt machine RT bin.m . . .
19.70.4 machines\prt machine gpclap.m . . .
19.70.5 machines\prt machine gpml.m . . . .
19.70.6 machines\prt machine gpr.m . . . . .
19.70.7 machines\prt machine krr.m . . . . .
19.70.8 machines\prt machine rvr.m . . . . .
19.70.9 machines\prt machine sMKL cla.m . .
19.70.10 machines\prt machine sMKL reg.m .
19.70.11 machines\prt machine svm bin.m . . .
19.70.12 machines\prt rvr.m . . . . . . . . . .
19.70.13 machines\prt weights.m . . . . . . . .
19.70.14 machines\prt weights bin linkernel.m
19.70.15 machines\prt weights gpclap.m . . . .
19.70.16 machines\prt weights sMKL cla.m . .
19.70.17 machines\prt weights sMKL reg.m . .
19.70.18 machines\prt weights svm bin.m . . .
19.71 utils . . . . . . . . . . . . . . . . . . . . . . .
19.71.1 utils\prt centre kernel.m . . . . . . . .
19.71.2 utils\prt checkAlphaNumUnder.m . .
19.71.3 utils\prt normalise kernel.m . . . . . .
19.71.4 utils\prt utils update mask.m . . . . .

V

Bibliography

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

184
184
184
185
185
186
186
187
187
188
188
189
189
190
190
190
191
191
192
192
193
193
193
194
194
194
194
195
195
196
197
198
198
199
199
200
200
201
202
202
202
202
203
203
203
203
203
203

205

Chapter 1

Introduction
Contents
1.1
1.2

1.3

1.4

1.5
1.6

1.1

Background . . . . . . . . . . . . .
Methods . . . . . . . . . . . . . . .
1.2.1
Inputs and preprocessing . .
1.2.2
Machine learning algorithms
Installing & launching the toolbox .
1.3.1
Installation . . . . . . . . . .
1.3.2
Launching and batching . .
1.3.3
Troubleshooting . . . . . . .
What’s new? . . . . . . . . . . . . .
1.4.1
Version 2.1 . . . . . . . . . .
1.4.2
Version 2.0 . . . . . . . . . .
1.4.3
Version 1.1 . . . . . . . . . .
1.4.4
Version 1.0 . . . . . . . . . .
1.4.5
How to cite . . . . . . . . . .
Main contributors . . . . . . . . . .
Acknoweldgements . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

9
10
10
11
11
12
12
12
14
14
14
15
15
15
15
17

Background

Advances in neuroimaging techniques have radically changed the way neuroscientists address
questions about functional anatomy, especially in relation to behavioural and clinical disorders.
Many questions about brain function, previously investigated using intracranial electrophysiological recordings in animals can now be addressed non-invasively in humans. Such studies have
yielded important results in cognitive neuroscience and neuropsychology. Amongst the various
neuroimaging modalities available, Magnetic Resonance Imaging (MRI) has become widely used
due to its relatively high spatial and temporal resolution, and because it is safe and non-invasive.
By selecting specific MRI sequence parameters, different MR signals can be obtained from different tissue types, giving images with high contrast among organs, between normal and abnormal
tissues and/or between activated and deactivated brain areas. MRI is often sub-categorized into
structural MRI (MRI) and functional MRI (fMRI). Examples of other of imaging modalities
that measure brain signals are Positron Emission Tomography (PET), ElectroEncephaloGraphy
(EEG) recordings and MagnetoEncephaloGraphy (MEG) recordings. Neuroimaging data are inherently multivariate, since each measure (scan or recording) contains information from thousands
of locations (e.g. voxels in MRI or electrodes in EEG). Considering that most brain functions
are distributed processes involving a network of brain regions, it would seem desirable to use the
spatially distributed information contained in the data to give a better understanding of brain
functions in normal and abnormal conditions.
9

10

CHAPTER 1. INTRODUCTION

The typical analysis pipeline in neuroimaging is strongly rooted in a mass-univariate statistical
approach, which assumes that activity in one brain region occurs independently from activity in
other regions. Although this has yielded great insights over the years, specially in terms of
function localization, and continues to be the tool of choice for data analysis, there is a growing
recognition that the spatial dependencies among signal from different brain regions should be
properly modelled. The effect of interest can be subtle and spatially distributed over the brain
- a case of high-dimensional, multivariate data modelling for which conventional tools may lack
sensitivity.
Therefore, there has been an increasing interest in investigating this spatially distributed
information using multivariate pattern recognition approaches, often referred to as multi-voxel
pattern analysis (MVPA) (see [12], [7] and [13]). Where pattern recognition has been used in
neuroimaging, it has led to fundamental advances in the understanding of how the brain represents
information and has been applied to many diagnostic applications. For the latter, this approach
can be used to predict the status of the patient scanned (healthy vs. diseased or disease A vs.
B) and can provide the discriminating pattern leading to this classification. Pattern recognition
techniques can also be used to identify relationships between patterns of brain structure or activity
and continuous measures such as age or a clinical score. Such information can then be used to
predict individual-level measures for new individuals (i.e. regression models).
Several active areas of research in machine learning are crucially important for the difficult
problem of neuroimaging data analysis: modelling of high-dimensional multivariate time series,
sparsity, regularisation, dimensionality reduction, causal modelling, and ensembling to name a
few. However, the application of pattern recognition approaches to the analysis of neuroimaging
data is limited mainly by the lack of user-friendly and comprehensive tools available to the
fundamental, cognitive, and clinical neuroscience communities. Furthermore, it is not uncommon
for these methods to be used incorrectly, with the most typical case being improper separation
of training and testing datasets.

1.2

Methods

PRoNTo (Pattern Recognition for Neuroimaging Toolbox) is a toolbox based on pattern recognition techniques for the analysis of neuroimaging data. Statistical pattern recognition is a field
within the area of machine learning which is concerned with automatic discovery of regularities in
data through the use of computer algorithms, and with the use of these regularities to take actions
such as classifying the data into different categories [3]. In PRoNTo, brain images are treated
as spatial patterns and statistical learning models are used to identify statistical properties of
the data that can be used to discriminate between experimental conditions or groups of subjects
(classification models) or to predict a continuous measure (regression models).
PRoNTo is Matlabbased and includes five main modules: ‘Data & Design’, ‘Prepare feature
set’, ‘Specify model’, ‘Run model’ and ‘Compute weights’. The results can displayed in terms of
the performance of the estimated model, as well as in terms of model parameters. Additional
review options enable the user to review information about the data, features and models. All
modules were implemented using a graphical user interface (GUI) and the MATLAB Batch System. Using the MATLAB Batch System the user can run each module as batch jobs, which
enables a very efficient analysis framework. All information about the data, experimental design,
models and results are saved in a structure called PRT. PRoNTo also creates additional files
during the analysis that are described in details in the next chapters.
The toolbox code will be distributed for free, but as copyright software under the terms of
the GNU General Public License as published by the Free Software Foundation.

1.2.1

Inputs and preprocessing

In terms of neuroimaging modalities, PRoNTo accepts NIFTI files. Mostly designed to analyse structural and functional Magnetic Resonance Imaging and PET, it can be used on any
dataset converted to a NIFTI file. It assumes that the neuroimaging data has been previously
pre-processed using SPM (http://www.fil.ion.ucl.ac.uk/spm/) or a similar software for neuroimaging analysis. In general, raw fMRI data should be previously corrected for movement

1.3. INSTALLING & LAUNCHING THE TOOLBOX

11

artefact (realigned) and time difference in slice acquisition (slice time correction), mapped to a
common template (normalized) and spatially smoothed. The normalisation and spatial smoothing
steps might not be necessary for single subject analysis.
In addition the General Linear Model (GLM) can be applied as a pre-processing step for
pattern recognition analysis. In this case, the GLM coefficients (e.g. beta images in SPM) will
correspond to the spatial patterns. Using beta images should be preferred in the case of eventrelated designs with short inter-stimulus time and/or event duration to better take into account
the Haemodynamic Response Function (HRF). Important note: Beta images output by SPM
contain NaNs (Not a Number) values in some voxels. For better performance of the model, a
mask should be created to exclude the corresponding voxels from the analysis. Practically, a
mask should be built (e.g. using SPM imcalc batch or a script provided in ‘utils’) to specify
which voxels have scalar values (1 in mask) or NaN or 0 value (0 in mask). The updated mask
should then be input in the Data & Design window. To ease this extra preprocessing step, a script
is provided in your PRoNTo folder/utils. It takes as inputs the images (i.e. beta images in nifti
format), the mask to update (e.g. SPMnoeyes.nii provided in your PRoNTo folder/mask ) and the
directory to save the updated mask. Inputs are optional, and just typing in the Matlabcommand:
prt utils update mask will ask for the different inputs using file and path selectors.
Raw structural MRI data should be previously mapped to a common template (normalized)
and spatially smoothed. Raw PET data should be realigned, normalized and smoothed. PET
data is also usually scaled. This operation can be performed before hand or during the building
of the feature set.

1.2.2

Machine learning algorithms

In PRoNTo different pattern recognition algorithms correspond to different machines. The machine library in PRoNTo v2 includes four classification models: Support Vector Machine ([4],
[11]), Gaussian Process Classifier (binary and multiclass, [15], [9]) and L1-Multiple Kernel Learning [14]. Four regression models are available: Kernel Ridge Regression [19], Relevance Vector
Regression [20], Gaussian Process Regression [15] and L1-Multiple Kernel regression [14]. New
machines will be added to the library in future versions of the toolbox.
PRoNTo should facilitate the interaction between machine learning and neuroimaging communities. On one hand the machine learning community should be able to contribute to the
toolbox with novel published machine learning models. On the other hand the toolbox should
provide a variety of tools for the neuroscience and clinical neuroscience communities, enabling
them to ask new questions that cannot be easily investigated using existing statistical analysis
tools.

1.3

Installing & launching the toolbox

In order to work properly, PRoNTo requires 2 other softwares:
• a recent version of Matlab. We used versions 7.5 (R2007b) to 8.3 (R2014a) to develop
PRoNTo, and it will not work with earlier versions1 .
• SPM8[10] installed on your computer2 . SPM12b and SPM12 are also supported.
PRoNTo latest public version can be downloaded, after registration, from the following address: http://www.mlnl.cs.ucl.ac.uk/pronto/prtsoftware.html.
Important note: if you already have a version of SPM on your computer, please download
the latest updates from the website. Various bugs, especially in terms of weight visualization,
arise from out of date SPM versions.
1 Any

later Matlab version should work, in theory.
can be dowloaded from the following website: http://www.fil.ion.ucl.ac.uk/spm/software/. You
should install it in a suitable directory, for example C:\SPM8\, then make sure that this directory is on the
Matlabpath. No need to include the subdirectories!
2 SPM8

12

CHAPTER 1. INTRODUCTION

1.3.1

Installation

After downloading the zipped file containing PRoNTo, the installation proceeds as follow:
1. Uncompress the zipped file in your favourite directory, for example C:\PRoNTo\;
2. Launch Matlab;
3. Go to the “File” menu → “Set path”;
4. Click on the “Add folder” button and select the PRoNTo folder, i.e. C:\PRoNTo\ if you
followed the example;
5. Click on save.
Some routines, in particular the ’machines’, are written in C++ (.cpp files) for increased
efficiency. We are trying to provide these compiled routines for the usual Operating Systems
(OS’s) such as: Windows XP (32 bits), Windows 7 (64 bits), Mac OS 10, Linux (32 and 64 bits).
If your OS is not listed or routines do not work properly then you should compile the routines
for your specific OS3 .

1.3.2

Launching and batching

Once installed, there are three ways to call up PRoNTo functionalities. To launch the toolbox
GUI, just type prt or pronto at the Matlab prompt and the main GUI figure will pop up, see
Fig. 1.1. From there on simply click on the processing step needed (see Part I of this manual).
Most functions of PRoNTo have been integrated into the matlabbatch batching system [5] (like
SPM8) and the batching GUI is launched from the main GUI by clicking on the Batch button
(see Part II of this manual). Of course most tools can also be called individually by calling them
directly from the Matlab prompt, or for scripting in a .m file (see Part IV of this manual).

Figure 1.1: Main GUI interface: each button launches a specific processing step.

1.3.3

Troubleshooting

Compiling libsvm
Some problems when using SVMs might arise due to libsvm, in which case, you might need
to compile it on your own. The first thing that needs to be done is to download the desired
3 you

can also contact us and we’ll try to come up with a solution for your system.

1.3. INSTALLING & LAUNCHING THE TOOLBOX

13

libsvm version (usually the latest one) from the following website: http://www.csie.ntu.edu.
tw/~cjlin/libsvm/. Then, the process will depend on your operating system.
If the steps described bellow do not work, please refer to the README file that comes with
libsvm.
Microsoft Windows
• Make sure you have a C++ compiler install. If not, you can install Microsoft Visual C/C++;
• Copy the libsvm folder to the ‘machines’ directory of your PRoNTo instalation
(e.g. C:\PRoNTo\machines\);
• Open a DOS command window and change to the libsvm folder in the previous step (cd
C:\PRoNTo\machines\libsvm-3.17\). If the environment variables of VC++ have not
been set, run the following command:
C:\Program Files\Microsoft Visual Studio 10.0\VC\bin\vcvars32.bat. This command might be different, depending on the path of your Visual Studio installation;
• In the libsvm folder run the command: nmake -f Makefile.win clean all
• If no errors appear, open MATLAB;
• Change to the ‘matlab’ folder inside the libsvm folder (e.g. C:\PRoNTo\machines\libsvm3.17 \matlab\);
• Run make in the MATLAB Command Window. If there are no errors, you have just
successfully compiled libsvm to be used with MATLAB.
Remember, if you want to use the version that you have just compiled, you have to add the
libsvm folder to your path in MATLAB. If you have more than one libsvm folder inside the
‘machines’ folder, please remove one of them from the MATLAB path. You should only have one
libsvm folder in your path.
Unix (Mac OS or Linux)
• Make sure you have a C++ compiler installed. If you are using Mac OS, please install
‘Xcode’. On Linux systems, you should already have ‘gcc’ installed;
• Copy the libsvm folder to the ‘machines’ directory of your PRoNTo instalation
(e.g. /home//PRoNTo/machines/);
• Open a terminal window and change to the ‘machines’ directory: cd PRoNTo/machines/
• Compile libsvm by running the following command: make
• If no errors appear, open MATLAB;
• Change to the ‘matlab’ folder inside the libsvm folder (e.g. PRoNTo/machines/libsvm3.17/matlab/);
• Run make in the MATLAB Command Window. If there are no errors, you have just
successfully compiled libsvm to be used with MATLAB.
Remember, if you want to use the version that you have just compiled, you have to add the
libsvm folder to your path in MATLAB. If you have more than one libsvm folder inside the
‘machines’ folder, please remove one of them from the MATLAB path. You should only have one
libsvm folder in your path.

14

CHAPTER 1. INTRODUCTION

1.4
1.4.1

What’s new?
Version 2.1

This version of the toolbox is released mainly to provide bug fixes for version v2.0. One feature
is also added: removing confounds in kernel space for predictions and in the original space for
weight computation. A detailed description can be find: http://www.cs.ucl.ac.uk/fileadmin/
UCL-CS/research/Research_Notes/RN_17_09_ARJMM.pdf.

1.4.2

Version 2.0

This version of the toolbox (2015) aims at providing multiple new functionalities, including:
• Build one kernel per modality: When there are multiple modalities in the same dataset,
it is now possible to build one kernel per modality and use them in the same model later on.
If this option is not chosen, the selected modalities can still be concatenated as additional
samples/examples or sessions (if they have the same dimensions) or used separately.
• Build one kernel per region: It is also possible to specify an atlas, comprising regions
of interest (ROIs) as defined by values in the atlas ranging from 1 to number of ROIs. For
each anatomically defined ROI, a kernel will be built taking into account the multivariate
pattern within the region. It is possible to use this feature in combination with the previous
one (e.g. one kernel per region and per modality, or concatenate modalities as additional
samples and build one kernel per region).
• New machines: The list of available machines was extended, and now includes L1-Multiple
Kernel Learning (MKL, classification and regression). The latter corresponds to a hierarchical model defined by weights at two levels: the voxel level and the kernel level (i.e. ROI
and/or modality).
• Flexible cross-validation: A GUI is provided to manually specify a custom cross-validation
matrix. It allows to either specify a basis (e.g. leave-one-subject-out), load a .mat or specify
the number of folds. Then the user can, for each fold, select which examples are part of the
training set, test set, or won’t be used. The resulting matrix can be saved for further use.
• Nested cross-validation for hyperparameter optimization: It is now possible to
optimize the hyperparameter(s) of some machines (e.g. the soft-margin parameter, C,
in SVM) using a nested cross-validation (CV) framework. The number of folds in the
nested cross-validation, used only to estimate the value of the hyperparameter leading to
the highest performance, does not need to be the same as the ‘outer’ cross-validation (i.e.
the one estimating the final model performance). For example, to decrease computational
expenses, the nested CV can be a 4-fold CV while the outer CV can be a leave-one-out.
• Display results: The display of the results was divided into two modules (Display Results
and Display Weights). This allows to review the model performance in one window, with
all the statistics. A new graph displaying the effect of the hyperparameter (if optimized)
was also included.
• Weights per ROI: For each model it is possible to build images representing the weights
per voxel and also images summarising the weights per regions of interest as defined by an
atlas. If an MKL model was built on ROIs, the contribution of each ROI (regional weight)
is explicitly derived. On the other hand, if a simple kernel model was selected (e.g. SVM
on the whole brain), the weights per voxel will be averaged (in absolute value) within each
region, as defined by an atlas specified by the user. In both cases, an additional image, with
weights per ROI, is created and saved.
• Display weights: The weights of each model can be displayed at the voxel level in this
window. If weights per ROI were derived (either summarized or from an MKL on ROIs),
weights per region can be displayed as an image, as well as in a sorted list of regions.
The same applies for MKL models on multiple modalities. A histogram is also displayed

1.5. MAIN CONTRIBUTORS

15

representing the contribution/weight of each ROI/modality to the model. The table can be
exported as text for future use in publications/communications.

1.4.3

Version 1.1

In 2012, PRoNTo v1.1 was released mainly to provide bug fixes for version v1.0. Two features
were also added:
• automatic compiling of the machines (in particular: no more issues with SVM, nor with
Matlab toolboxes and paths).
• k-folds Cross-Validation (CV): specify the number of folds or set it to 1 for half-half CV
(train on first half, test on second).

1.4.4

Version 1.0

Launched in 2011, this version of PRoNTo allows to perform all the analysis steps, from Data &
Design to computing the weights for three classification machines (SVM, binary and multi-class
GP) and two regression machines (KRR, SVR).

1.4.5

How to cite

Please cite [16] when using PRoNTo analyses in any type of publication. In addition, Multiple
Kernel Learning analyses should refer to a second paper (submitted, tba)4 . Meanwhile, please
cite [18] for MKL analyses and [17] for a posteriori weight summarization using atlas-defined
regions of interest.

1.5

Main contributors

PRoNTo is developed by the Machine Learning & Neuroimaging Laboratory, Computer Science
department, University College London, UK (http://www.mlnl.cs.ucl.ac.uk) and associated
researchers.
The main contributors, in alphabetical order, are:
Dr. John Ashburner is a Professor of Imaging Science at the Wellcome Trust Centre for Neuroimaging at the University College London Institute of Neurology. He is mainly interested
in modelling brain anatomy from MR scans, and more recently in applying pattern recognition methods to make predictions about individual subjects. He is a co-developer of the
SPM software (intra- and inter-subject registration, tissue classification, visualization and
image file formats), which is used internationally by thousands of neuroimaging researchers.
He has a Web of Science h-index of 63. He did not contribute any actual code to PRoNTo,
but he did attend many of the meetings;
Dr. Carlton Chu is a research scientist at Google DeepMind. Before joining DeepMind, he was
a research fellow in brain imaging at the National Institute of Mental Health (NIMH), NIH.
He received the B.Eng. degree (1st class Honours) from Auckland University, in 2002 and
the master of Biomedical Engineering from University of New South Wales, in 2004. Carlton
obtained a PhD in Neuroimaging method from University College London in 2009, working
in the statistical methods group at the Wellcome Trust Centre for Neuroimaging, creators of
the famous “SPM” program. There he developed innovative pattern recognition methods to
automatically detect the early stages of neurodegenerative diseases such as Alzheimer’s and
Huntingdon’s from structural brain images. In 2007, Carlton won the first prize in the 2nd
Pittsburgh Brain Activity Interpretation Competition (PBAIC), a prestigious international
competition involving the application of machine learning to the problem of classification
of brain activity. He led a small research team to victory, acclaim from peers in the field,
and the $10K first prize. His current research interests include image segmentation using
4 When the paper gets accepted, this section will be updated and the paper will be listed in the documentation
section of the website

16

CHAPTER 1. INTRODUCTION

convolutional neural networks and applications of deep-learning. Carlton was involved in
the development of PRoNTo version 1.0;
Dr. Andre Marquand is an Junior Principal Investigator at the Donders Institute for Brain
Cognition and Behaviour and a Lecturer at the Institute of Psychiatry, Psychology and
Neuroscience, King’s College London. His research focuses on the application of probabilistic
machine learning techniques to neuroimaging data, particularly for clinical applications. His
recent work includes the development of multi-class, multi-task and multi-modality pattern
classification methods that offer many advantages over current techniques including more
sensitive and specific detection of disease effects;
Dr. Janaina Mourao-Miranda is a Wellcome Trust Senior Research Fellow at the Centre
for Computational Statistics and Machine Learning (CSML), UCL. Over the past years
her research has involved developing and applying pattern recognition methods to analyze
neuroimaging data, in particular brain activation and structural patterns that distinguish
between controls and patients. Her current research focuses on developing machine-learning
models to investigate complex relationships between neuroimaging data and multidimensional descriptions of mental health disorders. She has been coordinating the PRoNTo
developments and contributed to PRoNTo versions 1.0, 1.1 and 2.0;
Mr. Joao M. Monteiro is a former MPhil/PhD Student at University College London under
the supervision of Prof. John Shawe-Taylor and Dr. Janaina Mourao-Miranda. His research
focuses on the application of unsupervised machine learning methods to neuroimaging. He
contributed to PRoNTo versions 1.1 and 2.0;
Dr. Christophe Phillips is FRS-FNRS Research Associate at the Cyclotron Research Centre
and adjunct Assistant Professor at the Department of Electrical Engineering and Computer
Science, University of Liège, Belgium. His research focuses on the processing of multi-modal
neuroimaging data. Recent work within the field of “brain decoding” aimed at distinguishing
between levels of consciousness in unresponsive patients or between typical and atypical
Parkinson Disease patients using Positron Emission Tomography (PET) imaging, as well
as tracking mnesic traces in trained healthy subjects with fMRI;
Dr. Jonas Richiardi is a Marie Curie researcher at the University of Geneva (Department
of Fundamental Neuroscience). His research interests include the combination of imaging
modalities with other biological information sources including genomic data, learning with
graphs, machine learning for neuroimaging, brain connectivity / resting-state data analysis,
interpretability of brain decoding results, and functional biomarkers. He has been involved
with the organization of the workshop on Pattern Recognition in NeuroImaging since the
beginning;
Dr. Jane Rondina is a research fellow at the University College London Institute of Neurology. Previously, she was a post-doctoral research associate at the Centre for Neuroimaging
Sciences, King’s College London. In the past years, her research has involved application
of pattern recognition methods to neuroimaging data and development of a stability-based
method for feature selection and mapping in neuroimaging. Her current research focuses
on prognosis and prediction of treatment response, mainly addressing approaches to combine complementary information from different imaging modalities and other sources of
data (clinical, demographic and genetic). She contributed to the development of PRoNTo
version 1.0. She is also involved in incorporation of new resources for version 3.0;
Dr. Maria J. Rosa is an imaging scientist at IXICO, plc. Before, she was a Post-Doctoral
Research Fellow at the Institute of Psychiatry, King’s College London (KCL) and a Wellcome Trust post doctoral research associate at the Centre for Computational Statistics and
Machine Learning (CSML), UCL. Maria’s main area of work is the development and application of machine learning and multivariate methods to neuroimaging data. She did her
PhD at the Wellcome Trust Centre for Neuroimaging, UCL. She contributed to PRoNTo
versions 1.0, 1.1 and 2.0;

1.6. ACKNOWELDGEMENTS

17

Dr. Jessica Schrouff got her PhD from the University of Liège, under the supervision of Dr.
C. Phillips. She is a post-doctoral researcher at the Laboratory of Behavioral and Cognitive
Neuroscience, Stanford University. Her research focuses on the detection and characterization of memory traces in resting state wakefulness using machine learning techniques,
based on fMRI data (Ph.D. thesis) and intracranial EEG recordings. She contributed to
PRoNTo versions 1.0, 1.1 and 2.0. She has also started adapting the toolbox for analysis of
electrophysiological recordings (version 3.0);
Dr. Tong Wu got her PhD from the University of Queensland in Australia, under supervision
of Dr. Tianzi Jiang, Dr. David Reutens and Dr. Simone Bosshard. She is a post-doctoral
researcher in Centre for Medical Image Computing (CMIC) at the University College London. Her research focus during PhD was resting-state mouse fMRI in both wild type mice
and a schzophrenia mouse model. She contributed to PRoNTo version 2.1.
We also thank students and post-docs for their help in testing the software and writing this
manual: Dr. Liana Lima Portugal, Liane Canas, Dr. Ana Regia Neves, Rochelle Silva, Dr.
Orlando Fernandes Junior and Dr. Qinquan Gao.

1.6

Acknoweldgements

PRoNTo is the deliverable of a Pascal Harvest project coordinated by Dr. J. Mourao-Miranda
and its development was possible with the financial and logistic support of
• the Department of Computer Science, University College London (http://www.cs.ucl.
ac.uk);
• the Wellcome Trust (http://www.wellcome.ac.uk/) under grants no. WT086565/Z/08/Z
and no. WT102845/Z/13/Z;
• PASCAL2 (http://www.pascal-network.org) and its HARVEST programme;
• the Fonds de la Recherche Scientifique-FNRS (http://www.fnrs.be), Belgium;
• Fundação para a Ciência e Tecnologia - FCT (http://www.fct.pt), Portugal;
• Swiss National Science Foundation (PP00P2-123438) and Center for Biomedical Imaging
(CIBM) of the EPFL and Universities and Hospitals of Lausanne and Geneva.
• The EU Marie Curie Action under grant FP7-PEOPLE-2011-IOF # 299500.
• The Laboratory of Behavioral and Cognitive Neuroscience, Stanford University.

18

CHAPTER 1. INTRODUCTION

Part I

Description of PRoNTo tools

19

Chapter 2

Data & Design
Contents
2.1
2.2

2.3

2.4

2.1

Introduction . . . . . . . . . .
Methods . . . . . . . . . . . .
2.2.1
Data and design input
2.2.2
Data and design output
2.2.3
Review . . . . . . . . .
2.2.4
HRF correction . . . .
Graphical User interface . . .
2.3.1
PRT directory . . . . .
2.3.2
Groups . . . . . . . . .
2.3.3
Subjects . . . . . . . .
2.3.4
Modalities . . . . . . .
2.3.5
Masks . . . . . . . . . .
2.3.6
Review . . . . . . . . .
2.3.7
Load, Save and Quit .
matlabbatch interface . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

21
22
22
22
22
22
24
24
24
25
25
28
29
29
30

Introduction

The first step in a statistical analysis of neuroimaging data, whether it’s in a pattern recognition
or general linear model (GLM) framework, usually entails providing to the analysis software all
the information regarding the data and experimental design. PRoNTo is no exception. After
preprocessing the data (if required), the analysis in PRoNTo starts with the ‘Data and Design’
module. It is important to note that PRoNTo does not perform any spatial or temporal preprocessing, and if not performed with another software, pattern recognition might be affected by
misalignment and noise in the data.
In the ‘Data and design’ module the user can enter the image/scan files, experimental conditions (TR, durations and onsets of events), as well as other parameters, covariates and regression
values. PRoNTo supports multi-modality datasets and therefore it allows the user to enter more
than one data modality, such fMRI, MRI, PET and ASL, per analysis. This module is therefore
essential for the rest of the framework and stores all the information that is needed from the data
to be used by the rest of the software modules, such as feature set preparation, model specification
and estimation.
Below is a summary of what the ‘Data and Design’ module does. The Methods section
discusses how the module is organised and what its main output is. It also mentions a few issues
that need to be taken into consideration when entering the information and how they affect
subsequent steps. This chapter then presents the graphical user interface (GUI) that is used
to enter the data and design information and how it is used. Finally, the chapter finishes by
mentioning the corresponding ‘Data and Design’ matlabbatch module, and particular issues that
do not apply to the GUI.
21

22

2.2
2.2.1

CHAPTER 2. DATA & DESIGN

Methods
Data and design input

PRoNTo provides two types of interfaces for entering the data and design information, a PRoNTospecific graphical user interface (GUI) and the matlabbatch system that is also currently used
by SPM. These two interfaces are also available for the other modules, as discussed in the Introduction chapter.
The information that needs to be entered is almost exactly the same for both the GUI and
batch (the small differences are explained below in the matlabbatch section) and, more importantly, the output is exactly the same. Therefore it is up to the user to decide which system is
best suited for his/her analyses. For instance, the GUI can be used as a first approach to the
toolbox and by users not familiar with SPM, whilst the batch can be used by more advanced or
SPM users, who know how to take advantage of the batch system to optimise their analyses.
As mentioned, PRoNTo supports multi-modality analyses. Therefore the data and design
module is prepared to receive as input the following types of data: fMRI, sMRI, PET and beta
images (created from a previous GLM analysis). Other types of data can also be entered at the
user’s risk, as long as they comprise nifti files.
Regardless of which interface the user chooses to enter the data and design (GUI or batch),
the organisation is very similar and starts (after choosing the directory to save PRT.mat) with
the definition of Groups. In neuroimaging datasets, it is common to have a few subjects with a
lot of images/scans per subject, such as the time-series in fMRI. However, the opposite is also
common: lots of subjects with one image per subject, such as encountered in PET or MRI studies.
Therefore, for each group, PRoNTo provides two ways of entering the rest of the information, i.e.
subjects, modalities and design, which are referred to as the ‘select by subject’ or ‘select by scans’
option, respectively (as is shown below). If one chooses to enter the data by ‘scans’, PRoNTo
allows the user to enter, for each modality, all subjects (one image/scan per subject) at once,
which is a lot quicker than entering each subject at a time. It is important to note that when
using Regression models this is the only way of inputing the data. As explained below, only the
‘select by scans’ option allows the users to enter regression values (one value per subject/image).
This option however is not appropriate for modalities which have an experimental design and
more than one image per subject, such as fMRI. For these datasets the user should choose the
‘subjects’ option. For each subject one can specify the modalities, experimental conditions and
enter more than one image/scan. Both options are valid and produce exactly the same output
structure (if used with the same dataset).

2.2.2

Data and design output

The output of the ‘Data and Design’ module is the PRT structure (as discussed in the Introduction). This structure contains subfields with all the information that is needed from the data for
the subsequent analysis steps and it is saved in a ‘PRT.mat’ file. For advanced users the fields of
this structure can be edited directly and saved, therefore bypassing the need to use the GUI or
matlabbatch to create the PRT. However, this structure is the core of PRoNTo and should be
carefully created because it affects everything else.

2.2.3

Review

The ‘Data and Design’ module also allows the user to review the information that has been
entered (through the GUI, batch or manually). The main aim of the ‘Review’ function is to check
if the data and design has been correctly specified. It can also be used to inspect if the design
is appropriate for subsequent analysis. For example, the review window shows the number of
subjects in each group, and for modalities with experimental design, it can be used to show and
alter the number of used and unused scans (see below).

2.2.4

HRF correction

For datasets such as fMRI, there is a very important issue that needs to be carefully addressed
when specifying the data and design. As is well known, the hemodynamic response function

2.2. METHODS

23

(HRF) is a delayed and dispersed version of the underlying neuronal response to an experimental
event (Figure 2.1). This means that, depending on the TR, the effect of the HRF can be felt
over multiple scans, and therefore the acquired scans are not independent and might contain
information from both past and present events. This can confound subsequent analyses and
needs to be accounted for. For instance, in SPM, the stimulus time-series are convolved with a
canonical HRF. Although convenient in the GLM framework, the convolution approach is not
appropriate in the pattern recognition context. Therefore, the solution used in PRoNTo is to
discard all overlapping scans. This is done as follows: PRoNTo allows the user to control a delay
(time it takes for the hemodynamic response to peak after the stimulus) and overlap (width of
the response) parameter that determine the shape of the HRF. As can be seen in Figure 2.1, the
delay means that the scans corresponding to a given condition are actually shifted in time, and
the overlap means that the number of independent scans, for which the signal corresponds only
to a given condition, is smaller than the total number of acquired scans for each condition. Given
the delay, PRoNTo finds which scans correspond to each condition and discards the last scans in
the time-series for which the response has not yet peaked. It then uses the overlap to determine
which consecutive scans contain information from only one condition (i.e. the response does not
overlap with the response from the previous condition) and discards the ones for which there is
overlap (as shown in Figure 2.1, bottom right). The discarded scans are not actually deleted but
are not used in further analyses.
When using the GUI, the default value for the HRF parameters is 0 seconds and can only be
changed in the ‘Review’ window (as shown below). Therefore, for fMRI, the user should review
the data and design and change these parameters to a more appropriate value (e.g. 6 seconds
each). In the matlabbatch, the default value for these parameters is also 0 seconds but can be
changed directly within the batch (no need to open Review window). Again, for fMRI, these
values should be changed (e.g. to 6 seconds).
Importantly, if one wants to avoid discarding scans and having to correct for the shape of the
HRF, as explained in the above paragraph, one should use as input the beta (coefficients) images
obtained by first running a GLM analysis on the original data. The GLM analysis accounts for
the HRF delay and overlap using the convolution approach. This is normally the best approach
in case of rapid event-related design experiments, in which there can be a lot of overlap, i.e. the
number of discarded scans can be very high.

Figure 2.1: HRF correction. On the left is the standard HRF response. On the right is the effect
of the delay and overlap on the number of independent scans (C1, C2 and C3 correspond to
three different experimental conditions and the blue boxes correspond to various scans acquired
during each condition). In fMRI datasets, the nature of the HRF (i.e. being a delayed and
dispersed version of the neuronal response to an experimental event) might lead to less independent scans/events than the ones originally acquired. In PRoNTo, this issue is accounted for by
discarding overlapping scans.

The steps to specify the information relative to the data and design using both the GUI and
the matlabbatch system are described in the following sections.

24

2.3

CHAPTER 2. DATA & DESIGN

Graphical User interface

The graphical user interface to specify the data and design is presented in Figure 2.2. This GUI
can be launched by typing ‘prt’ in the Matlab window and then clicking the first button on the
left, in the main steps panel.

Figure 2.2: Data and design graphical user interface. This interface allows the user to enter all the
information relative to the data, including the experimental design and masks. After introducing
all the fields, PRoNTo creates the PRT structure, which is saved in the specified directory, as
‘PRT.mat’ file.

2.3.1

PRT directory

The first thing the user should specify is the directory in which to save the PRT structure. This can
be done by browsing existing directories (previously created by the user) from the top of the data
and design interface (Figure 2.2). It is recommended to have different directories for different
datasets (note that a dataset can include different modalities in case of multimodal analysis)
because PRoNTo overwrites an existing PRT in the selected directory. The later modules in
PRoNTo will then add more fields to this structure with further information, such as the models,
features and kernels used in subsequent analyses. The file created is called ‘PRT.mat’.

2.3.2

Groups

The group panel allows one to add or remove a group of subjects. The minimum number of
groups is one, but there is no maximum number. When ‘Add’ is clicked, the user should provide
a name to the group. Any alphanumeric string is sufficient and there should be no spaces in the
string (this applies to all names throughout the toolbox). The name of the group can be later
modified by right clicking on the name. When ‘Remove’ is clicked, all the information relative to
this group (including all subjects and corresponding data) is deleted. PRoNTo does not restore
the deleted information and it can only be re-entered again by clicking ‘Add’.
The following panel after ‘Groups’ is ‘Subjects/Scans’. Here, as mentioned above, there are
two ways of entering the data: by ‘subjects’ or by ‘scans’. The former is chosen by clicking ‘Add’
under the ‘Subjects/Scans’ panel and filling in the fields for each added subject at a time. The
latter is done by clicking the tick box ‘Scans’ under the ‘Subjects/Scans’ panel. The subjects
panel is then de-activated and the user can enter the modalities and files straight away. The fields
to be filled under these two options are described below.

2.3. GRAPHICAL USER INTERFACE

2.3.3

25

Subjects

Select by scans The ‘select by scans’ option allows the users to skip the subject step. To
identify that this option has been selected, PRoNTo writes ‘scans’ in the subjects panel (Figure
2.3). The user can then add modalities and for each modality a new window will appear (bottom
of Figure 2.3).
It is important to remember that when the ‘scans’ box is clicked all the information in the
subjects panel is automatically deleted. Unselecting the ‘scans’ box also deletes all the information!
Select by subjects The ‘Subjects/Scans’ panel allows the user to add/remove subjects. This
panel works exactly like the groups panel, but the subject name is automatically generated. This
name can be later modified by right clicking on it. For each subject one can then specify the
modalities in the next panel.

Figure 2.3: Data and design graphical user interface. If one chooses to specify everything using
the ‘Scans’ option (tick box below the ‘Subjects/Scans’ panel), one can introduce the data for
all subjects at once for each modality, but one cannot specify any design. This is the optimal
approach when one has a lot of subjects with only one image/scan per subject, such can be the
case of MRI and PET datasets.

2.3.4

Modalities

The modalities panel works like the group and subjects panel, but allows one to add and remove
modalities. When a modality is added, a name needs to be provided (unless the modality has
already been defined for a previous subject or through the masks menu, see below). It is important to note that a different modality can be a different type of data, such as fMRI and PET,
or a different session of the same type of data, e.g. different runs/sessions of the same fMRI
experiment. This way the different sessions can be integrated later into the same model and
analysis.
The steps to enter the modality information are slightly different if one ticks the ‘scans’ box
or not.

26

CHAPTER 2. DATA & DESIGN

Select by scans Here the data is assumed to have been acquired without an experimental
design, and therefore the ‘No design’ option is automatically selected and cannot be changed
(bottom window in Figure 2.3). However, in select by scans, the user can also introduce ‘Covariates’, i.e. a variable that covaries with the data (subjects) but of no interest to the subsequent
analyses. This option will be functional in version 2.1 of PRoNTo. It requires the input of a matrix, with one row per image/subject. This matrix can either be entered as a Matlab command
in the editable box, or the full path to a .mat containing the matrix should be provided (matrix
named ‘R’). This last option is recommended to input a matrix (i.e. more than one covariate).
The last empty field can be used to enter ‘Regression targets’ (Figure 2.3). This option allows
the users to introduce a real number per subject to be used later for regression if that is the case.
As mentioned above, this is the only way of entering the data and regression values when doing
Regression models.
Select by subjects When entering the data by subjects, the modality window allows one to
specify the experimental design (Figure 2.3). Here there are three options. The last option is
simply ‘No design’, which means that for this modality there are no experimental conditions
(this option is normally used when there is only one image per subject e.g. structural MRI or
beta images from GLM analysis). The first option is to load an SPM.mat with a previously
specified design. This option can be chosen if the user has created an SPM structure containing
all the experimental fields using the SPM software. In this case, the user does not need to specify
anything else, only the files (scans/images) for this subject/modality. The design information is
extracted directly from the SPM structure and saved in PRT.mat. Finally, the ‘Specify design’
option allows one to introduce all the conditions (durations and onsets), TR and other parameters
corresponding to the experimental paradigm used for this subject and modality (this option is
normally used with time series data, e.g. fMRI). After the design of the first subject has been
specified, a new option will appear in the menu that allows to ‘Replicate design of subject 1’,
for the same modality and group. This facilitates design specification for groups of subjects with
controlled (i.e. non-random) event onset and duration.

Figure 2.4: Data and design graphical user interface. The design menu in the modality window
(when one uses the select by subject option) allows one to load a previously specified design
from an SPM.mat file, create a new design or simply select no design, which usually applies to
modalities where there is no experimental task, such as MRI or PET.

Design To create a new design one selects the option ‘Specify design’ as explained in the
previous paragraph (Figure 2.4). This will then open another window (after choosing how many

2.3. GRAPHICAL USER INTERFACE

27

conditions you have) (Figure 2.5). In this window one can then write the names, onsets, and
durations of each condition. The units in which this information is read is specified below. There
are two options ‘Scans’ or ‘Seconds’. If the unit scans is selected, it is good to bear in mind
that PRoNTo follows the convention, adopted in SPM, that the first scan is scan 0. In the
durations field, one can introduce as many values as the number of onsets or just simply one
value, which assumes the events all have the same duration. In this window there is also the
option of introducing the Interscan Interval (TR), which is always read in seconds.
One issue to have in mind when specifying the design is the following: if there are more scans
than experimental events, these extra scans will not be used in later analyses. They are not
deleted and the corresponding indexes can be found in the PRT structure:
PRT.group(g).subject(s).modality(m).design.conds(c).discardedscans.

Figure 2.5: Data and design graphical user interface. The ‘specify conditions’ window is available
from the modality interface when the user chooses to enter the data by subjects and clicks ‘specify
design’. This window is used to enter the conditions (names, onsets and durations) as well as the
units of design, TR and covariates.

Modify design The user can later modify a design by loading a PRT.mat in the Data and
Design window. Please note that if feature sets or models have been previously computed, they
will be discarded if changes are performed to the dataset. If the user wants to keep those, he/she
should change the directory before saving any modification to the design.
After loading a previously saved PRT, any change can be performed: subjects, groups or files
can be added or removed. If the design needs to be modified, a right-click (ctrl+left-click in Mac)
on the name of the concerned modality proposes to re-open the modality definition window. To
review or modify the onsets/durations/blocks, the user can access their definition via the ‘specify
design option’. Similar right-clicks (ctrl+left-click in Mac) allow renaming groups or subjects.
To modify the HRF parameters (delay or overlap), there is no need to load the PRT in Data
and Design. Loading it within the Data Review allows the user to keep all previously computed
feature sets and models. However, if the HRF parameters are changed, feature sets have to
be computed anew since they do not correspond to the modified design. Changing the desired
parameter (e.g. replacing ‘0’ by ‘6’) and hitting the ‘return’ key updates the PRT directly in
terms of scans selected for modelling. Please remember to keep an eye on the Matlab window,
since important information are displayed on the workspace!

28

CHAPTER 2. DATA & DESIGN

Files Finally, independent of the way the user entered the information (by subjects or scans)
the ‘Files’ option allows one to choose which image files to use (Figure 2.6). This will open
another window that shows all image files available in each directory. These can be selected one
by one or all at once, by using the mouse’s right button on the right panel of the window (or shift
key).

Figure 2.6: This window is called when one clicks ‘Files’ and is used to select the scans/images
for each subject/modality.

All that is needed for each group, subject and modality has been specified and can now be
viewed on the main window (Figure 2.7) under each panel. The last panel shows which files have
been entered for each modality and can be modified directly (click Modify). When Modify is
clicked and no files are then selected all the previous files are deleted! Figure 2.7 shows how the
data and design interface should look like once all the fields have been specified (using select by
subject). The design and files for each modality can also be modified by right clicking on the
modality name in the modality panel. This option can be useful to visualise the design (onsets
and durations) that has been previously entered and change it if necessary. For instance, one can
check the design of the first subject and if changes are needed these can then be replicated for all
other subjects as explained above.

2.3.5

Masks

This popdown menu on the bottom of the main data and design window is where the user enters a
binary image mask for each modality. This mask can be previously created by the user or simply
chosen from a list of default masks available in the masks directory of PRoNTo. Every modality
has to have a mask, which can be the same for all modalities. This is a first-level mask and is used
simply to optimise the prepare feature set step by discarding all uninteresting features, such as
voxels outside the brain. Later in the analysis one can choose another mask (second-level mask)
that is more relevant to the scientific question and that can, for example, restrict the analysis
to certain areas of the brain. To specify the mask one needs only to select the modality and
then enter an image file. If the modalities have not yet been created, then one can create the
modalities here, which will then appear in the modality panel.
Important note: If the first-level mask overlaps with areas that do not have values (e.g.
either 0 or NaN) in the specified images, those areas will still be taken into account for further
analysis. This might affect the results if those areas are not the same across images (typically,
performance will be lower). We therefore advise the user to check the overlap between the first-

2.3. GRAPHICAL USER INTERFACE

29

Figure 2.7: Data and design graphical user interface. After filling in all the fields using the select
by subject option (the select by scans case is very similar) the data and design interface should
look like this example figure.

level mask and his/her data. In most cases, this will not be needed, but for the acquisition of
e.g. specific slices, this is recommended. We provide a script to update the mask automatically
for beta images derived from a SPM GLM analysis (see 1, Inputs and preprocessing).

2.3.6

Review

The ‘Review’ button allows one to review the data and design for each modality (Figure 2.8).
On the top right is the information relative to the number of groups and modalities that have
been entered. The plot on the left displays the number of subjects per group. This is particularly
important to check if the design is too unbalanced in terms of subjects. Then on the bottom right
panel is the design information for each modality (if the selected modalities have an experimental
design). Here, the user can view the number of conditions and can also edit the parameters that
control the HRF delay and overlap (as explained above). The user can change the default value of
0 seconds and the effect is immediately seen on the number of scans plotted on the left (number
of selected scans and number of discarded scans for each condition). The higher the value of
the HRF peak and overlap, the higher the number of discarded scans. One can also read on the
main Matlab window information regarding which group/subjects have had some scans discarded.
The information below the HRF parameters corresponds to the interval between successive scans
before and after the HRF delay/overlap correction. These values also change according to the
changes entered in the boxes above. Please note, as mentioned in the section ‘Modify design’, that
information regarding the PRT being updated after changing the HRF parameters is written on
the main Matlab window. Once again, if you have previously computed feature sets and models,
you have to recompute them because they do not correspond to the data anymore (changing the
HRF delay and overlap parameters changes the data). The information regarding which scans
have been removed or not from the analysis can be found in the PRT structure:
PRT.group(g).subject(s).modality(m).design.conds(c).hrfdiscardedscans.

2.3.7

Load, Save and Quit

The ‘Save’ button allows the user to create the PRT.mat file with the PRT structure containing
all the information that has been specified here (Figure 2.7). Incomplete information cannot be
saved. At least one group should have all the required fields so that PRT.mat can be created.
‘Load’ allows the user to load the data and design information from a previously saved PRT.mat.

30

CHAPTER 2. DATA & DESIGN

Figure 2.8: Data and design graphical user interface - ‘Review’ window. This window allows the
user to check the data and design, including the number of subjects per group. It also allows the
user to change the HRF delay and overlap parameters that control the number of discarded scans
(appropriate only for modalities such as fMRI). When there is no experimental design only the
top plot and information is shown.

The user can then edit the fields and update PRT by clicking again the ‘Save’ button. It’s very
important to click ‘Save’ because all the other steps in the analysis rely on the PRT structure.
Without this structure one cannot proceed. However, when the PRT.mat contains fields that have
been added by the ‘Prepare feature set’ or other modules, if the Save button is clicked, these
fields will be deleted. The option ‘Quit’ allows the user to leave the interface without saving the
information. This is also the case when the user closes the window without first using the Save
button.

2.4

matlabbatch interface

The ‘Data and Design’ module in the matlabbatch is called either by first typing ‘prt’ and clicking
the ‘Batch’ button or by typing ‘prt batch’. The user can then find on top of the batch a PRoNTo
menu and under this menu the first module corresponds to the data and design module.
The options presented in the ‘Data and Design’ GUI, mentioned above, are all available in
the matlabbatch interface (Figure 2.9). However, there are a few things in the batch that differ
from the GUI. One issue to note here is that, when using the batch one needs to be very careful
with the names of the modalities specified for each subject (or using select by scans) and specified
for each mask. The number of modalities should be exactly the same for each group and subject
and the names should be consistent between groups/subjects and correspond to the names of the
modalities under the masks field. In the GUI the names are made automatically consistent. The
names of the conditions should also be the same across subjects and will be later used to define
classes in the ‘Specify model’ batch module.
Another issue is the HRF delay and overlap correction values. In the batch, the user can

2.4. MATLABBATCH INTERFACE

31

directly alter these values (instead of having to use the ‘Review’ window) but the default is 0
seconds and should be changed (e.g. to 6 seconds) for modalities that depend on the HRF, such
as fMRI.
As mentioned in the Introduction, the batch job can be saved as a .mat, and loaded again
whenever needed, or as a .m that can be edited using the Matlab editor. This is a powerful tool
that can make the specification of the data and design a lot easier and quicker, for example by
editing and scripting existing batch files (for further information see the matlabbatch chapter
below).

Figure 2.9: Data and design module in matlabbatch. The matlabbatch contains two extra
options relative to the Data and Design interface. These options allow one to specify the delay
and overlap of the HRF response (in the GUI it can only be changed in the ‘Review’ window),
and which are then used to determine the number of discard scans.

32

CHAPTER 2. DATA & DESIGN

Chapter 3

Prepare feature set
Contents
3.1
3.2
3.3
3.4

3.1

Introduction . . . . . . .
Methods and resources .
Graphical User interfaces
matlabbatch interface . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

33
33
35
37

Introduction

One of the main inputs of a machine learning algorithm consists in a Nsamples × Nf eatures data
matrix, containing the values of selected features for each sample. This matrix can either be
input directly into the machine or be used to compute a “similarity matrix”, or kernel, of size
Nsamples × Nsamples , which is then input into the classification/regression algorithm [see “kernel
trick” [8, 1]]. PRoNTo computes a linear kernel (i.e. dot product) between the samples. The
‘Prepare Feature Set’ step computes both the feature and linear kernel matrices from one or more
modalities, as defined in the previously built dataset (see chapter 2). It allows detrending the
features in the case of time series (such as fMRI) and scaling each image by a constant factor
(input by the user) in the case of quantitative modalities (such as PET). Masks can be specified
to perform the classification/regression on specific voxels only (e.g. Regions of Interest).
Multiple runs entered as different modalities (e.g. modality 1 is ‘fMRI run1’, modality 2 is
‘fMRI run2’,...) can be concatenated in terms of samples during this step. The images from
different runs should then have the same number of features (i.e. selected voxels). In addition,
version 2.0 allows to build multiple kernels, either from multiple modalities, or based on different
anatomically labelled regions as defined by an atlas. In the case of multiple modalities, it is
required that the selected modalities have the same number of samples, i.e. images.

3.2

Methods and resources

After the selection of the dataset and of which modality to include in the feature set (further
referred to as FS), the toolbox accesses each image, i.e. it gets the value of the voxels which
are comprised in the first level mask selected for that modality (mask specified at the data and
design step, see chapter ‘Data and Design’). This access is performed by ‘blocks’ of features, not to
overload the RAM memory. In the case of time-series, the user can specify detrending methods
and parameters to apply to the time course of each feature. Methods comprise a polynomial
detrending (parameter: order of the polynomial) or a Discrete Cosine Transform high-pass filter
(SPM, parameter: frequency cutoff in seconds). An example of a linear detrending (polynomial
detrending of order 1) is shown in Fig. 3.1.
For each modality, the (detrended) features are then written in a file array (SPM, with a
‘.dat’ extension), on the hard drive (in the same directory as the dataset). Please note that in
the case of large datasets, this operation may require many Gb of free space on the hard drive
33

34

CHAPTER 3. PREPARE FEATURE SET

Figure 3.1: Example of detrending: the original signal over time of one feature (in blue) was
approximated by a polynomial of order 1 (red line), which was then substracted from the original
signal to give the detrended signal (in green).

and long computational times. Therefore, if the first condition can’t be fulfilled, we recommend
the use of external drives for the whole analysis. Regarding the computational expenses, we
tried to minimize their effect by computing the features only once per modality: when preparing
other feature sets using the same modality and detrending parameters, the built file array will be
accessed for the next steps.
Be careful that using the same modality but different detrending methods and/or parameters
will force the re-computation of the file array for the considered modality. In the same way,
changing the dataset (PRT.mat) from directory might lead to the re-computation of the feature
sets if the file arrays were not moved accordingly.
From the feature set(s), a linear kernel can then be computed. Different options can be
specified:
• All scans/ All conds: In ‘all scans’ the kernel matrix will be computed between all scans
within the time series of all subjects and in ‘all conds’ the kernel matrix is computed
only between the scans corresponding to the specified conditions of interest (see ‘Data and
Design’). By default, the toolbox will use all scans to compute the kernel. With large
datasets however, computational expenses can be reduced by selecting the last option.
• Scaling: allows the specification of constant values to scale each scan. The user has to enter
a .mat containing a variable called ‘scaling’ and of the same size as the number of scans in
that modality. In case of quantitative modalities such as PET, this step is required since it
insures the convergence of the machine learning algorithm.
• Additional mask for selected modality: this option allows the specification of a ‘second-level’
mask, which would for example define Regions of Interest (ROIs) on which the classification/regression can be performed. In this case, the voxels used to compute the kernel (and
only the kernel) would be the ones contained in both the first and second-level masks.
Therefore, using one first-level mask and two second-level masks would create two kernels
but only one file array.
• Build one kernel per region: starting from version v2.0, PRoNTo allows to build one kernel
per region as defined by an anatomically defined atlas, specified by the user1 . One atlas
(Anatomical Automatic Labelling, AAL) is provided in your PRoNTo folder/atlas. Atlases
can be generated easily through SPM, or manually by the user. There are no constraints
on how regions are built, as long as all the voxels within each region have a specific integer
1 The atlas corresponds to a mask, except that the value of the voxels in each defined area correspond to a
unique value, e.g. all voxels in fusiform have the value 3, and all voxels in orbito-frontal have the value 50.

3.3. GRAPHICAL USER INTERFACES

35

value. The toolbox will identify the different regions based on the values in the voxels. Each
region will then act as a second-level mask and one kernel will be built for each region. The
kernels are all saved in a same feature set and will then all be used at the modelling stage.
These options are performed at the kernel level only. This means that any change in one of
these options would lead to the computation of a new kernel but not to the (re)computation of
the file arrays. The use of different second-level masks or scaling parameters can therefore be
easily envisaged.
ProNTo version 2.0 allows to build multiple kernels. These kernels can be derived from multiple
modalities or from multiple regions of interest as defined by an atlas within each modality. These
two options are not mutually exclusive and it is also possible to build multiple kernels within
each modality and then combine those modalities as multiple kernels. The number of kernels
would hence become number of modalities × number of regions. In the same way, it is possible
to concatenate multiple runs of an experiment while building one kernel per region.
The PRT.mat structure saves all information linked to the file arrays in a fas field (standing
for “File Array Structure”), which size corresponds to the number of selected modality in all
feature sets. The selected options and the link to the kernel (saved on the disk as a .mat) are
stored in a fs field (standing for “Feature Set”), which size corresponds to the number of feature
sets defined by the user.

3.3

Graphical User interfaces

After clicking on the “Prepare Feature Set” button in the main interface (see Fig. 3.2), a second
window will appear (Fig. 3.3), allowing the user to select a saved PRT.mat, to name the FS and
to define the number of modalities which should be included in the FS.

Figure 3.2: Main interface: button to launch the ’Prepare Feature Set’ step.

To define the number of modalities to include, the user should click in the appropriate edit
box, type the number and then ‘return’. This will launch a third window (Fig. 3.4), allowing
the specification of the different options and parameters for each modality. When the dataset
contains only one modality, this window is launched directly and (Fig. 3.3) is filled automatically
expect for the feature set name.
In this third window, the user has to choose which modality to include based on its name
(first pull-down menu) and which scans to use to build the kernel (all or only those linked to the
design). All other options are facultative. The first panel refers to operations to perform on the
features:

36

CHAPTER 3. PREPARE FEATURE SET

Figure 3.3: Interface of the ‘Prepare Feature Set’ step. Top: Dataset selection: type the full
name (with path) or browse to select the dataset to prepare. Feature set name: Type the FS
name, which will be used to save the kernel as a .mat on the hard drive. Modalities: Number
of modalities to select with the list containing the names of the modalities included in the FS
(no user interaction possible). A checkbox allows to build one kernel per modality if multiple
modalities are present/have been selected in the FS. Build kernel/data matrix: builds the feature
set and kernel(s).

Figure 3.4: Specification of options and parameters for each modality. Modality: Select the
modality name from a pull-down menu. Conditions: choose to build All scans or All conditions.
Parameters: Detrend to perform with its parameter, as well as Scaling of the scans or not.
Features: Selection of a second-level mask and/or of an atlas to build one kernel per region.

• the detrending parameters: by default, the parameter is set to ‘No detrending’. However,

3.4. MATLABBATCH INTERFACE

37

we recommend to perform a detrending in the case of time series data such as fMRI (and
only in that case). When selecting polynomial, the ‘order’ parameter will appear, with a
default value of 1. Changing this value will increase the order of the polynomial used to
fit the data. If ‘Discrete Cosine Transform’ is selected, the editable parameter corresponds
to the cutoff frequency (in seconds) of the high-pass filter. Please note that, when including more than one run (‘modality’) into a feature set, nothing will prevent the user from
using different detrending methods/parameters. We however highly recommend to use a
consistent detrending in the same FS.
• the scaling: ‘no scaling’ is the default option. However, when dealing with quantitative
modalities such as PET, the user should provide one value per scan, stored in a vector in a
.mat file under the variable name ’scaling’.
As previously mentioned, the detrending is performed before the features are saved in the file
array, while the scaling is performed only when building the kernel.
The second panel allows to select a subset of the saved features to build the kernel. Two
options are available:
• the specification of a second-level mask: type the full name (with path) of the mask or
browse to select the mask image. When left empty or untouched, voxels are selected from
the first-level mask specified in the data and design step. Otherwise, voxels within both the
first and second-level masks will be selected to build the kernel.
• the building of one kernel per region: when selecting this option, the user should load an
atlas which defines regions in terms of their anatomy (in MNI space). Each region will then
act as a second-level mask and one kernel will be built for each region.
When working with Graphical User Interfaces (GUIs), some messages might appear in Matlabworkspace. These can display information about the operations currently performed or explain why the toolbox does not do as the user expected (e.g. when a file could not be loaded or
if information was input in a wrong format). Therefore we strongly encourage the user to have a
look at Matlabprompt when using GUIs.

3.4

matlabbatch interface

The matlabbatch system allows the input/selection of all parameters and options aforementioned.
Just note that the batch is based on the names of the modalities and/or conditions. Therefore,
for the batch to work properly, names should be consistent across all steps, starting from data
and design to the model specification and running. The hierarchy for the case of a feature set
containing one fMRI modality is displayed in (Fig. 3.5). For this feature set, we chose to load an
atlas, and build multiple kernels based on the regions it defines.
Important note: Defining all important steps in one batch and running that batch will
overwrite the PRT.mat previously created and thus delete the links between the PRT.mat and
the computed kernel(s) and feature set(s). The file arrays would then be recomputed each time
the batch is launched. For large datasets, we therefore recommend splitting the batch in two
parts: a data and design and prepare feature set part and a second part comprising the model
specification, run model and compute weights modules. This would indeed allow changing, e.g.
model parameters, without recomputing the feature sets and kernels.

38

CHAPTER 3. PREPARE FEATURE SET

Figure 3.5: matlabbatch GUI for feature set building.

Chapter 4

Model Specification and
Estimation
Contents
4.1
4.2
4.3
4.4

4.5
4.6
4.7

4.1

Introduction . . . . . . . . . . . . . . . . . .
Beginning a model specification . . . . . . .
Feature set . . . . . . . . . . . . . . . . . . .
Model type / pattern recognition algorithm
4.4.1
Classification . . . . . . . . . . . . . .
4.4.2
Regression . . . . . . . . . . . . . . .
4.4.3
Hyper-parameter optimization . . . .
Cross-validation . . . . . . . . . . . . . . . .
Specify / Run model . . . . . . . . . . . . .
Batch interface . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

39
39
40
40
41
42
42
43
46
47

Introduction

The specification of a model is the core step of the pattern recognition pipeline and entails setting
up the combination of the different components making up the analysis. For example, model
specification is where you select which data features to use as input (i.e. a feature set), the type
of prediction to perform (e.g. classification or regression), which machine learning algorithm to
employ (e.g. support vector machines, Gaussian processes, ...), which cross-validation strategy to
employ (e.g. leave one subject out, leave one run out, ...) and which operations or manipulations
to apply to the kernel before the algorithm is trained. The framework provided by PRoNTo is
highly flexible and supports most types of pattern recognition analysis typically performed in
neuroimaging. This chapter provides an overview of each of the components making up a model
in PRoNTo. The presentation will focus on the user interface although it is important to note that
the batch system provides several advanced options not available in the user interface (described
further).

4.2

Beginning a model specification

To begin a model specification with the PRoNTo user interface, select ‘Specify model’ from the
main PRoNTo window. This will launch the model specification window (Figure 4.1)
Next, select the PRT.mat containing your experimental parameters. Note that at least one
feature set must be defined in this structure before a model can be created. See chapter 3 for
details on constructing feature sets.
Enter a unique name to identify the model, which is used internally in PRoNTo, by the batch
system and for display purposes. It is a good idea to select a meaningful but short name (without
39

40

CHAPTER 4. MODEL SPECIFICATION AND ESTIMATION

Figure 4.1: Model specification graphical user interface

spaces). Note: the PRT.mat data structure retains a permanent record of all models created
but if a model with the specified name already exists in the PRT.mat data structure, it will be
automatically overwritten.

4.3

Feature set

The drop-down list entitled ‘Feature set’ will be populated once a PRT.mat containing one or
more feature sets is selected. Select the appropriate feature set from the drop-down list. Note
that a single feature set may contain more than one data modality (see chapter 3), which can
be combined to build multimodal classification and regression models (e.g. L1-Multiple Kernel
Learning models). This might also be useful if more than one run/session is available for each
subject, in which case each run could be input as an independent modality in the data and design
step and a single-subject classifier might be specified using leave-one-run-out cross-validation.
In the current release of PRoNTo, only kernel machines are supported via the user interface.
The capability to support non-kernel techniques will be added in a future release. Thus, the ‘Use
kernel’ radio button should always be set to true.

4.4

Model type / pattern recognition algorithm

In this part of the model specification input form, select the pattern recognition algorithm to employ (referred to in PRoNTo as a ‘machine’). In the current release, three classification algorithms
are supported (binary support vector machines, Gaussian processes (binary and multiclass) and
L1-Multiple Kernel Learning) and four multivariate regression methods (Gaussian process regression, kernel ridge regression 1 , relevance vector regression and L1-Multiple Kernel Learning).
Note: if a feature set contains multiple kernels (either from regions of interest or based on
different modalities) but the selected classification/regression technique is a single kernel method
(e.g. SVM or KRR), the kernels will first be summed before entering the classification/regression
1 Kernel ridge regression is equivalent to a maximum a posteriori approach to Gaussian process regression with
fixed prior variance and no explicit noise term

4.4. MODEL TYPE / PATTERN RECOGNITION ALGORITHM

41

phase. This corresponds to concatenating the features before building the kernel. For regions of
interest in a single modality, the summed kernel is hence equivalent to a whole brain model.
The PRoNTo user interface provides a mechanism for flexible definition of which components
of the data design should be used for each classification or regression model. Note that this will
not necessarily be the whole experiment; for example, in a complex fMRI experiment there may
be several groups, each containing multiple subjects, each in turn having multiple experimental
conditions (e.g. corresponding to different subprocesses of a cognitive task). In such cases, it is
usually desirable to ask several different questions using the data, such as discriminating between
groups for a given experimental condition (“between group comparison”), discriminating between
experimental conditions for a fixed group (‘between-task comparison’) or training independent
pattern recognition models for different subsets of subjects. All of these can be easily defined via
the user interface by clicking the ‘Define classes’ button (for classification) or ‘Select subjects/scan’
(for regression).

4.4.1

Classification

The class selection panel is displayed in figure 4.2. First, define the number of classes, noting that
some classification algorithms (e.g. support vector machines) are limited to binary classification,
while other classification algorithms (e.g. Gaussian processes) support more than two classes.
Enter a name for each class - again, it is a good idea to make these names informative but short.
Notice that immediately after the number of classes has been specified, the group-, subject- and
condition selection panels are greyed out. To enable them, simply select one of the classes from
the drop-down menu.
For each class, select the subjects and conditions (if any) that collectively define that class.
It is possible to select multiple experimental conditions in the same class, but this complicates
model interpretability and potentially also model performance (since by definition conditions are
not identically distributed). If a condition or subject is erroneously selected, click on it in the
‘selected subject(s)’ or ‘selected condition(s)’ panel and it will be removed from the list. The
performance of classification models is evaluated based on measures such as total accuracy, class
accuracies and positive predictive values (representing the sensitivity and specificity).

Figure 4.2: Subject / condition selection panel for classification models

42

4.4.2

CHAPTER 4. MODEL SPECIFICATION AND ESTIMATION

Regression

Regression is a generic term for all methods attempting to fit a model to observed data in order
to quantify the relationship between two groups of variables. Traditionally in neuroimaging
massively univariate strategies (e.g. GLM) have been largely used, where data for each voxel
are independently fitted with the same model. Statistics test are used to make inferences on
the presence of an effect at each voxel (e.g. t-test). Multivariate regression, on the other hand,
takes into account several input variables (voxels) simultaneously, thus modelling the property of
interest considering existing relations among the voxels.
Although most studies exploring predictive analyses in neuroimaging have been related to
classification, regression analysis has aroused interest in neuroscience community for its ability
to decode continuous characteristics from neuroimaging data. This approach has potential to be
used when the examples (patterns) can be associated to a range of real values. The objective is
to predict a continuous value instead of predicting a class to which the example belongs. These
values usually refer to demographic, clinical or behavioural data (as age, blood pressure or scores
resulting from a test, for example). For validation, different metrics can be used to compute
the agreement between the predicted values and the actual ones, such as Pearson’s correlation
coefficient (r) and Mean Squared Error (MSE).
The specification of which subjects and scans to include in regression models is similar to that
for classification, see Figure 4.3 and for the purposes of model specification in PRoNTo, regression
can be thought of as a classification problem with a single class. In the current release, regression
is only supported if there is a single scan per subject (e.g. structural images or beta images from a
GLM analysis). In a future release it will be possible to perform regression where an independent
regression target is supplied for each trial, block or condition. To perform a regression, the
regression targets are specified during the design stage. It is important to emphasize that in
the current implementation, regression is only supported using the ”select by scans” option (see
chapter 2).

Figure 4.3: Subject / condition selection panel for regression models

4.4.3

Hyper-parameter optimization

From PRoNTo version 2.0, it is possible to optimize hyper-parameters of the machine learning
models. For example, the soft-margin (a.k.a. C) hyper-parameter in SVM can be optimized, using

4.5. CROSS-VALIDATION

43

a nested cross-validation scheme. In this case, there are two loops in the cross-validation scheme.
The inner loop is used for parameter optimization and the outer loop is used for assessing the
model’s performance. More specifically, the data is divided into training and testing sets according
to the cross-validation scheme selected (outer loop). For each fold of the outer loop the training set
is further divided into training and testing sets according to the cross-validation scheme selected
(inner/nested loop). The inner loop is used to train and test the model with each value of the
hyper-parameter specified by the user. The parameter leading to the highest performance in the
inner/nested loop (balanced accuracy for classification and Mean Squared Error for regression)
is then used in the outer loop. For each fold of the outer loop, the model is trained using the
’optimal’ value of the hyper-parameter and tested on the data that was left out (and which was
not used for parameter optimization). This nested CV procedure can lead to different values of
the hyper-parameter to be selected in each fold. These are stored in the outputs of the model
and can be reviewed in the ‘Display Results’ panel.
Optimizing the hyper-parameter might lead to improved results compared to fixed values.
This will usually depend on the number of features selected to model the data: for example, for
whole brain models based on SVM classifiers, with many more features than images/trials, it
is reasonable to assume that changing the hyper-parameter won’t affect the model performance
significantly due to the high dimensionality of the data with respect to the number of examples.
However, when using (e.g.) regions of interest (in a second-level mask or in a MKL model),
the ratio between the number of features and the number of examples will be much smaller. In
this case, different values of the hyper-parameter might lead to different decision functions and
optimizing the hyper-parameter is desirable.
Performing a nested cross-validation can be computationally expensive. For computational
efficiency, PRoNTo allows to specify different cross-validation schemes for the ‘outer’ and the
‘nested’ CV2 .
In the current version of PRoNTo, the soft-margin parameter can be optimized for SVM and
for MKL (classification and regression). In the same way, it is possible to optimize the λ ridge parameter for KRR. If no value is provided, those parameters will take the values 0.01, 0.1, 1, 10, 100
and 1000, i.e. 10.[−2:3] .

4.5

Cross-validation

In the final part of the specify model input form, select the type of cross-validation to employ.
Cross-validation is a crucial part of the pattern recognition modelling and is used to assess the
generalisation ability of the model and to ensure the model has not overfit to the data. Typically
this is done by partitioning the data into one or more partitions: a ‘training set’, used to train
the model (e.g. fit parameters) and a ‘testing set’ used to assess performance on unseen data. By
repeatedly repartitioning the data in this way, it is possible to derive an approximately unbiased
estimate of the true generalisation error of the model.
The most common cross-validation schemes in neuroimaging applications are leave-one-subject
out (LOSO; exclude one subject for testing, train with the remaining), leave-one-run-out (LORO;
leave one fMRI run out for testing, train with the remainder) and leave-one-block-out (LOBO;
leave out a single block or event and train with the remainder). LOSO is suitable for multisubject designs, while LORO and LOBO are suitable for single subject designs, where the former
is better suited to designs having multiple experimental runs and the latter is appropriate if there
is only a single run. The current release of PRoNTo supports each of these, and also supports
leave-one-subject-per-group-out (LOSGO), which is appropriate if the subjects in each group are
paired or for repeated measures experimental designs. Versions 1.1 and later allows k-fold crossvalidation for each of the available schemes. This means that the user specifies the number of
folds (‘k’) and that the data is partitioned according to that number. For example, specifying
k = 4 will use 25% of the data to test the model, and 75% to train it. Note: k = 1 splits the data
in half, training the model on the first half and testing on the second, i.e. there is no circular
partitioning.
2 For example, the outer CV could have more folds, to use as much data as possible in each fold for prediction
(e.g. leave-one-out), while the nested CV would not need as many folds to select the ‘optimal’ value of the
hyper-parameter (e.g. k-folds CV).

44

CHAPTER 4. MODEL SPECIFICATION AND ESTIMATION

In version 2.0, a GUI allows the user to fully specify his/her cross-validation scheme. First, a
‘basis’ needs to be specified (Figure 4.4). Three options are available:
• Load a .mat containing a previously computed CV matrix (needs to contain the variable
‘CV’).
• Select a basis from the pop-down list (contains the same options as for the outer CV).
• Specify the number of folds.

Figure 4.4: Specify basis to build custom cross-validation
When an option has been selected, a new window will appear (Figure 4.5). The top panel
of this window is a table that can be edited. Each row refers to a trial/image selected in the
definition of the classes (or to perform regression on). Each column represents a fold. For each
column, the different trials can have a value of 2 (test set), 1 (train set) or 0 (unused in this fold).
Setting a whole fold to 0 takes it out of the CV matrix. Note: it is possible to change the value
of multiple trials by changing the value of the last trial to modify, then shift-select the first one
to modify. This also works across folds. The bottom panel displays the structure of the data
selected for further classification or regression, along with a preview of the built CV matrix.
The resulting CV matrix can be saved in a .mat, alongside the PRT (name: model name
CV.mat). This matrix can loaded as a custom CV in the batch, if exactly the same trials were
selected for modelling. Note: The ‘custom’ CV option is not available as a nested/inner crossvalidation scheme.
Information concerning the cross-validation structure is stored internally in matrix format, and
can be visualised by clicking ‘Review Kernel and CV’ from the main ProNTo window (see 4.6 for
an example). In the left panel, this figure indicates which group, subject, modality and condition
each scan in the feature set belongs to. On the right, each cross-validation fold (partition) is
displayed as a separate column and each scan is colour coded according to whether it is in the
training or testing set (or if it is unused).
It should be emphasised that the type of cross-validation selected should be appropriate for
the experimental design. For example, it is nonsensical to select a leave-one-subject-out crossvalidation approach for single subject designs. It is also important to ensure that the training and
testing sets are completely independent to avoid the cross-validation statistics becoming biased.
This is particularly important for fMRI, where successive scans in time are highly correlated.
For example, if a leave-one-block-out approach is employed and the blocks are too close together
then the independence of the training and testing set will be violated, and the cross-validation
statistics will be biased (technically this is governed by the autocorrelation length of the fMRI
timeseries and the temporal blurring induced by the haemodynamic response function). This can
be avoided if care is taken to ensure that overlapping scans are discarded from the design (see
chapter 2), but it is a very important issue, and the user should still be careful to ensure that
cross-validation folds are sufficiently far apart in time (especially for LOBO cross-validation).
During this part of the model specification, it is also possible to select one or more operations
to apply to the data. Each of these operations is defined below:
1. Sample averaging (within blocks): constructs samples by computing the average of all
scans within each block or event for each subject and condition.

4.5. CROSS-VALIDATION

45

Figure 4.5: For each fold, specify which images/trials are part of the training and test set, or are
unused

Figure 4.6: Review cross-validation matrix

2. Sample averaging (within subjects): constructs samples by computing the average of
all scans within all blocks for each subject and condition.
3. Mean centre features using training data: subtract the voxel-wise mean from each
data vector.
4. Divide data vectors by their norm: scales each data vector (i.e. each example) to lie
on the unit hypersphere by dividing it by its Euclidean norm.
A crucial point to note is that all operations are embedded within the cross-validation structure
such that they are applied independently to training and testing sets. This prevents a very
common mistake in pattern recognition from occurring, whereby parameters are computed using
the whole data set prior to cross-validation. Observing a complete split between training and
testing sets during all phases of analysis ensures that accuracy measures are an appropriate
reflection of the true generalisation ability of the machine and are not biased because of improper
applications of preprocessing operations to the entire dataset.

46

CHAPTER 4. MODEL SPECIFICATION AND ESTIMATION

Other points to note include: (i) the order of operations is potentially important. For example,
subtracting the mean then dividing each data vector by its norm is not the same as performing
the operations the other way around. (ii) operations (1) and (2) have no effect if no design is
specified or for events with a length of one TR.
At a minimum, we recommend that features should be mean centered over scans during crossvalidation. In addition, for multiple kernel learning, we advise the user to normalize each kernel.
This will compensate for the fact that different kernels might be computed from examples/samples
with different number of features (e.g. different regions contain different numbers of voxels).
The different operations selected for a specific model can then be reviewed using the ‘Review
Kernel and CV’ (starting from version 2.0). The selected operations will be listed below the
kernels (‘Show kernel’).

4.6

Specify / Run model

Using the GUI, it is possible to either ‘Specify’ the model, or ‘Specify and Run’ the model. The
first option saves all the parameters of the model in the PRT structure. This information can
be found in PRT.model(m ).input, where m is the index of the model. The second saves all the
parameters of the model and then runs the model. In this case, the inputs, which include the
cross-validation matrix, the target values or labels, and the machine (e.g. binary SVM, Gaussian
Process, etc.), are fed to the estimation routines, which will then add to the PRT an output field
(PRT.model(m ).output) containing the estimated parameters, statistics, and other information
from the learning process.
In some cases (e.g. multiple models to run or models with nested CV and/or slower machines),
it would be desirable to estimate models later on (e.g. just before lunch break or at the end of
the day). The ‘Run model’ option allows to select multiple models and run them one at a time
automatically (Figure 4.7). The first thing that needs to be done using this window is to specify
which PRT we would like to work with. PRoNTo will then read the available models from this
structure and display the list of models on the left panel. These models can be selected (the
selected models will show on the right panel) by clicking each model individually or by clicking
the ‘Select all’ button in the middle of the panels. Finally, to estimate the model(s), one needs
only to click the bottom button ‘Run model(s)’.

Figure 4.7: Choose models to be estimated

It is useful to have a look at what is displayed in the Matlabcommand space when the model
is being estimated. Information such as the number of folds can help double-check that everything
is going as expected. Furthermore, if some options were specified (e.g. using a feature set with
multiple kernels) that are not available at the modelling step (e.g. to be modelled with SVM),
warnings will be displayed, as in Figure 4.8.
Important note: multiple kernel feature sets can be modelled using any machine. For
machines not supporting multiple kernels, these will first be added before the model is estimated.
This corresponds to concatenating the features from the different regions and/or modalities.

4.7. BATCH INTERFACE

47

Figure 4.8: Matlabworkspace displaying warnings when kernels are added

4.7

Batch interface

The batch module provides all the functionality provided in the user interface and allows complex
analyses to be scripted in advance. As noted, the batch module also provides functionality
not available in the user interface. The most important difference is that the batch module
allows customised Matlab functions to be used as prediction machines. This functionality allows
PRoNTo to be easily extended to allow many types of classification and regression algorithms not
provided under the current framework. This can be achieved by selecting ‘Custom machine’ under
the ‘Model Type’ heading. This allows a function name to be specified (i.e. any *.m function
in the Matlab search path). The behaviour of this custom machine can then be controlled
by a free-format argument string. See the developer documentation and the examples in the
machines/ subdirectory of the PRoNTo distribution for more information. Another important
difference between the batch and user interfaces is that mean centering data vectors across scans
is enabled by default in the batch. Also, flexible CV is available in the batch only in the form of
‘load a .mat’. This .mat must contain a variable called ‘CV’, specifying the CV matrix for the
selected trials/images.
An example of the batch window for model specification is provided in figure 4.9.
As displayed in Figure 4.9, the batch does not allow to specify and run the model directly.
Instead, the user had to add a ‘Run model’ module. The batch has the advantage of allowing to
perform permutations along model estimation (option available in the ‘Display results’ window
in the GUI). Furthermore, it is possible to save the predictions and the balanced accuracy (for
classification) for each permutation, to perform further statistical tests if needed. The ‘Run
model’ batch module is presented in Figure 4.10. Note: Dependencies on the model name are
available when performing a ‘Model specification’ followed by a ‘Run model’ module.

48

CHAPTER 4. MODEL SPECIFICATION AND ESTIMATION

Figure 4.9: Batch interface to specify a model

Figure 4.10: Batch interface to run a model

Chapter 5

Computing Feature and Region
Contributions
Contents

5.1

5.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

49

5.2

Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

49

5.3

Graphical user interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

5.4

matlabbatch interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

Introduction

The previous module allows the user to specify one or more models. These include the machine
to be used, the cross-validation scheme and the classification/regression problem. The estimation
of those models led to predictions on unseen/test data (in each fold), from which measures of
performance of the model can be derived.
In addition, as PRoNTo uses linear models it provides the option of recovering the model
weights in the original feature (voxel) space, and transforming the weights vector into an image,
or map. These maps contain at each voxel the corresponding weight of the linear model (together
defining the optimal decision function), and which related to how much this particular voxel
contributed to the classification/regression task in question. The weights can later be displayed
using the ‘Display results’ module (described below).
Furthermore, the MKL machine estimates contribution of each kernel to the final decision
function. This means that there will be one value per region of interest as defined by an atlas
and/or per modality (depending on how multiple kernels were built). Therefore, it is possible to
build maps at the region level, in addition to the maps at the voxel level. Regions/modalities can
then be ranked according to their contribution. Since L1-MKL is a sparse algorithm (i.e. only
some kernels will have a non-null contribution to the model), this eases model interpretation.

5.2

Methods

The output of the linear kernel models in PRoNTo include the coefficients of the dual representation, i.e. the coefficients of the training examples. These coefficients are then multiplied by the
training examples to obtain the model weights. The vector of model weights has the same dimensions of the original voxel space, and can therefore be converted to a 3D image. This computation
is done for each fold. The resulting 3D images for all folds are then assembled into a single 4D
NIFTI file with dimensions [3D x (number of folds + 1)], where 1 corresponds to an extra 3D
image with the averaged weights over all folds. The NIFTI file is saved in the same directory as
PRT.mat. In the case of multi-class classification, one image will be built for each class, the index
of the class being saved in the image name (e.g. image name c1.img). In the case of multiple
49

50

CHAPTER 5. COMPUTING FEATURE AND REGION CONTRIBUTIONS

modalities being considered as multiple kernels (i.e. not concatenated in samples), one image will
be built for each modality, the modality name being appended to the image name.
In addition, it is possible to build an image containing the contributions of each region of
interest as defined by an atlas (same dimensions as for the voxel weight images). Two options
are available:
• Contributions of kernels estimated through MKL: In this case, the contributions of each
region to the model are derived from the contributions of each kernel. This option is available
for MKL modelling of feature sets containing multiple kernels based on ROIs defined by an
atlas.
• Summarizing the weights according to ROIs: If a whole brain feature set was used, or if
the kernels were added to perform single kernel modelling (e.g. SVM, GP, KRR, RVR), it
is possible to select an atlas and summarize a posteriori the weights in each anatomically
defined ROI. The contribution of each region is then simply the sum of the absolute values
of the weights within that region, divided by the number of voxels in that region (see [17]
for details).
In both cases, the contribution of each region is divided by the total contribution of all regions.
The derived values can then be seen as percentages of contribution of each region to the decision
function. The contributions can be ranked, leading to a list of regions sorted by descending
contribution to the model. This list can also be computed if multiple modalities were built and
used in an MKL model. In this case, each modality has a contribution to the model, that can be
normalized and a sorted list can be derived.
Furthermore, for MKL models - which are sparse in the number of kernels contributing to the
model - a bar graph can be built, representing the number of kernels with a non-null contribution
to the model. The same graph bar will depict the contribution of each region to the decision
function in the case of summarized weights. In this case, the bar graph will not be sparse.

5.3

Graphical user interface

If the user wants to create images of the weights, using the GUI, the user first needs to click the
‘Compute weights’ button on the main PRoNTo window. This will launch the window shown in
Fig 5.1. To estimate the weights and create the weight maps the user needs to select a PRT.mat
file. The window is then divided in two panels: a ‘Feature weights’ panel and a ‘Atlas-based
weights’ panel. In the first panel, PRoNTo will show the list of available models, and the user can
choose one model for which to estimate the weights. If the selected model is the MKL modelling
of ROI-based kernels, the ‘Atlas-based weights’ panel will be automatically updated, and the
name of the selected atlas at the feature set step will appear. Otherwise, those fields will stay
blank. In the ‘Feature weights’ panel, it is also possible to define the name of the created image
file, which is saved in the same directory as PRT.mat. Alternatively, if left empty, PRoNTo
will name the file according to the model name, class (if multi-class machine) and/or modality
(if MKL on modalities). The ‘Atlas-based weights’ allows to estimate the contribution of each
anatomically defined region to the model. If the model refers to an MKL machine estimated
on kernels per region, the ‘Atlas name’ field will be filled automatically and the contribution of
each region is derived from the contribution of each kernel. If this is not the case (single kernel
machine or feature set), an atlas should be loaded (using the browse option). The weights will
then be summarized for each region a posteriori.

5.4

matlabbatch interface

The matlabbatch module to compute the weights has the same options as the GUI. One main
difference being that instead of listing the available models in a given PRT, it will ask for the
name (string) of the model to be used. As for estimating models, the name of the model should
be exactly the name given in ‘Specify model’. Another difference is that the batch allows to
compute weight images for each permutation. In this case, only the average across folds will be

5.4. MATLABBATCH INTERFACE

51

Figure 5.1: Weights computation GUI.

saved in a nifti. This potentially allows for statistical tests to be performed on weights and/or on
the derived ranking of the regions (for MKL modelling of anatomically defined ROIs).

Figure 5.2: Weights computation GUI.

52

CHAPTER 5. COMPUTING FEATURE AND REGION CONTRIBUTIONS

Chapter 6

Display Model Performance
Contents
6.1
6.2
6.3
6.4

6.5

6.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . .
Launching results display . . . . . . . . . . . . . . . . . .
The main results display window . . . . . . . . . . . . .
Analysing a machine’s performance graphically . . . . .
6.4.1
Classification . . . . . . . . . . . . . . . . . . . . .
6.4.2
Regression . . . . . . . . . . . . . . . . . . . . . .
6.4.3
Influence of the hyper-parameter on performance
Statistical analysis of a machine’s performance . . . . . .
6.5.1
Classification . . . . . . . . . . . . . . . . . . . . .
6.5.2
Regression . . . . . . . . . . . . . . . . . . . . . .
6.5.3
Permutation testing . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

53
53
54
54
55
56
57
58
59
59
61

Introduction

Once a machine (e.g. a classifier or a regression function) has been specified, its parameters have
been estimated over training data, and its performance has been evaluated over a testing set using
cross-validation, it is necessary to examine the outcome of the whole procedure in detail. The
results windows enables the user to see the model’s performance evaluated by different metrics.
Examining model output and parameters is helpful in diagnosing the potentially bad performance of a particular model. For example, if the machine cannot perform above chance, it could
be due to an inappropriate experimental paradigm, noisy data, insufficient amount of data, wrong
choice of features, or the wrong choice of machine. It is important to recognise that any of these
factors could cause the modelling to fail.
Model performance can be reviewed using the ‘Display Results’ GUI. Alternatively, all computed statistics are saved within the PRT structure, in the PRT.model(m).output.stats field,
with m, being the index of the model to review.

6.2

Launching results display

Make sure all previous steps have been performed (Data and Design, Chapter 2; Prepare feature
set, Chapter 3; Specify Model and Run Model, Chapter 4 ).
In the Review Options panel of the main PRoNTo window, press Display Results. At the
‘Select PRT.mat’ window, navigate to where your PRT.mat file is stored (using the left column),
and select it. The main results window will open and look as represented in Figure 6.1. In the
Model panel in the top-right corner, the list of models that have been successfully estimated will
appear. Note: there will be a ‘beep’ if one or more models were specified but not estimated (‘Run
model’) and their name will appear in the command window.
53

54

CHAPTER 6. DISPLAY MODEL PERFORMANCE

Figure 6.1: Initial state of the results display main window.

6.3

The main results display window

The window is divided into three panels; going clockwise from top left to bottom left, they are:
Plot : This panel displays the plots for the various analyses that can be performed on test
results. With the exception of the confusion matrix plot, these cannot be interacted with.
Model : This panel allows the user to select the model to visualize, whether to visualize a
particular fold or all folds at once, and which plot to produce.
Stats : The stats panel allows the user to visualize a variety of performance metrics (based on
the selected fold), including accuracy statistics for classifiers and MSE for regression models.
In addition, p-values for these metrics based on permutation tests can also be visualized.
To populate the ‘Plot’ panel, first click on a model in the Model selector, then on ‘all folds’
(or a particular fold) in the Fold selector, and finally on a plot in the Plot selector. The next
section details the plots available.
The window also comprises ‘Edit plot’, ‘Help’ and ‘Quit’ buttons. The ‘Edit plot’ button
exports the displayed plot in an extra window, such that it can be edited and easily saved. The
‘Help’ button provides information on each panel of the window (not as detailed as in this manual)
and the ‘Quit’ button closes the results window. In addition, the GUI menu comports a ‘Save
figure’ option (on the top left) that acts as a ‘printscreen’ of this window (with white background),
which can be saved for records/publications.

6.4

Analysing a machine’s performance graphically

Looking at a machine output’s graphically can often yield insights into the performance of the
machine. In PRoNTo, plots are different for classification and regression.

6.4. ANALYSING A MACHINE’S PERFORMANCE GRAPHICALLY

6.4.1

55

Classification

Predictions plot
A prediction plot displays, for a particular fold (y-axis), the output value of the machine’s decision
function for each test sample (x-axis, e.g., for a linear SVM, this could be wT xi + b, for a
probabilistic classifier this could be a posterior probability P (Ω = ω|xi )). The decision threshold
is displayed by a vertical line at the centre of the plot. A well-performing classifier will yield very
different function values for samples of different classes, i.e. samples from different classes will
fall on different sides of the decision threshold. The inspection, in each fold, of the overlap of
function values between classes, can help to identify which of the test blocks/subjects/conditions
is atypical with respect to the training set. This plot is available for binary classification.
On the plot, each class is represented by a different marker and color, and indicated in the
legend. Figure 6.2 shows an example predictions plot.

Figure 6.2: Example predictions plot for a two-class problem modelled by an SVM.

Receiver Operating Characteristic (ROC) plot
In two-class classification, there is always a trade-off between class 1 and class 2 errors. Indeed, a
classifier predicting class 1 regardless of input would have excellent accuracy on class 1, but bad
accuracy on class2. This is also known as the sensitivity / specificity trade-off. The ROC curve
is a graphical depiction of this trade-off, showing how one error rate varies as a function of the
other. An ideal classifier would have an ROC passing through the top-left corner. The area under
curve (AUC) is a summary measure of classifier performance, where higher is better (1 represents
perfect performance, 0.5 represents random performance). As with all summary measures, the
AUC is but one way of comparing performance of machines, and cannot be used alone to declare
a machine statistically significantly superior to another on a given dataset.
Figure 6.3 shows an example of such a plot (Haxby dataset, Faces versus Houses, 4-folds
nested CV, LOBO outer CV).
Histogram plot
The histogram plot is a smoothed density version of the predictions plot, showing how function
values are distributed. A good classifier would have minimal overlap between the densities. The
error rate of the classifier is proportional to the area of the overlap. The ROC curve can be
thought of as the result of sweeping a decision threshold over the range of functional values, and
recording the joint sensitivity/specificity values for each decision threshold setting. A typical
linear SVM would have a decision threshold at 0.
Figure 6.4 shows an example of such a plot for a binary classification (SPM EEG dataset,
Faces versus Scrambled, LOBO CV).

56

CHAPTER 6. DISPLAY MODEL PERFORMANCE

Figure 6.3: Example ROC curve for a two-class problem modelled by an MKL on ROIs.

Figure 6.4: Example function values histogram curve for a binary problem modelled by L1-MKL.

Confusion matrix plot
The confusion matrix shows counts of predicted class labels ω̂n = f (xn ) (in rows) versus true
class labels ωn (in columns). An ideal confusion matrix is diagonal: all predicted class labels
correspond to the truth. Off-diagonal elements represent errors. It is important to check that
none of the classes is “sacrificed” to gain accuracy in other classes - in other words, if all classes
are equally important to classify, no class should have more off-diagonal than on-diagonal entries.
Many summary statistics, including class accuracy, total accuracy, sensitivity, and specificity, can
be computed from the confusion matrix.
Figure 6.5 shows an example of a confusion matrix (Haxby dataset, Faces versus Houses versus
Scissors, LOBO CV).

6.4.2

Regression

Predictions (scatter)
This plot represents the predicted values (x-axis) against the real values or targets (y-axis). A
perfect correspondence between targets and predictions would be represented by a diagonal on
this plot. Figure 6.6 displays such a plot for the prediction of age from sMRI images (IXI dataset,
15 images acquired across 3 centres).

6.4. ANALYSING A MACHINE’S PERFORMANCE GRAPHICALLY

57

Figure 6.5: Example confusion matrix for all folds of a three-class problem modelled by GP.

Figure 6.6: Example of scatter prediction plot on 15 data points modelled by KRR.

Predictions (bar)
This plot displays, for each image/subject, the target and the prediction in bar plots. An example
plot is displayed in Figure 6.7 for the same model.
Predictions (line)
This plot displays, for each fold, the target and the prediction, each in line plots. An example
plot is displayed in Figure 6.8 for the same model.

6.4.3

Influence of the hyper-parameter on performance

This plot will be present in the list if hyper-parameter optimization was performed. When
displaying the average across folds, for each value of the hyper-parameter, it displays the average
model performance (balanced accuracy for classification and MSE for regression, line on the plot)
across nested folds, with an error bar representing the standard deviation of model performance.
The frequency of selection of a hyper-parameter value (i.e. the number of times this value was
returned as ‘optimal’ to the outer CV fold) is represented with a gray bar plot on the right-side
y-axis. An example of such a plot is displayed in Figure 6.9 for the optimization of the softmargin parameter in L1-MKL (Haxby dataset, Faces versus Houses, 4-folds nested CV, LOBO
outer CV). When selecting a specific fold, this plot displays the model performance for each

58

CHAPTER 6. DISPLAY MODEL PERFORMANCE

Figure 6.7: Example of bar prediction plot on 15 data points modelled by KRR.

Figure 6.8: Example of line prediction plot on 15 data points modelled by KRR.

value of the hyper-parameter, and represents the optimal value (i.e. the one leading to highest
performance) in red.

6.5

Statistical analysis of a machine’s performance

One of the main questions to ask of a model is how precise its predictions are. In regression,
goodness-of-fit is often assessed via mean squared error and coefficient of determination (R2 ). In
classification, a common practice is to compute prediction accuracy, both for each class and for
all test data. Once a specific performance metric has been obtained, it is also possible to obtain
a p-value for the metric, reflecting how certain we are that the result is not due to chance.
The statistics table gives a summary of the model’s performance. Model performance is estimated differently for classification and for regression. In PRoNTo, classification performance
is assessed using total accuracy (TA), balanced accuracy (BA), class accuracies (CA) and class
predictive values. For regression, the goodness-of-fit is assessed based on the coefficient of determination (R2 ), the mean squared error (MSE), the normalized mean squared error, and the
correlation between the targets and the predictions.

6.5. STATISTICAL ANALYSIS OF A MACHINE’S PERFORMANCE

59

Figure 6.9: Example performance curve depending on the hyper-parameter value with frequency
of selection of each hyper-parameter.

6.5.1

Classification

The accuracy acc is the total number of correctly classified test samples divided by the total
number of test samples N , irrespective of class. The accuracy is exactly equivalent to
acc = 1 −

1 X
l01 (yn , f (xn )),
N n

(6.1)

where l01 (yn , f (xn )) is a 0-1 loss function that counts each classification error as costing 1 and
each classification success as costing 0:

l01 (yn , f (xn )) =

0
1

yn = f (xn )
yn 6= f (xn )

(6.2)

Balanced accuracy takes the number of samples in each class into account, and gives equal
weight to the accuracies obtained on test samples of each class. In other words, the class-specific
accuracy is computed by restricting the sum of equation 6.1 to be taken over C disjoint subsets
of the whole testing data, where each subset contains only test samples from one class. This
produces a set of class-specific accuracies {acc1 , . . . , accC }, from which the balanced accuracy
can be computed as
accbal =

1 X
accc .
C

(6.3)

Balanced accuracy is the measure of choice when there is class imbalance (one class, called
the majority class, has much more data than others).
The stats panel also gives the class accuracies {acc1 , . . . , accC }, useful to check whether the
model favours some classes over others. If class 1 represents control subjects, and class 2 represents
patients, then class 1 accuracy is equivalent to specificity, and class 2 accuracy is equivalent to
sensitivity. In the same way, the figure displays class positive predictive value, which represents
the number of false positives for each class. An example of classification stats is displayed in
Figure 6.10.

6.5.2

Regression

As previously mentioned, model performance for regression is assessed by the correlation between
the predictions and the targets (linear correlation), the coefficient of determination (R2 ), the mean
square error (MSE), and the normalized MSE. An example of such stats window is displayed in
Figure 6.11.

60

CHAPTER 6. DISPLAY MODEL PERFORMANCE

Figure 6.10: Example statistics for all folds of a two-class problem modelled by an L1-MKL.

Figure 6.11: Example statistics for all folds of a KRR modelling on 15 data points.

The mean-squared error is calculated as:
M SE =

1 X
(yn − f (xn ))2
N n

(6.4)

This is the standard measure when assessing goodness-of-fit for regression models. Since the
magnitude of the MSE depends on the scale of y, we also calculate the Normalised Mean Square
Error, ‘Norm. MSE’:
M SE
N orm.M SE =
(6.5)
(ymax − ymin )
where we divide the MSE by the range of the targets over the data. This gives a scale invari-

6.5. STATISTICAL ANALYSIS OF A MACHINE’S PERFORMANCE

61

ant measure of prediction accuracy. In addition, the correlation coefficient of the targets and
predictions are determined:
P
n (yn − µy )(f (xn ) − µf )
CORR = P
(6.6)
1
P
{ n (yn − µy )2 n (f (xn ) − µf )2 } 2
in which µy and µf are the sample means of the targets and predictions respectively. The resulting
measure −1 < CORR < 1 provides a measure of the strength of the linear dependence between the
targets and the predicted targets, with values close to zero indicating no relationship, values close
to 1 indicating a positive relationship, and values close to -1 indicating a negative relationship.
Values of CORR less than zero would imply that the model has performed poorly, as this would
mean that targets with large values tend to be given smaller predicted values than targets with
small values. However, it should be remembered that a large positive value of CORR does not
necessarily mean that the model is giving accurate predictions, since a global scaling and shifting
of the predictions gives the same value for CORR. We would therefore recommend examining
both CORR and M SE, as well as the scatterplots, to verify that the model is performing well.
For completeness, the statistics table also include the ‘coefficient of determination’ R2 , which is
given by
(6.7)
R2 = CORR2

6.5.3

Permutation testing

Much of statistical theory and machine learning theory rests on the assumption that the data
is IID (independently and identically distributed). However, in functional neuroimaging this assumption is often not met, due to e.g. within-run correlations and haemodynamic effects. Therefore, classical estimates of confidence intervals (such as the binomial confidence interval) may not
always be appropriate. Permutation testing is a non-parametric procedure that allows to obtain
meaningful confidence intervals and p-values in this case. Because it requires retraining the model
a number of times, which can be costly in computation time, this is not done by default. After
filling in the repetitions field with a number of repetitions R, pressing the Permutation test
button will estimate the model for the specified number of times with permuted labels/targets,
and produce a p-value for performance statistics (see Figure 6.10). The smallest increment in
p-value is equal to 1/R (e.g. 20 permutations gives you increments of 0.05), with a minimum
value of 1/R (i.e. running 10 permutations will never lead to statistically significant result at
the commonly used threshold of p < 0.05). Usually, we would recommend computing several
hundreds to a thousand permutations.
For both classification and regression models, the p-value associated with each performance
measure can be estimated using permutations. Until they have been estimated, ‘N.A.’ will be
displayed (standing for ‘not available’).
Important note: This step is essential to assess model performance! It is not methodologically sound to simply assume that the chance level is close to 50% and that any balanced accuracy
higher than that threshold is significant. Please report p-values as computed from permutations
along with model performance.

62

CHAPTER 6. DISPLAY MODEL PERFORMANCE

Chapter 7

Display voxel and region
contribution
Contents
7.1
7.2

7.1

Introduction . . . . . . . . . .
Displaying weights . . . . . .
7.2.1
Select image to display
7.2.2
Weights map . . . . . .
7.2.3
Anatomical image . . .
7.2.4
Additional plots . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

63
64
64
65
65
66

Introduction

Another important aspect of pattern recognition modelling when applied to neuroimaging is
trying to interpret the models’ parameters or weights. Some brain areas are probably more
informative about class membership/regression targets than others. For example, in a visual
task, we would expect discriminative information in the occipital lobe. This can be seen as
information mapping, and it can be helpful to evaluate a specific model - if the discriminative
weight of a machine is concentrated in the eyes, for example, it is important to correct the mask
used in the analysis to exclude them. In the case of linear kernels, the classifier/regression weight
vector is a linear combination or weighted average of the training examples, and can be plotted
as an image representing a weight map. The weight map is therefore a spatial representation of
the decision function, i.e. every voxel within the mask contributes with a certain weight to the
decision function. Pattern recognition models (classifiers or regression functions) are multivariate,
i.e. they take into account correlations in the data. Since the discrimination or prediction is based
on the whole brain pattern, rather than on individual regions or voxels, all voxels contribute to
the classification or regression and no conclusions should be drawn about a particular subset of
voxels in isolation.
Starting from PRoNTo v2.0, it is possible to derive weights at the region level (as anatomically defined by an atlas, from MKL or from summarizing the weights). This window allows to
display maps of voxel and of region contribution. Furthermore, the region contributions can be
ranked in descending order, yielding a sorted list of regions according to their contribution to the
classification/regression model. We hope this will help the interpretation of model parameters in
terms of cognitive neuroscience.
Important note: The implemented version of MKL ([14]) is sparse in the kernel combination.
This means that only a few regions/modalities will contribute to the model. However, this
selection of regions/modalities might depend on the dataset, and small variations in the dataset (as
induced by cross-validation) might lead to different subsets of regions/modalities begin selected.
Therefore, care should be taken when reporting selected regions/modalities and each fold should
be looked at separately. We also provide a quantification of the variability across folds of the
63

64

CHAPTER 7. DISPLAY VOXEL AND REGION CONTRIBUTION

ranking of the regions/modalities (‘Expected Ranking’, see further) to provide some insights on
this issue.

7.2

Displaying weights

To launch the ‘Display weights’ window, make sure that weight maps have been computed for at
least one model (Compute Weights, Chapter 5).
In the Review Options panel, press Display weights. At the ‘Select PRT.mat’ window,
navigate to where your PRT.mat file is stored (using the left column), and select it in the right
column. The display window then opens (Figure 7.1).

Figure 7.1: Display weights main window after selection of PRT.mat.
The window is divided into four panels; going from top left to bottom left, they are:
Display : This panel allows the user to choose which model and image to display, as well as
whether to display the voxel weights or the region contributions.
Weights map : This displays three projections of the selected weight map and allows to navigate
it.
Anatomical img : If an anatomical image has been loaded, this will display three projections,
and the cross-hair will be synchronised with the weight map.
Additional plots : The blank area at the bottom of the window will display additional information about the model parameters, such as a sorted list of the regions according to their
contribution (if weights per region were computed, in the form of a table) and a bar plot of
the relative region contribution. If MKL modelling was performed based on multiple modalities, the same table and bar plot will display the relative contributions of each modality to
the decision function.

7.2.1

Select image to display

The Display panel shows the models for which weights were computed and weight images were
found in the same folder as the PRT. For each model, the list of images available is displayed

7.2. DISPLAYING WEIGHTS

65

in the Weights pop-down list. Typically, one image will be created for a binary comparison or
regression with only one modality or multiple modalities concatenated in samples (e.g. multiple
runs). On the other hand, multiclass classification models will return one image per class (with
the index of the class appended to the name of the image). In the same way, multiple modalities
used in multiple kernels will lead to the building of number of modalities images. For each image,
the weight map can be displayed for each fold or for their average. In PRoNTo v2.0, it is possible
to display the contributions of each voxel (weights per voxel) or of each region (weights per
region, if previously computed).
Note: the weight images (per voxel and per region) are automatically detected in the list of
files in the PRT folder according to the name specified in the ‘Compute weights’ step (Chapter
5). Modifying the image name afterwards or moving the images might lead to warning messages
and the images will not be listed in the GUI.
To display a weight image, select a model, an image and a fold. If only weights per voxel were
estimated, the window will look similar to Figure 7.2.

Figure 7.2: Displaying weight image for class 1 of a three-class GP model (Haxby dataset).

7.2.2

Weights map

The weight map is displayed with a cross-hair and a colorbar. The colorbar indicates the relative
importance of the voxel in the decision function. This value is also indicated in the intensity
field of the Anatomical img panel. Note that all voxels in the mask contribute to the decision
function, since the analysis is multivariate. Contrary to common practice in Statistical Parametric
Mapping, which is a mass-univariate approach, it does not make sense to isolate part of the pattern
and report only on the peaks of the distribution of the decision function’s weight map, unless they
have perfectly null contribution (as might happen with sparse models such as L1-MKL modelling).
Below the displayed image, an edit box and a ‘browse’ ([...]) button also allow to load a
weight image (.img) that is not linked to a model (e.g. from a previous PRoNTo version).

7.2.3

Anatomical image

By clicking on the [...] button next to the Load anatomical img field, a dialogue opens that
allows you to select an anatomical images .img file that was co-registered with the data images.

66

CHAPTER 7. DISPLAY VOXEL AND REGION CONTRIBUTION

In this panel, the cross-hair position is displayed in voxels and in mm. It can also be reset
to the origin of the image. For each position, the corresponding voxel weight is displayed in the
‘intensity’ field.

7.2.4

Additional plots

Additional information will be displayed in two main cases:
• Multiple Kernel Learning modelling: MKL modelling, based on modalities or on
regions as defined by an atlas, will provide weights at two levels: the kernel level and the
voxel level. The kernel contributions, which sum to 1, can then be ranked in descending
order.
• Summarizing weights per region: in PRoNTo v2.0, weights can be summarized within
regions of interest as defined by an atlas (user-specified). For each region, a normalized
contribution can be defined, and those contributions can then be ranked in descending
order.
In both cases, the region/modality contributions will be displayed in a table as well as in a
bar plot, for each fold and for their average (according to the selected fold in the Display panel).
An example is displayed in Figure 7.3, for weight summarization after GP modelling.

Figure 7.3: Displaying ROI contributions for class 1 of a three-class GP model (Haxby dataset).

In the case where kernels were built both at the modality and at the region level (i.e. multiple
modalities with each multiple regions as defined by an atlas), two tables will be displayed (one for
regions, one for modalities). The table for modalities will sum the contributions of each region
within that modality.
Sorted table of region/modality contributions
The displayed table comprises one row per region and 5 columns (for an example on ROIs, as
displayed in Figure 7.3):
• Index of the ROI: The first column displays the ranking of the region of interest in the
selected fold, according to its contribution.

7.2. DISPLAYING WEIGHTS

67

• ROI Label: When using the atlas provided in your PRoNTo folder/atlas, the labels of
each region will be loaded automatically from a .mat, stored alongside the atlas. If using
another atlas, the labels can be loaded through the ‘Load Labels’ button. In this case, the
user should select a .mat comprising a cell array of size ( number of regions,1), with the
label for each region in the corresponding cell (in characters). The cell array should be
saved under the variable name ‘ROI names’. Otherwise, generic names will be used (e.g.
ROI 1).
• ROI weight: The (normalized) contribution of each region is displayed in the third column
(in %). The rows of the table are sorted in descending order according to this value.
• ROI size: This column displays the size of the ROI in voxels. This gives indications on
the overlap between the atlas and the data.
• Expected Ranking: This measure reflects how stable the ranking of the region is across
folds. It is computed from the ranking in each fold (see [17] for details), and is therefore
the same, whether the user is displaying fold 1, or the average of all folds. If the Expected
Ranking (ER) is close to the ranking in the selected fold, then it reflects that this region has
a similar ranking across folds. On the contrary, if the ER is quite different from the ranking
shown for the selected fold, this means that the ranking might be variable across folds. This
variability can come from the fact that the region did not have the same contribution to
slightly different datasets. It might also happen that it is not selected at all in some folds
(as can happen with L1-MKL since it will not select kernels with correlated information).
When selecting a specific region label in the table, the weight map will only display colored
voxel or region weights (according to which plot was selected) for this region, the rest of the
image being in grey scale. This allows e.g. to look closely at the voxel weights within a region
that highly contributes to the selected model and fold (Figure 7.4).

Figure 7.4: Displaying fusiform weights for binary MKL model (Haxby data).

Finally, the table can be exported as a text file using the ‘Export Table’ button.

68

CHAPTER 7. DISPLAY VOXEL AND REGION CONTRIBUTION

Bar plot of contributions
The bar plot displays the third column of the table, i.e. the contribution of each ROI or modality
to the decision function. The x-axis represents the index of the ROI in the table (i.e. first
column of the table), in the selected fold, while the y-axis displays the contribution of each
region/modality. The bar graph provides insights on how sparse or dense the region/modality
contributions are.

Part II

Batch interfaces

69

Chapter 8

Data & Design

Specify the data and design for each group (minimum one group).

8.1

Directory

Select a directory where the PRT.mat file containing the specified design and data matrix will be
written.

8.2

Groups

Add data and design for one group. Click ’new’ or ’repeat’ to add another group.

8.2.1

Group

Specify data and design for the group.
Name
Name of the group. Example: ’Controls’.
Select by
Depending on the type of data at hand, you may have many images (scans) per subject, such as
a fMRI time series, or you may have many subjects with only one or a small number of images
(scans) per subject , such as PET images. If you have many scans per subject select the option
’subjects’. If you have one scan for many subjects select the option ’scans’.
Subjects Add subjects/scans.
Subject Add new modality for this subject.
Modality Add new modality.
Name Name of modality. Example: ’BOLD’. The names should be consistent accross subjects/groups and the same names specified in the masks.
Interscan interval Specify interscan interval (TR). The units should be seconds.
Scans Select scans (images) for this modality. They must all have the same image dimensions,
orientation, voxel size etc.
Data & Design Specify data and design.
Load SPM.mat Load design from SPM.mat (if you have previously specified the experimental
design with SPM).
71

72

CHAPTER 8. DATA & DESIGN

Specify design Specify design: scans (data), onsets and durations.
Units for design The onsets of events or blocks can be specified in either scans or seconds.
Conditions Specify conditions. You are allowed to combine both event- and epoch-related
responses in the same model and/or regressor. Any number of condition (event or epoch) types
can be specified. Epoch and event-related responses are modeled in exactly the same way by
specifying their onsets [in terms of onset times] and their durations. Events are specified with
a duration of 0. If you enter a single number for the durations it will be assumed that all
trials conform to this duration.For factorial designs, one can later associate these experimental
conditions with the appropriate levels of experimental factors.
Condition Specify condition: name, onsets and duration.
Name Name of condition (alphanumeric strings only).
Onsets Specify a vector of onset times for this condition type.
Durations Specify the event durations. Epoch and event-related responses are modeled in
exactly the same way but by specifying their different durations. Events are specified with a
duration of 0. If you enter a single number for the durations it will be assumed that all trials
conform to this duration. If you have multiple different durations, then the number must match
the number of onset times.
Multiple conditions Select the *.mat file containing details of your multiple experimental
conditions.
If you have multiple conditions then entering the details a condition at a time is very inefficient.
This option can be used to load all the required information in one go. You will first need to
create a *.mat file containing the relevant information.
This *.mat file must include the following cell arrays (each 1 x n): names, onsets and durations. eg. names=cell(1,5), onsets=cell(1,5), durations=cell(1,5), then names2=’SSent-DSpeak’,
onsets2=[3 5 19 222], durations2=[0 0 0 0], contain the required details of the second condition.
These cell arrays may be made available by your stimulus delivery program, eg. COGENT. The
duration vectors can contain a single entry if the durations are identical for all events.
Time and Parametric effects can also be included. For time modulation include a cell array
(1 x n) called tmod. It should have a have a single number in each cell. Unused cells may contain
either a 0 or be left empty. The number specifies the order of time modulation from 0 = No Time
Modulation to 6 = 6th Order Time Modulation. eg. tmod3 = 1, modulates the 3rd condition by
a linear time effect.
For parametric modulation include a structure array, which is up to 1 x n in size, called pmod.
n must be less than or equal to the number of cells in the names/onsets/durations cell arrays.
The structure array pmod must have the fields: name, param and poly. Each of these fields is in
turn a cell array to allow the inclusion of one or more parametric effects per column of the design.
The field name must be a cell array containing strings. The field param is a cell array containing
a vector of parameters. Remember each parameter must be the same length as its corresponding
onsets vector. The field poly is a cell array (for consistency) with each cell containing a single
number specifying the order of the polynomial expansion from 1 to 6.
Note that each condition is assigned its corresponding entry in the structure array (condition
1 parametric modulators are in pmod(1), condition 2 parametric modulators are in pmod(2), etc.
Within a condition multiple parametric modulators are accessed via each fields cell arrays. So for
condition 1, parametric modulator 1 would be defined in pmod(1).name1, pmod(1).param1, and
pmod(1).poly1. A second parametric modulator for condition 1 would be defined as pmod(1).name2,
pmod(1).param2 and pmod(1).poly2. If there was also a parametric modulator for condition
2, then remember the first modulator for that condition is in cell array 1: pmod(2).name1,
pmod(2).param1, and pmod(2).poly1. If some, but not all conditions are parametrically modulated, then the non-modulated indices in the pmod structure can be left blank. For example, if
conditions 1 and 3 but not condition 2 are modulated, then specify pmod(1) and pmod(3). Similarly, if conditions 1 and 2 are modulated but there are 3 conditions overall, it is only necessary
for pmod to be a 1 x 2 structure array.
EXAMPLE:
Make an empty pmod structure:
pmod = struct(’name’,”,’param’,,’poly’,);
Specify one parametric regressor for the first condition:
pmod(1).name1 = ’regressor1’;

8.3. MASKS

73

pmod(1).param1 = [1 2 4 5 6];
pmod(1).poly1 = 1;
Specify 2 parametric regressors for the second condition:
pmod(2).name1 = ’regressor2-1’;
pmod(2).param1 = [1 3 5 7];
pmod(2).poly1 = 1;
pmod(2).name2 = ’regressor2-2’;
pmod(2).param2 = [2 4 6 8 10];
pmod(2).poly2 = 1;
The parametric modulator should be mean corrected if appropriate. Unused structure entries
should have all fields left empty.
No design Do not specify design. This option can be used for modalities (e.g. structural
scans) that do not have an experimental design.
Scans Depending on the type of data at hand, you may have many images (scans) per subject,
such as a fMRI time series, or you may have many subjects with only one or a small number of
images (scans) per subject, such as PET images. Select this option if you have many subjects
per modality to spatially normalise, but there is one or a small number of scans for each subject.
This is a faster option with less information to specify than the ’select by subjects’ option. Both
options create the same ’PRT.mat’ but ’select by scans’ is optimised for modalities with no design.
Modality Specify modality, such as name and data.
Name Name of modality. Example: ’BOLD’. The names should be consistent accross subjects/groups and the same names specified in the masks.
Files Select scans (images) for this modality. They must all have the same image dimensions,
orientation, voxel size etc.
Regression targets (per scans) Enter one regression target per scan. or enter the name
of a variable. This variable should be a vector [Nscans x 1], where Nscans is the number of
scans/images.
Covariates Select a .mat file containing your covariates (i.e. any other data/information you
would like to include in your design). This file should contain a variable ’R’ with a matrix of
covariates. On covariate per image is expected.

8.3

Masks

Select first-level (pre-processing) mask for each modality. The name of the modalities should be
the same as the ones entered for subjects/scans.

8.3.1

Modality

Specify name of modality and file for each mask. The name should be consistent with the names
chosen for the modalities (subjects/scans).
Name
Name of modality. Example: ’BOLD’. The names should be consistent accross subjects/groups
and the same names specified in the masks.
File
Select one first-level mask (image) for each modality. This mask is used to optimise the prepare
data step. In ’specify model’ there is an option to enter a second-level mask, which might be used
to select only a few areas of the brain for subsequent analyses.

74

8.4

CHAPTER 8. DATA & DESIGN

fMRI Des

fMRI design specific parameters, HRF overlap and delay.

8.4.1

HRF overlap

If using fMRI data please specify the width of the hemodynamic response function (HRF). This
will be used to calculate the overlap between events. Leave as 0 for other modalities (other than
fMRI).

8.4.2

HRF delay

If using fMRI data please specify the delay of the hemodynamic response function (HRF). This
will be used to calculate the overlap between events. Leave as 0 for other modalities (other than
fMRI).

8.5

Review

Choose ’Yes’ if you would like to review your data and design in a separate window.

Chapter 9

Feature set/Kernel

Compute feature set according to the design specified

9.1

Load PRT.mat

Select data/design structure file (PRT.mat).

9.2

Feature/kernel name

Target name for kernel matrix. This should containonly alphanumerical characters or underscores
( ).

9.3

Modalities

Add modalities

9.3.1

Modality

Specify modality, such as name and data.
Modality name
Name of modality. Example: ’BOLD’. Must match design specification
Scans / Conditions
Which task conditions do you want to include in the kernel matrix? Select conditions: select
specific conditions from the timeseries. All conditions: include all conditions extracted from the
timeseries. All scans: include all scans for each subject. This may be used for modalities with
only one scan per subject (e.g. PET), if you want to include all scans from an fMRI timeseries
(assumes you have not already detrended the timeseries and extracted task components)
All scans No design specified. This option can be used for modalities (e.g. structural scans)
that do not have an experimental design or for an fMRI designwhere you want to include all scans
in the timeseries
All Conditions

Include all conditions in this kernel matrix
75

76

CHAPTER 9. FEATURE SET/KERNEL

Voxels to include
Specify which voxels from the current modality you would like to include
All voxels Use all voxels in the design mask for this modality
Specify mask file Select a mask for the selected modality.
Detrend
Type of temporal detrending to apply
None Do not detrend the data
Polynomial detrend
trend)

Perform a voxel-wise polynomial detrend on the data (1 is linear de-

Order Enter the order for polynomial detrend (1 is linear detrend)
Discrete cosine transform Use a discrete cosine basis set to detrend the data.
Cutoff of high-pass filter (second)
(same as SPM)

The default high-pass filter cutoff is 128 seconds

Scale input scans
Do you want to scale the input scans to have a fixed mean (i.e. grand mean scaling)?
No scaling

Do not scale the input scans

Specify from *.mat Specify a mat file containing the scaling parameters for each modality.
Use atlas to build ROI specific kernels
Select an atlas file to build one kernel per ROI. The AAL atlas (named ’aal 79x91x69.img’) is
available in the ’atlas’ subdirectory of PRoNTo

9.4

Use one kernel per modality

Select ”Yes” to use one kernel per modality.

Chapter 10

Specify model

Construct model according to design specified

10.1

Load PRT.mat

Select data/design structure file (PRT.mat).

10.2

Model name

Name for model

10.3

Use kernels

Are the data for this model in the form of kernels/basis functions? If ’No’ is selected, it is assumed
the data are in the form of feature matrices

10.4

Feature sets

Enter the name of a feature set to include in this model. This can be kernel or a feature matrix.

10.5

Model Type

Select which kind of predictive model is to be used.

10.5.1

Classification

Specify classes and machine for classification.
Classes
Specify which elements belong to this class. Click ’new’ or ’repeat’ to add another class.
Class

Specify which groups, modalities, subjects and conditions should be included in this class

Name

Name for this class, e.g. ’controls’
77

78

CHAPTER 10. SPECIFY MODEL

Groups Add one group to this class. Click ’new’ or ’repeat’ to add another group.
Group Specify data and design for the group.
Group name Name of the group to include. Must exist in PRT.mat
Subjects Subject numbers to be included in this class. Note that individual numbers (e.g.
1), or a range of numbers (e.g. 3:5) can be entered
Conditions / Scans Which task conditions do you want to include? Select conditions: select
specific conditions from the timeseries. All conditions: include all conditions extracted from the
timeseries. All scans: include all scans for each subject. This may be used for modalities with
only one scan per subject (e.g. PET), if you want to include all scans from an fMRI timeseries
(assumes you have not already detrended the timeseries and extracted task components)
Specify Conditions Specify the name of conditions to be included
Condition Specify condition:.
Name Name of condition to include.
All Conditions Include all conditions in this model
All scans No design specified. This option can be used for modalities (e.g. structural scans)
that do not have an experimental design or for an fMRI designwhere you want to include all scans
in the timeseries
Machine
Choose a prediction machine for this model
SVM Classification

Binary support vector machine.

Optimize hyper-parameter Whether to optimize C, the SVM hyper-parameter, or not. If
Yes, than provide a range of possible values for C, in the form min:step:max. Examples: 10.ˆ[-2:5]
or 1:100:1000 or 0.01 0.1 1 10 100. If not, a default value will be used (C=1).
Soft-margin hyper-parameter Value(s) for prt machine svm bin: soft-margin C. Examples: 10.ˆ[-2:5] or 1:100:1000 or 0.01 0.1 1 10 100.
Cross-validation type for hyper-parameter optimization Choose the type of crossvalidation to be used
Leave one subject out Leave a single subject out each cross-validation iteration
k-folds CV on subjects k-partitioning of subjects at each cross-validation iteration
k Number of folds/partitions for CV. To create a 50
Leave one subject per group out Leave out a single subject from each group at a time.
Appropriate for repeated measures or paired samples designs.
k-folds CV on subjects per group K-partitioning of subjects from each group at a time.
Appropriate for repeated measures or paired samples designs.
k Number of folds/partitions for CV. To create a 50
Leave one block out Leave out a single block or event from each subject each iteration.
Appropriate for single subject designs.
k-folds CV on blocks k-partitioning on blocks or events from each subject each iteration.
Appropriate for single subject designs.
k Number of folds/partitions for CV. To create a 50
Leave one run/session out Leave out a single run (modality) from each subject each
iteration. Appropriate for single subject designs with multiple runs/sessions.
Gaussian Process Classification Gaussian Process Classification
Arguments
Multiclass GPC
Arguments

Arguments for prt machine gpml
Multiclass GPC
Arguments for prt machine gpclap

10.5. MODEL TYPE

79

L1 Multi-Kernel Learning Multi-Kernel Learning. Choose only if multiple kernels
were built during the feature set construction (either multiple modalities or ROIs).
It is strongly advised to ”normalize” the kernels (in ”operations”).
Optimize hyper-parameter Whether to optimize C, the SVM hyper-parameter, or not. If
Yes, than provide a range of possible values for C, in the form min:step:max. Examples: 10.ˆ[-2:5]
or 1:100:1000 or 0.01 0.1 1 10 100. If not, a default value will be used (C=1).
Arguments Arguments for prt machine sMKL cla (same as for SVM)Examples: 10.ˆ[-2:5]
or 1:100:1000 or 0.01 0.1 1 10 100.
Cross-validation type for hyper-parameter optimization Choose the type of crossvalidation to be used
Leave one subject out Leave a single subject out each cross-validation iteration
k-folds CV on subjects k-partitioning of subjects at each cross-validation iteration
k Number of folds/partitions for CV. To create a 50
Leave one subject per group out Leave out a single subject from each group at a time.
Appropriate for repeated measures or paired samples designs.
k-folds CV on subjects per group K-partitioning of subjects from each group at a time.
Appropriate for repeated measures or paired samples designs.
k Number of folds/partitions for CV. To create a 50
Leave one block out Leave out a single block or event from each subject each iteration.
Appropriate for single subject designs.
k-folds CV on blocks k-partitioning on blocks or events from each subject each iteration.
Appropriate for single subject designs.
k Number of folds/partitions for CV. To create a 50
Leave one run/session out Leave out a single run (modality) from each subject each
iteration. Appropriate for single subject designs with multiple runs/sessions.
Custom Load a cross-validation matrix comprising a CV variable
Custom machine
Function

Choose a function that will perform prediction.

Arguments

10.5.2

Choose another prediction machine

Arguments for prediction machine.

Regression

Add group data and machine for regression.
Groups
Add one group to this regression model. Click ’new’ or ’repeat’ to add another group.
Group Specify data and design for the group.
Group name

Name of the group to include. Must exist in PRT.mat

Subjects Subject numbers to be included in this class. Note that individual numbers (e.g.
1), or a range of numbers (e.g. 3:5) can be entered
Machine
Choose a prediction machine for this model
Kernel Ridge Regression Kernel Ridge Regression.

80

CHAPTER 10. SPECIFY MODEL

Optimize hyper-parameter Whether to optimize K, the KRR hyper-parameter, or not. If
Yes, than provide a range of possible values for K, in the form min:step:max. Examples: 10.ˆ[-2:5]
or 1:100:1000 or 0.01 0.1 1 10 100. If not, a default value will be used.
Regularization
0.01 0.1 1 10 100.

Regularization for prt machine krr. Examples: 10.ˆ[-2:5] or 1:100:1000 or

Cross-validation type for hyper-parameter optimization Choose the type of crossvalidation to be used
Leave one subject out Leave a single subject out each cross-validation iteration
k-folds CV on subjects k-partitioning of subjects at each cross-validation iteration
k Number of folds/partitions for CV. To create a 50
Leave one subject per group out Leave out a single subject from each group at a time.
Appropriate for repeated measures or paired samples designs.
k-folds CV on subjects per group K-partitioning of subjects from each group at a time.
Appropriate for repeated measures or paired samples designs.
k Number of folds/partitions for CV. To create a 50
Leave one block out Leave out a single block or event from each subject each iteration.
Appropriate for single subject designs.
k-folds CV on blocks k-partitioning on blocks or events from each subject each iteration.
Appropriate for single subject designs.
k Number of folds/partitions for CV. To create a 50
Leave one run/session out Leave out a single run (modality) from each subject each
iteration. Appropriate for single subject designs with multiple runs/sessions.
Custom Load a cross-validation matrix comprising a CV variable
Relevance Vector Regression Relevance Vector Regression. Tipping, Michael E.; Smola,
Alex (2001).
”Sparse Bayesian Learning and the Relevance Vector Machine”. Journal of Machine Learning
Research 1: 211?244.
Gaussian Process Regression Gaussian Process Regression
Arguments

Arguments for prt machine gpr

Multi-Kernel Regression

Multi-Kernel Regression

Optimize hyper-parameter Whether to optimize C, the MKL hyper-parameter, or not. If
Yes, than provide a range of possible values for C, in the form min:step:max. Examples: 10.ˆ[-2:5]
or 1:100:1000 or 0.01 0.1 1 10 100. If not, a default value will be used (C=1).
Arguments

Arguments for prt machine sMKL reg

Cross-validation type for hyper-parameter optimization Choose the type of crossvalidation to be used
Leave one subject out Leave a single subject out each cross-validation iteration
k-folds CV on subjects k-partitioning of subjects at each cross-validation iteration
k Number of folds/partitions for CV. To create a 50
Leave one subject per group out Leave out a single subject from each group at a time.
Appropriate for repeated measures or paired samples designs.
k-folds CV on subjects per group K-partitioning of subjects from each group at a time.
Appropriate for repeated measures or paired samples designs.
k Number of folds/partitions for CV. To create a 50
Leave one block out Leave out a single block or event from each subject each iteration.
Appropriate for single subject designs.

10.6. CROSS-VALIDATION TYPE

81

k-folds CV on blocks k-partitioning on blocks or events from each subject each iteration.
Appropriate for single subject designs.
k Number of folds/partitions for CV. To create a 50
Leave one run/session out Leave out a single run (modality) from each subject each
iteration. Appropriate for single subject designs with multiple runs/sessions.
Custom Load a cross-validation matrix comprising a CV variable
Custom machine
Function

Choose a function that will perform prediction.

Arguments

10.6

Choose another prediction machine

Arguments for prediction machine.

Cross-validation type

Choose the type of cross-validation to be used

10.6.1

Leave one subject out

Leave a single subject out each cross-validation iteration

10.6.2

k-folds CV on subjects

k-partitioning of subjects at each cross-validation iteration
k
Number of folds/partitions for CV. To create a 50

10.6.3

Leave one subject per group out

Leave out a single subject from each group at a time. Appropriate for repeated measures or
paired samples designs.

10.6.4

k-folds CV on subjects per group

K-partitioning of subjects from each group at a time. Appropriate for repeated measures or
paired samples designs.
k
Number of folds/partitions for CV. To create a 50

10.6.5

Leave one block out

Leave out a single block or event from each subject each iteration. Appropriate for single subject
designs.

10.6.6

k-folds CV on blocks

k-partitioning on blocks or events from each subject each iteration. Appropriate for single subject
designs.
k
Number of folds/partitions for CV. To create a 50

82

10.6.7

CHAPTER 10. SPECIFY MODEL

Leave one run/session out

Leave out a single run (modality) from each subject each iteration. Appropriate for single subject
designs with multiple runs/sessions.

10.6.8

Custom

Load a cross-validation matrix comprising a CV variable

10.7

Include all scans

This option can be used to pass all the scans for each subject to the learning machine, regardless
of whether they are directly involved in the classification or regression problem. For example, this
can be used to estimate a GLM from the whole timeseries for each subject prior to prediction.
This would allow the resulting regression coefficient images to be used as samples.

10.8

Data operations

Specify operations to apply

10.8.1

Mean centre features

Select an operation to apply.

10.8.2

Other Operations

Include other operations?
No operations
No design specified. This option can be used for modalities (e.g. structural scans) that do not
have an experimental design or for an fMRI designwhere you want to include all scans in the
timeseries
Select Operations
Add zero or more operations to be applied to the data before the prediction machine is called.
These are executed within the cross-validation loop (i.e. they respect training/test independence)
and will be executed in the order specified.
Operation Select an operation to apply.

Chapter 11

Run model

Trains and tests the predictive machine using the cross-validation structure specified by the model.

11.1

Load PRT.mat

Select PRT.mat (file containing data/design structure).

11.2

Model name

Name of a model. Must match your entry in the
’Specify model’ batch module.

11.3

Do permutation test?

Perform a permutation test on accuracy, or not

11.3.1

No permutation test

Do not perform permutation test

11.3.2

Permutation test

Perform a permutation test.
Number of permutations
Enter the number of permutations to perform
Save permutations parameters
Set to Yes to save the parameterss obtained from eachpermutation.

83

84

CHAPTER 11. RUN MODEL

Part III

Data processing examples

85

Chapter 12

Block design fMRI dataset
Contents
12.1

GUI analysis . . . . . . . . . . . . . . .
12.1.1 Data & Design . . . . . . . . . .
12.1.2 Prepare feature set . . . . . . .
12.1.3 Specify model . . . . . . . . . .
12.1.4 Display model (optional step) .
12.1.5 Compute weights (optional step)
12.1.6 Display results . . . . . . . . . .
12.1.7 Display weights . . . . . . . . .
12.2
Batch analysis . . . . . . . . . . . . . .
12.2.1 Data & Design . . . . . . . . . .
12.2.2 Feature set / Kernel . . . . . . .
12.2.3 Specify model . . . . . . . . . .
12.2.4 Run model . . . . . . . . . . . .
12.2.5 Compute weights (optional step)

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

87
88
90
92
94
95
96
96
97
97
101
103
106
106

This chapter will describe the steps necessary to perform a classification using PRoNTo.
The dataset1 used in this chapter can be found in PRoNTo’s website http://www.mlnl.cs.
ucl.ac.uk/pronto/prtdata.html (data set 1) and the whole2 dataset is available in http:
//data.pymvpa.org/datasets/haxby2001/.
This fMRI dataset originates from a study on face and object representation in human ventral
temporal cortex [6]. In this study, the subject was shown a set of grey scale images of 8 categories
(faces, houses, cats, chairs, bottles, scissors, shoes and scrambled pictures), with 12 runs/blocks.
Each image was displayed for 500 ms and was followed by a 1500 ms rest interval. This experiment
consisted on a block-design of 9 scans of task followed by 6 scans of inter-stimulus interval. Images
were acquired with a TR of 2.5 s. The full-brain fMRI data was made up by 1452 volumes with
40 x 64 x 64 voxels, each of which with dimensions of 3.5 x 3.75 x 3.75 mm.
For simplicity, in this example we will use PRoNTo to predict if the subject is viewing an
image of a Face or a House based on the fMRI scans. We will classify the whole brain images
using Support Vector Machines, and a leave one block out cross-validation scheme.

12.1

GUI analysis

We will first analyse the data using PRoNTo’s GUI and then repeat the analysis using the
matlabbatch system.
To start, create a new directory in which to save the results of the analysis, then start up
MATLAB and type ‘prt’ or ‘pronto’ in the MATLAB prompt. This will open the main interface
of PRoNTo (Figure 12.1).
1 Pre-processed
2 Not

(realigned and normalised) data from participant 1.
pre-processed.

87

88

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.1: Main interface of PRoNTo.

12.1.1

Data & Design

• In PRoNTo’s main window, click on ‘Data & Design’ and a new window will open, ‘Data
and design’ (Figure 12.2). Then, browse the directory in which to save the PRT structure
(saved as ‘PRT.mat’);

Figure 12.2: ‘Data and design’ GUI.
• In the panel ‘Groups’, click on ‘Add’ and provide a name to the group (we only have one
group/subject), with no spaces, e.g. ‘G1’;
• Add a subject in the ‘Subject/Scans’ option, e.g. ‘s1’, and leave the ‘Scans’ tick box below
the panel unchecked. See Chapter 2 of the manual for more information on this option;

12.1. GUI ANALYSIS

89

• In the ‘Modalities’ panel, click on ‘Add’ and provide a name to the modality, e.g. ‘fMRI’.
In the ‘Design’ field, choose the option ‘Load SPM.mat’ (Figure 12.3). This file is available
with the Haxby dataset on PRoNTo’s website3 inside the folder Haxby dataset/design/;

Figure 12.3: ‘Specify modality’ GUI allows one to load a specified design from an ‘SPM.mat’ file.
– In case there is no ‘SPM.mat’ file available to use, create a new design by selecting
the option ‘Specify design’. Choose how many conditions you have, which in this case
are 8 conditions (corresponding to the 8 categories of images). This will open another
window that allows the user to write the names, onsets and durations (if the duration
is the same for all events only one value is required) of each condition (Figure 12.4).
The unit in which the onsets/durations are read in this case is ‘scans’ and the interscan
interval (TR) is 2.5 seconds. The design information (names, onsets and durations)
can be found inside the ‘Haxby design.pdf’ file in the Haxby dataset folder.

Figure 12.4: ‘Specify design’ GUI to enter the conditions, the units of design, TR and covariates.
• Finally, load all the image files available in the fMRI directory (Haxby dataset/fMRI/).
You can select all the files by using the right mouse button and clicking on the option
‘Select All’ (Figure 12.5). When all the images are selected, click on the ‘Done’ button;
• In the ‘Masks’ field, on the bottom left of the ‘Data and design’ window, select the
‘whole brain’ mask for the modality specified (Figure 12.6). The mask is available in the
masks directory inside the folder Haxby dataset/masks/;
3 http://www.mlnl.cs.ucl.ac.uk/pronto/prtdata.html

90

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.5: ‘Files’ field is used to select the scans/images for the selected subject.

Figure 12.6: This window is called when one clicks ‘Masks’.
• Click on ‘Review’ button to check the data and the design inserted in this modality (Figure
12.7). For more information on what one can do with the Review option please see Chapter
2;
• The ‘Data and design’ window should look similar to the Figure 12.8. Click on ‘Save’
button to create ‘PRT.mat’ file with the structure containing the information that has been
previously specified. If no errors are shown in the MATLAB command window, leave the
‘Data and design’ window by clicking ‘Quit’.

12.1.2

Prepare feature set

• In PRoNTo’s main window, click on ‘Prepare feature set’ and a new window will open,
‘Prepare feature set’ (Figure 12.9);
• Select the ‘PRT.mat’ file previously created in the ‘Data & Design’ step and another window

12.1. GUI ANALYSIS

91

Figure 12.7: ‘Review’ GUI allows the user to check the data and design.

Figure 12.8: ‘Data and design’ GUI final configuration.
will open, ‘Specify modality to include’ (Figure 15.4), to set the specification of different
parameters and options for each modality, which are:
– ‘Modality’ field: select the modality previously specified in the ‘Data & Design’ step,
‘fMRI’;
– ‘Conditions’ field: select ‘All scans’;
– Parameters box: select the polynomial detrend with order 1 and the ’No scaling’
option;

92

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.9: ‘Prepare feature set’ GUI.
– Features box: leave the additional mask field as it is and the ‘build one kernel per
region’ tick box unchecked. Then, click on the ‘Done’ button.
∗ As an optional step, in the ‘Additional mask for selected modality’ field, the user
can specify a ‘second-level’ mask, which can be used to select regions of interest
(ROIs) on which the classification can be performed. For instance, we can enter
the ‘fusiform gyrus’ mask available with this dataset;

Figure 12.10: ‘Specify modality to include’ GUI.
• In the ‘Prepare feature set’ window, provide a name for the feature set, e.g. ‘HaxbyFeatures’;
• Click on ‘Build kernel/data matrix’ to build the feature set and kernel (Figure 12.11). It
takes a few minutes.

12.1.3

Specify model

• In PRoNTo’s main window, click on ‘Specify model’ and a new window will open, ‘Specify
model’ (Figure 12.12);

12.1. GUI ANALYSIS

93

Figure 12.11: Preparing feature set.

Figure 12.12: ‘Specify Model’ GUI.
• Select the ‘PRT.mat’ file and provide a name to the model, e.g. ‘svmFacesHouses’;
• Select one of the ‘Feature Set’ previously defined. In this case, there is only one ‘HaxbyFeatures’;
• Leave the option ‘Use kernels’ tick box as it is, i.e. ‘Yes’;
• Select the ‘Classification’ model type and click on ‘Define classes‘ button. A new window
will open, ‘Specify classes’ (Figure 12.13), to define the number of classes and a name for
each class. We will define 2 classes. For ‘Class 1’ select subject ‘S1’ and the condition
‘Faces’ and, similarly, for ‘Class 2’ select subject ‘S1’ and the condition ‘Houses’. Click on
‘Done’;

94

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.13: ‘Specify classes’ GUI.
• Select the ‘Binary support vector machine’ option, in the ‘Machine’ field;
• Leave the option ‘Optimize hyper-parameter’ tick box unchecked and ‘Cross-Validation
Scheme’ (internal loop) as it is;
• Select the ‘Leave One Block Out’ cross-validation scheme (external loop);
• In the ‘Data operations’ box, select the ‘Sample averaging (within block)’ option, which
corresponds to a temporal compression of the data within each block, and ‘Mean centre
features using training data’ option. Then, the ‘Specify model’ window should look similar
to Figure 16.5;
• Click on ‘Specify and run model’ and the model will be immediately estimated, therefore
there is no need to use the ‘Run model’ module in this case;
• If you do not wish to average the scans within each block (i.e. to do temporal compression),
go back to the ‘Specify model’ window, give another name to the model and select the
same options mentioned above, except in the data operations part. Here, choose only the
‘Mean centre features using training data’ option. Finish by clicking on the ‘Specify and
run model’ button.

12.1.4

Display model (optional step)

• To review the model specification, in the main PRoNTo GUI, click on ‘Review kernel &
CV’ and a new window will open, ‘Review Model Specification’ (Figure 12.15);
• Select the model, ‘svmFacesHouses’, from the list at the top and click on ‘Review model’;
then, select one class from the list of ‘Class’ to see which groups, subjects and conditions
this class comprises (Figure 12.16);
• To review the data and cross-validation matrix click on ‘Review CV’ (Figure 12.17). For
more information on what these matrices mean, please consult the previous chapters of the
manual;
• To review the kernel, click on ‘Show kernel’ (Figure 12.18).

12.1. GUI ANALYSIS

95

Figure 12.14: ‘Specify model’ GUI final configuration.

Figure 12.15: Review CV & kernel window.

12.1.5

Compute weights (optional step)

• In PRoNTo’s main window, click on ‘Compute weights’ and a new window will open, ‘Compute weights’ (Figure 12.19);
• Select the ‘PRT.mat’ file;
• Select the model from the list to ‘Models computed in PRT’, ‘svmFacesHouses’ model;
• Leave the option ‘Compute average/kernel weight per region’ tick box unchecked;
• Click on ‘Compute weights’ button. Computations will be displayed on the MATLAB
command window.

96

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.16: Review model specification window to Class 1.

Figure 12.17: Data and cross-validation matrix from ‘Review CV’ option.

12.1.6

Display results

• In PRoNTo’s main window, click on ‘Display results’ and select the ‘PRT.mat’ file. This
will open the main results window (Figure 15.7);
• In the ‘Model’ panel, select the model that you want to view, ‘svmFacesHouses’. The
performance should be similar to the one in Figure 12.21;
• In the ‘Results’ window, one can select a different plot in the ‘Plots’ list;
• Finally, to check the significance of the results, run a permutation test by clicking on
‘Permutation test’ button with 100 repetitions. Results will be displayed on the MATLAB
prompt (Figure 12.22). Please note that 100 repetitions is a small amount, if possible, this
should be greater (e.g. 1000).

12.1.7

Display weights

• In PRoNTo’s main window, click on ‘Display weights’ and select the ‘PRT.mat’ file. This
will open the ‘Model interpretation’ window (Figure 12.23);

12.2. BATCH ANALYSIS

97

Figure 12.18: Kernel matrix used for classification.

Figure 12.19: ‘Compute weights’ GUI.
• By clicking on ‘Model’, svmFacesHouses, an image will appear in the ‘Weights map’ box;
and to show the ‘Anatomical img’ you have to load an anatomical image for reference. A
template image can be found in SPM’s canonical folder (‘single subj T1‘ file). The final
window will look similar to the one shown in the Figure 12.24.

12.2

Batch analysis

This tutorial will now show how to analyse the same data but using the matlabbatch system.
Once again, create a new directory where you wish to save the results. On the main interface
of PRoNTo click on the ‘Batch’ button to open the ‘matlabbatch’. Alternatively, type ‘prt batch’
on the MATLAB command window. On the menu bar of the batch, there is a PRoNTo menu
with the 5 options shown in the main steps interface (Figure 12.25).

12.2.1

Data & Design

• Click on ‘Data & Design’ in the PRoNTo menu (Figure 12.26);
• In the ‘Directory’ field, select a directory where the ‘PRT.mat’ file will be saved. There
are three ways of editing all fields in matlabbatch: (i) by using the right mouse button

98

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.20: ‘Results’ GUI.

Figure 12.21: Summary for model’s performance.
and clicking on the current option, (ii) clicking on current button in the window or (iii) by
double clicking;
• In the ‘Groups’ field:
– Add one group;
– In the field ‘Name’, provide a name without spaces to that group, e.g. ‘G1’;
– In the field ‘Select by’, select the ‘Subjects’ option and add one subject;
– Add one modality for this subject and provide a name, e.g. ‘fMRI’; define the interscan
interval of 2.5 seconds; and in the field ‘Scans’, select all the image files available in
the fMRI directory of the Haxby dataset;
– In the ‘Data & Design’ field, choose ‘Load SPM.mat’ option. This file is available with
the Haxby dataset on PRoNTo’s website4 inside the folder Haxby dataset/design/.
The batch editor should look similar to the one in Figure 12.27;
∗ In case there is no ‘SPM.mat’ file available to use, create a new design by selecting
the option ‘Specify design’. Choose how many conditions you have, which in this
case are 8 conditions (corresponding to the 8 categories of images). The unit
4 http://www.mlnl.cs.ucl.ac.uk/pronto/prtdata.html

12.2. BATCH ANALYSIS

99

Figure 12.22: Sample of the MATLAB window after the permutation test (100 repetitions).

Figure 12.23: ‘Model interpretation’ GUI.
in which the onsets/durations are read is ‘Scans’. Write the names, onsets and
durations of each condition (Figure 12.28);
• In the ‘Masks’ field, add a new modality and provide the same modality name, ‘fMRI’; and
select the ‘whole brain’ mask available in the masks directory of the Haxby dataset. The
name of the modality here has to be exactly the same as in ‘Modalities’, otherwise it will
not work;

100

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.24: ‘Model interpretation’ GUI with results.

Figure 12.25: Menu PRoNTo in the main matlabbatch window.

• Leave the ‘HRF overlap’ and the ‘HRF delay’ fields as default;
• In the ‘Review’ field, select ‘Yes’ if you would like to review your data and design in a
separate window. Otherwise, leave as it is, i.e. ‘No’.

12.2. BATCH ANALYSIS

101

Figure 12.26: Data and design module in matlabbatch.

Figure 12.27: Data and design module in matlabbatch.

12.2.2

Feature set / Kernel

• Click on ‘Feature set / Kernel’ option on PRoNTo’s matlabbatch menu (Figure 12.29);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Data & Design’ step (Figure 12.30) or click on the
‘Select files’ button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the ‘Feature/kernel’ set, e.g. ‘HaxbyFeatures’;

102

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.28: Data and design module. The ‘Specify design’ option.

Figure 12.29: Feature set/Kernel module in matlabbatch.
• Add one modality and select the modality name with the ‘Dependency’ button5 (Data &
Design:Mod#1 name);
– In the ‘Scans/Conditions’ field , select the ‘All scans’ option;
– In the ‘Voxels to include’ field, select ‘All voxels’ option, this means we are not entering
an additional second-level mask;
5 Or type it in manually, ‘fMRI’, but the name needs to be exactly the same as the one specified in the ‘Data
& Design’ module.

12.2. BATCH ANALYSIS

103

Figure 12.30: Feature set / Kernel module in matlabbatch. This window is called to establish a
dependency connection with the previous ‘Data and design’ module.
∗ This is an optional step. In the ‘Voxels to include’ options, the user can specify
a ‘second-level’ mask, which would define regions of interest (ROIs) on which the
classification can be performed. In this case, select the ‘fusiform gyrus’ mask;
– In the ‘Detrend’ field, select ‘Polynomial detrend’ option with order 1;
– In the ‘Scale input scans’ field, select ‘No scaling’ option;
– Leave ‘Load Atlas’ as default. After all these steps, the batch editor should look similar
to the one in Figure 12.31;
• Leave the ‘Generate Multiple Kernels’ and the ‘Use one kernel per modality’ fields as default.

12.2.3

Specify model

• Click on ‘Specify model’ option on PRoNTo’s matlabbatch menu (Figure 12.32);
• With ‘Load PRT.mat’ field selected, click on ‘Dependency’ button to associate the ‘PRT.mat’
file created in the previous ‘Feature set / Kernel’ step (Figure 12.33) or click on ‘Select files’
button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the model, e.g. ‘svmFacesHouses’;
• Leave the ‘Use kernels’ field as it is, i.e. ‘Yes’;
• In the ‘Feature sets’ field, select the feature set name with the ‘Dependency’ button6 ;
• Select the ‘Classification’ model type:
– Add 2 new classes;
– For Class (1) write ‘Faces’ on the name field and add one group. Select the group
name from the ‘Data & Design’ module (‘Data & Design:Group#1 name’) with the
‘Dependency’ button7 . Similarly, for Class (2) write ‘Houses’ on the name field and
add the group created in the ‘Data & Design’ module, ‘G1’;
6 or write it exactly as previously defined in the ‘Feature set / Kernel’ module (option ‘Feature set/Kernel:
Feature/kernel name’), here ‘HaxbyFeatures’.
7 Or write it exactly, as previously defined in the Data & Design’ module, here ‘G1’

104

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Figure 12.31: Feature set / Kernel module. Selected parameters in the Modality option.

Figure 12.32: Specify model module in matlabbatch.
– In the ‘Subjects’ field, type ‘1’ (only subject 1 is selected);
– In the ‘Conditions / Scans’ field, select the ‘Specify Conditions’ option and add a new
condition. Provide a name for this condition, i.e. for Class (1) ‘Faces’ and for Class
(2) ‘Houses’. Note that this name needs to be spelled exactly as specified in the ‘Data
& Design’ module: if you simply loaded an ‘SPM.mat’ file for the design, you must
know the names of the conditions;
– After all this steps, the batch editor should look similar to the one in Figure 12.34

12.2. BATCH ANALYSIS

105

Figure 12.33: Specify model module in matlabbatch. This window is called to establish a dependency connection with the previous ‘Feature set / Kernel’ module.

Figure 12.34: Specify model module. Selected parameters in the Class option.
• In the ‘Machine’ field:
– Select the ‘SVM Classification’ option;
– Leave the ‘Optimize hyper-parameter’ field as it is, i.e. ‘No’;
– Leave the ‘Soft-margin hyper-parameter’ field as it is, i.e. ‘1’;
– Leave the ‘Cross validation type for hyper-parameter optimization’ field as it is, i.e.
‘Leave one subject out’;

106

CHAPTER 12. BLOCK DESIGN FMRI DATASET

• In the ‘Cross-validation type’ field, select ‘Leave one block out’ option;
• Leave the ‘Include all scans’ field as it is, i.e. ‘No’;
• In the ‘Data operations’ field:
– Leave the ‘Mean centre features’ field as it is, i.e. ‘Yes’;
– Leave the ‘Other Operations’ field as it is, i.e. ‘No operations’;

12.2.4

Run model

• Click on the ‘Run model’ option on PRoNTo’s matlabbatch menu (Figure 12.35);

Figure 12.35: Run model module in matlabbatch.
• With the ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Specify model’ step;
• Select the model name from the ‘Specify model’ module with the ‘Dependency’ button8 ;
• In the field ‘Do permutation test?’, select ‘yes’ with 100 repetitions or leave as default (1000
repetitions).

12.2.5

Compute weights (optional step)

• Click on the ‘Compute weights’ option on PRoNTo’s matlabbatch menu (Figure 12.36);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Run model’ step;
• Select the model name from the ‘Specify model’ module with the ‘Dependency’ button;
• It’s optional to define a name for the image;
• Leave the ‘Build weights images for permutations’ field as it is, i.e. ‘No’;
8 Or

write it exactly, as previously defined in the ‘Specify model’ module, here ‘svmFacesHouses’

12.2. BATCH ANALYSIS

107

Figure 12.36: Compute weights module in matlabbatch.
• Leave the ‘Load Atlas’ field as default.
Finally, save the batch (e.g. as batch run all.m) and click on the ‘Run Batch’ option, in the
‘File menu’. The batch file created can then be opened and edited for further analyses. The
results will be the same as those obtained using the GUI (see Section 12.1.6 of this chapter).
Please note that in this case the ‘Sample averaging (within block)’ operation was not selected
when specifying the model. In order to obtain the same results as before, the model has to use
the same data operations.

108

CHAPTER 12. BLOCK DESIGN FMRI DATASET

Chapter 13

Regression dataset
Contents
13.1

GUI analysis . . . . . . . . . . . . . . . .
13.1.1 Data & Design . . . . . . . . . . .
13.1.2 Prepare feature set . . . . . . . .
13.1.3 Specify model . . . . . . . . . . .
13.1.4 Display results . . . . . . . . . . .
13.2
Batch analysis . . . . . . . . . . . . . . .
13.2.1 Data & Design . . . . . . . . . . .
13.2.2 Feature set/Kernel . . . . . . . .
13.2.3 Specify model (KRR) . . . . . . .
13.2.4 Run model (KRR) . . . . . . . . .
13.2.5 Specify and Run model (RVR and

. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
GPR)

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

. 109
. 109
. 111
. 111
. 113
. 113
. 113
. 114
. 115
. 116
. 116

This chapter will describe the steps necessary to perform a regression using PRoNTo. These
are similar to the ones in the previous chapter, thus, the reader is advised to complete the tutorial
in Chapter 12 before moving on, since the explanation of some steps will be less descriptive. The
dataset used in this chapter can be found on PRoNTo’s website http://www.mlnl.cs.ucl.ac.
uk/pronto/prtdata.html (data set 3).

13.1

GUI analysis

As in Chapter 12, the analysis of the data will start with the PRoNTo’s GUI. Please create a
folder in your computer to store the results and type ‘prt’ or ‘pronto’ on the MATLAB command
window. This will open the main interface of PRoNTo (see Figure 12.1 in the previous chapter).

13.1.1

Data & Design

• In PRoNTo’s main window, click on ‘Data & Design’. Like in the previous chapter, browse
the directory in which to save the PRT structure (saved as ‘PRT.mat’);
• In the panel ‘Groups’, click on ‘Add’ and provide a name to the group, e.g. ‘Aged’;
• Unlike the previous chapter, all the images in the dataset correspond to different subjects;
therefore, click on the ‘Scans’ tick box. This will lock the ‘Subjects/Scans’ field, allowing
you to skip to the third field;
• In the ‘Modalities’ panel, click on ‘Add’ and provide a name for the modality, e.g. ‘fMRI’;
• Load all the image files available in the directory (IXIdata/aged/Guys/). You can select all
the files by using the right mouse button and clicking on the option ‘Select All’. When all
the images are selected, click on the ‘Done’ button;
109

110

CHAPTER 13. REGRESSION DATASET

• Into the ‘Regression targets’ field, write (or paste) the list of target values available in the
‘Age old Guys’ file (IXIdata/aged/). The final window should look like to the Figure 13.1.
Press ‘OK’;

Figure 13.1: ‘Specify modality’ GUI.

• In the ‘Masks’ field, on the bottom left of the ‘Data and design’ window, select the
‘SPM mask noeyes’ mask for the specified modality. The mask is available in the path
where you have installed PRoNTo (PRoNTo/masks/);
• The ‘Data and design’ window should look similar to Figure 13.2. Click on the ‘Save’
button to create ‘PRT.mat’ file with the structure containing the information that has been
previously specified. If no errors are shown in the MATLAB command, leave the ‘Data and
design’ window by clicking ‘Quit’.

Figure 13.2: ‘Data and design’ GUI final configuration.

13.1. GUI ANALYSIS

13.1.2

111

Prepare feature set

• In PRoNTo’s main window, click on ‘Prepare feature set’ and a new window will open,
‘Prepare feature set’ (see Figure 12.9 in the previous chapter);
• Select the ‘PRT.mat’ file previously created in the ‘Data & Design’ step and another window
will open, ‘Specify modality to include’ (Figure 13.3). There is no need to change anything
for this example. Just click on the ‘Done’ button;

Figure 13.3: ‘Specify modality to include’ GUI.
• In the ‘Prepare feature set’ window, provide a name to the feature set, e.g. ‘Scalar Momentum’;
and click on ‘Build Kernel / data matrix’ to build the feature set and kernel;

13.1.3

Specify model

• In PRoNTo’s main window, click on ‘Specify model’ and a new window will open, ‘Specify
model’ (see Figure 12.12 in the previous chapter);
• Select the ‘PRT.mat’ file and provide a name to the model, e.g. ‘KRR’;
• Select one of the ‘Feature Set’ previously defined. In this case, there is only one: ‘Scalar Momentum’;
• Leave the option ‘Use kernels’ tick box as it is, i.e. ‘Yes’;
• Select the ‘Regression’ model type and click on the ‘Select subjects/scans‘ button. This
will open a new window, ‘Specify subjects/scans to regress’, click on the ‘Select all’ button
to use all the scans for the regression (Figure 13.4);
• Select the ‘Kernel Ridge Regression’ option, in the Machine field;
• Leave the option ‘Optimize hyper-parameter’ tick box unchecked and ‘Cross-Validation
Scheme’ (internal loop) as it is;
• Select the ‘Leave One Subject Out’ cross-validation scheme (external loop);
• In the ‘Data operations’ box, select the ‘Mean centre features using training data’ option.
The final ‘Specify model’ window should look similar to the Figure 13.5. Click on the
‘Specify and run model’ button;

112

CHAPTER 13. REGRESSION DATASET

Figure 13.4: ‘Specify subjects/scans to regress’ GUI.

Figure 13.5: ‘Specify model’ GUI final configuration.

• Repeat the process two times using the other two machines. To do this, just follow the same
steps in this section, but select the other options in the ‘Machine’ drop-down list (‘Relevance
Vector Regression’ and ‘Gaussian Process Regression’) and give different names to the each
model.

13.2. BATCH ANALYSIS

13.1.4

113

Display results

• In PRoNTo’s main window, click on ‘Display results’ and select the ‘PRT.mat’ file. This
will open the main results window similar to the Figure 13.6;

Figure 13.6: ‘Results’ GUI.
• In the ‘Results’ window, one can select the different regression models in the ‘Model’ list on
the upper right region. This will show the results obtained using each one of the regression
models.

13.2

Batch analysis

In this section, the previous experiment will be repeated using the ‘matlabbatch’ system. The
reader is advised to complete the tutorial in Section 12.2 before continuing, since the explanation
of each step will be less descriptive.
Once again, to analyse the data, create a new directory in which to save the results of the
analysis. On the main interface of PRoNTo click on the ‘Batch’ button to open the ‘matlabbatch’.
Alternatively, type ‘prt batch’ in the MATLAB prompt.

13.2.1

Data & Design

• Click on ‘Data & Design’ in the PRoNTo menu (see Figure 12.26 in the previous chapter);
• In the ‘Directory’ field, select a directory where the ‘PRT.mat’ file will be saved;
• In the ‘Groups’ field:
– Add one group;

114

CHAPTER 13. REGRESSION DATASET

– In the field ‘Name’, provide a name without spaces for this group, e.g. ‘Aged’;
– In the field ‘Select by’, select the ‘Scans’ option and add a new modality. For more
information on the Scans option please consult Chapter 2;
– Provide a name for this modality, e.g. ‘fMRI’; select the image files available in the
‘aged/Guy’ directory of the IXI dataset and write (or paste) the regression targets1 in
the ‘Regression targets (per scans)’ field;
– Leave ‘Covariates’ field as default. The batch editor should look similar to the Figure
13.7;

Figure 13.7: Data and design module in matlabbatch.
– In the ‘Masks’ field, add a new modality and provide the same modality name, ‘fMRI’;
and select the ‘SPM mask noeyes’ mask available in the path where you have installed
PRoNTo (PRoNTo/masks/). The name of the modality here has to be exactly the
same as in ‘Modalities’, otherwise it will not work;
– Leave the ‘HRF overlap’ and the ‘HRF delay’ fields as default;
– In the ‘Review’ field, select ‘Yes’ if you would like to review your data and design in a
separate window. Otherwise, leave as it is, i.e. ‘No’.

13.2.2

Feature set/Kernel

• Click on the ‘Feature set / Kernel’ option on PRoNTo’s matlabbatch menu (see Figure
12.29 in the previous chapter);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Data & Design’ step or click on the ‘Select files’
button to browse where ‘PRT.mat’ file was saved;
1 Available

in the ‘Age old Guys’ file (IXIdata/aged/)

13.2. BATCH ANALYSIS

115

• Provide a name to the ‘Feature/kernel’ set, e.g. ‘Scalar Momentum’;
• Add one modality and select the modality name with the ‘Dependency’ button2 (Data &
Design:Mod#1 name);
– In the ‘Scans/Conditions’ field, select ‘All scans’ option;
– In the ‘Voxels to include’ field, select ‘All voxels’ option, this means we are not entering
with an additional second-level mask;
– In the ‘Detrend’ field, select the ‘None’ option;
– In the ‘Scale input scans’ field, select the ‘No scaling’ option;
– Leave ‘Load Atlas’ as default. After all these steps, the batch editor should look similar
to the one in Figure 13.8;

Figure 13.8: Feature set / Kernel module. Selected parameters in the Modality option.
• Leave the ‘Generate Multiple Kernels’ and the ‘Use one kernel per modality’ fields as default.

13.2.3

Specify model (KRR)

• Click on the ‘Specify model’ option on PRoNTo’s matlabbatch menu (see Figure 12.32 in
the previous chapter);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Feature set / Kernel’ step or click on the ‘Select
files’ button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the model, e.g. ‘KRR’;
2 Or type it in manually, ‘fMRI’, but the name needs to be exactly the same as the one specified in the ‘Data
& Design’ module.

116

CHAPTER 13. REGRESSION DATASET

• Leave the ‘Use kernels’ field as it is, i.e. ‘Yes’;
• Select the feature set name with the ‘Dependency’ button3 ;
• Select the ‘Regression’ model type:
– Add a new group and call it ‘Aged’;
– In the ‘Subjects’ field, type ‘1:170’. This will instruct the program to use all the 170
scans, i.e. from scan 1 to scan 170;
• In the ‘Machine’ field:
– Select the ‘Kernel Ridge Regression’ option:
– Leave the ‘Optimize hyper-parameter’ field as it is, i.e. ‘No’;
– Leave the ‘Regularization’ field as it is, i.e. ‘1’;
– Leave the ‘Cross validation type for hyper-parameter optimization’ field as it is, i.e.
‘Leave one subject out’;
• In the ‘Cross-validation type’ field, select ‘Leave One Subject Out’ option;
• Leave the ‘Include all scans’ field as it is, i.e. ‘No’;
• In the ‘Data operations’ field:
– Leave the ‘Mean centre features’ field as it is, i.e. ‘Yes’;
– Leave the ‘Other Operations’ field as it is, i.e. ‘No operations’;

13.2.4

Run model (KRR)

• Click on the ‘Run model’ option on PRoNTo’s matlabbatch menu (see Figure 12.35 in the
previous chapter);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Specify model’ step;
• Select the model name from the ‘Specify model’ module with the ‘Dependency’ button4 ;
• In the field ‘Do permutation test?’, leave as it is, i.e. ‘No permutation test’

13.2.5

Specify and Run model (RVR and GPR)

The specification of the other models (‘Relevance Vector Regression’ and ‘Gaussian Process Regression’) follows the same procedure as the ‘KRR’. The only difference is that in the ‘Machine’
field of the ‘Specify model’ module, one has to choose the appropriate machine to use (‘Relevance
Vector Regression’ or ‘Gaussian Process Regression’). The parameters used for each machine
should be the default ones.
Note that when the ‘PRT.mat’ file is loaded in each module, the user should select the latest
option on the list.
When all the models are defined, the ‘Module List’ should contain 8 modules:
1. Data & Design;
2. Feature set/Kernel;
3. Specify model;
4. Run model;
3 or write it exactly as previously defined in the ‘Feature set / Kernel’ module (option ‘Feature set/Kernel:
Feature/kernel name’), here ‘Scalar Momentum’.
4 or write it exactly as previously defined in the ‘Specify model’ module, here ‘KRR’

13.2. BATCH ANALYSIS

117

5. Specify model;
6. Run model;
7. Specify model;
8. Run model.
Note that modules 3 and 4 correspond to the KRR model; 5 and 6 to the RVR model; 7 and
8 to the GPR model.
When all the modules are added, just click on the ‘Run Batch’ button. The resulting
‘PRT.mat’ file will be saved in the specified directory and the results can be viewed using the
process described in Section 13.1.4.

118

CHAPTER 13. REGRESSION DATASET

Chapter 14

Multiple Kernel Learning
Example
Contents
14.1

GUI analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
14.1.1

Data & Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

14.1.2

Prepare feature set . . . . . . . . . . . . . . . . . . . . . . . . . . .

121

14.1.3

Specify model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

122

14.1.4

Display model (optional step) . . . . . . . . . . . . . . . . . . . . .

123

14.1.5

Compute weights (optional step) . . . . . . . . . . . . . . . . . . . .

124

14.1.6

Display results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

124

14.1.7

Display weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

125

14.2

120

Batch analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
14.2.1

Data & Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

126

14.2.2

Feature set / Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . .

129

14.2.3

Specify model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

131

14.2.4

Run model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

132

14.2.5

Compute weights (optional step) . . . . . . . . . . . . . . . . . . . .

134

This chapter will describe the steps necessary to perform a classification with SimpleMKL
http://asi.insa-rouen.fr/enseignants/~arakoto/code/mklindex.html [14] using PRoNTo.
These are similar to the ones in Chapter 12, thus, the reader is advised to complete the tutorial
in Chapter 12 before moving on, since the explanation of some steps will be less descriptive.
Many practical learning problems involve multiple and heterogeneous data sources. In this
way, Multiple Kernel Learning (MKL) [2] has been proposed to simultaneously learn and combine
different models, represented by different kernels, in supervised learning settings. In MKL, the
kernel K can be considered as a linear combination of M ‘basis kernels’. For further details,
please refer to [2].
One example of a MKL approach based on SVM is the SimpleMKL algorithm [14]. Essentially,
the algorithm is based on a gradient descent on the SVM objective value and iteratively determine
the combination of kernels by a gradient descent wrapping [14]. For further details, please refer
to [14]
We will use the same dataset used in Chapter 12, this fMRI dataset originates from a study
on face and object representation in human ventral temporal cortex [6]. The dataset1 used in this
chapter can be found in PRoNTo’s website http://www.mlnl.cs.ucl.ac.uk/pronto/prtdata.
html (data set 1) and the whole2 dataset is available in http://data.pymvpa.org/datasets/
haxby2001/.
1 Pre-processed
2 Not

(realigned and normalised) data from participant 1.
pre-processed.

119

120

CHAPTER 14. MULTIPLE KERNEL LEARNING EXAMPLE

For simplicity, in this example we will use PRoNTo to predict if the subject is viewing an
image of a Face or a House based on the fMRI scans. We will classify the whole brain images
using SimpleMKL and a leave one block out cross-validation scheme.

14.1

GUI analysis

We will first analyse the data using PRoNTo’s GUI and then repeat the analysis using the
matlabbatch system.
To start, create a new directory in which to save the results of the analysis, then start up
MATLAB and type ‘prt’ or ‘pronto’ in the MATLAB prompt. This will open the main interface
of PRoNTo (Figure 14.1).

Figure 14.1: Main interface of PRoNTo.

14.1.1

Data & Design

• In PRoNTo’s main window, click on ‘Data & Design’. Like in the previous chapters, browse
the directory in which to save the PRT structure (saved as ‘PRT.mat’);
• In the panel ‘Groups’, click on ‘Add’ and provide a name to the group (we only have one
group/subject), with no spaces, e.g. ‘G1’;
• Add a subject in the ‘Subject/Scans’ option, e.g. ‘S1’, and leave the ‘Scans’ tick box below
the panel unchecked. See Chapter 2 of the manual for more information on this option;
• In the ‘Modalities’ panel, click on ‘Add’ and provide a name to the modality, e.g. ‘fMRI’. In
the ‘Design’ field, choose the option ‘Load SPM.mat’. This file is available with the Haxby
dataset on PRoNTo’s website3 inside the folder Haxby dataset/design/;
– In case there is no ‘SPM.mat’ file available to use, create a new design by selecting
the option ‘Specify design’. Choose how many conditions you have, which in this case
are 8 conditions (corresponding to the 8 categories of images). This will open another
window that allows the user to write the names, onsets and durations (if the duration
is the same for all events only one value is required) of each condition. The unit in
3 http://www.mlnl.cs.ucl.ac.uk/pronto/prtdata.html

14.1. GUI ANALYSIS

121

which the onsets/durations are read in this case is ‘scans’ and the interscan interval
(TR) is 2.5 seconds. The design information (names, onsets and durations) can be
found inside the ‘Haxby design.pdf’ file in the Haxby dataset folder.
• Finally, load all the image files available in the fMRI directory (Haxby dataset/fMRI/).
You can select all the files by using the right mouse button and clicking on the option
‘Select All’. When all the images are selected, click on the ‘Done’ button;
• In the ‘Masks’ field, on the bottom left of the ‘Data and design’ window, select the
‘whole brain’ mask for the specified modality. The mask is available in the masks directory
inside the folder Haxby dataset/masks/;
• Click on the ’Review’ button to check the data and the design inserted for this modality.
For more information on what one can do with the Review option, please see Chapter 2;
• The ‘Data and design’ window should look similar to the one in Figure 14.2. Click on the
’Save’ button to create the ‘PRT.mat’ file with the structure containing the information
that has been previously specified. If no errors are shown in the MATLAB command, leave
the ‘Data and design’ window by clicking ‘Quit’.

Figure 14.2: ‘Data and design’ GUI final configuration.

14.1.2

Prepare feature set

• In PRoNTo’s main window, click on ‘Prepare feature set’ and a new window called ‘Prepare
feature set’ will open (see Figure 12.9 in Chapter 12);
• Select the ‘PRT.mat’ file previously created in the ‘Data & Design’ step and another window will open, ‘Specify modality to include’ (see Figure 15.4 in Chapter 12), to set the
specification of different parameters and options for each modality, which are:
– ‘Modality’ field: select the modality previously specified in the ‘Data & Design’ step,
‘fMRI’;
– ‘Conditions’ field: select ‘All scans’;
– ‘Parameters’ box: select the polynomial detrend with order 1 and the ’No scaling’
option;

122

CHAPTER 14. MULTIPLE KERNEL LEARNING EXAMPLE

– ‘Features’ box: select the ‘Build one kernel per region’ tick box and load the ‘AAL’
atlas (named ’aal 79x91x69’) available in the PRoNTo directory (PRoNTo/atlas/).
Then, click on the ’Done’ button. The final ‘Specify modality to include’ window
should look similar to the one in Figure 14.3;

Figure 14.3: ‘Specify modality to include’ GUI final configuration.
• In the ‘Prepare feature set’ window, provide a name to the feature set, e.g. ‘HaxbyFeatures’;
• Click on ‘Build kernel/data matrix’ to build the feature set and kernel. It will take a few
minutes.

14.1.3

Specify model

• In PRoNTo’s main window, click on ‘Specify model’ and a new window called ‘Specify
model’ will open (see Figure 12.12 in Chapter 12);
• Select the ‘PRT.mat’ file and provide a name to the model, e.g. ‘mklFacesHouses’;
• Select one of the feature sets previously defined. In this case, there is only one: ‘HaxbyFeatures’;
• Leave the option ‘Use kernels’ tick box as it is, i.e. ‘Yes’;
• Select the ‘Classification’ model type and click on the ’Define classes‘ button. A new window
will open, ‘Specify classes’, to define the number of classes and a name for each class. We
will define 2 classes:
– for ‘Class 1’ select subject ‘S1’ and the condition ‘Faces’ and;
– for ‘Class 2’ select subject ‘S1’ and the condition ‘Houses’. Click on ‘Done’.
• Select the ‘L1- Multi-Kernel Learning’ option, in the Machine field;
• Select the ‘Optimize hyper-parameter’ tick box and leave the textbox empty. This will force
PRoNTo to use an array with default hyper-parameter values for the optimization. You
can choose other values by inputing them in this box (e.g. [0.1, 1, 100]);
• In the ‘Cross-Validation Scheme’ (internal loop) field, select the option ‘k-fold CV on Block’.
A window will appear asking to define the value of k, set it to 4;
• Select the ‘Leave One Block Out’ cross-validation scheme (external loop);

14.1. GUI ANALYSIS

123

• In the ‘Data operations’ box, select the ‘Mean centre features using training data’ and
‘Normalize samples’ options Then, the ‘Specify model’ window should look similar to the
one in Figure 14.4;

Figure 14.4: ‘Specify model’ GUI final configuration.
• Click on ‘Specify and run model’ and the model will be immediately estimated, therefore
there is no need to use the ‘Run model’ module in this case. It will take a few minutes to
complete.

14.1.4

Display model (optional step)

• To review the model specification, in the main PRoNTo GUI, click on ‘Review kernel & CV’
and a new window will open, ‘Review Model Specification’ (see Figure 12.15 in Chapter
12);
• Select the model, ‘mklFacesHouses’, from the list at the top and click on ‘Review model’;
then, select one class from the list to see which groups, subjects and conditions belong to
this class (see Figure 12.16 in Chapter 12);
• To review the data and cross-validation matrix click on ‘Review CV’ (Figure 14.5). For
more information on what these matrices mean, please consult the previous chapters of the
manual;
• To review the kernel, click on ‘Show kernel’ (Figure 14.6).

124

CHAPTER 14. MULTIPLE KERNEL LEARNING EXAMPLE

Figure 14.5: Data and cross-validation matrix from ‘Review CV’ option.

Figure 14.6: Kernel matrix used for classification.

14.1.5

Compute weights (optional step)

• In PRoNTo’s main window, click on ‘Compute weights’ and a new window will open, ‘Compute weights’ (see Figure 12.19 in Chapter 12);
• Select the ‘PRT.mat’ file;
• Select the model from the list to ‘Models computed in PRT’, ‘mklFacesHouses’ model;
• Check the tick box option ‘Compute average/kernel weight per region’;
• Click on the ’Compute weights’ button. Computations will be displayed on the MATLAB
prompt.

14.1.6

Display results

• In PRoNTo’s main window, click on ‘Display results’ and select the ‘PRT.mat’ file. This
will open the main results window. In the ‘Model’ panel, select the model that you want to
view, ‘mklFacesHouses’, and the results through performance will be similar to the one in
Figure 14.7;

14.2. BATCH ANALYSIS

125

Figure 14.7: Summary for model’s performance.
• In the ‘Results’ window, one can select the different plots in the ‘Plots’ list. Please note
that there is a new plot on the list, this displays information about the hyper-parameter
optimization, for more information, please refer to Chapter 6.

14.1.7

Display weights

• In PRoNTo’s main window, click on ‘Display weights’ and select the ‘PRT.mat’ file. This
will open the ‘Model interpretation’ window. By clicking on ‘Model’, mklFacesHouses, an
image will appear in the ‘Weights map’ box; and to show the ‘Anatomical img’ you have
to load an anatomical image for reference. A template image can be found in the SPM’s
canonical folder ‘single subj T1’. The final result window will look similar to that shown in
Figure 14.8.
• Since the machine used in the example was MKL with a kernel calculated for each brain
region, it is possible to see the contributions of each region. The labels for the regions can be
found in the same folder where the atlas is located (PRoNTo/atlas). For more information,
please refer to Chapter 7.

14.2

Batch analysis

This tutorial will now show how to analyse the same data but using the matlabbatch system.
Once again, to analyse the data, create a new directory in which to save the results of the
analysis, saved as ’PRT.mat’. On the main interface of PRoNTo click on the ’Batch’ button to

126

CHAPTER 14. MULTIPLE KERNEL LEARNING EXAMPLE

Figure 14.8: Summary for model’s performance.
open the ‘matlabbatch’. Alternatively, type ‘prt batch’ in the MATLAB prompt. On the menu
bar of the batch, there is a PRoNTo menu with the 5 options shown in the main steps interface
(Figure 14.9).

14.2.1

Data & Design

• Click on ‘Data & Design’ in the PRoNTo menu (Figure 14.10);
• In the ‘Directory’ field, select a directory where the ‘PRT.mat’ file will be saved. There
are three options to edit all fields in matlabbatch: (i) by using the right mouse button
and clicking on the current option, (ii) clicking on current button in the window or (iii) by
double clicking;
• In the ‘Groups’ field:
– Add one group;
– In the field ‘Name’, provide a name without spaces to that group, e.g. ‘G1’;
– In the field ‘Select by’, select the ‘Subjects’ option and add one subject;
– Add on modality for this subject and provide a name, e.g. ‘fMRI’; define the ‘Interscan
interval’ of 2.5 seconds; and in the field ‘Scans’, select all the image files available in
the fMRI directory of the Haxby dataset;

14.2. BATCH ANALYSIS

127

Figure 14.9: Menu PRoNTo in the main matlabbatch window.

Figure 14.10: Data and design module in matlabbatch.
– In the ‘Data & Design’ field, choose ‘Load SPM.mat’ option. This file is available with
the Haxby dataset on PRoNTo’s website4 inside the folder Haxby dataset/design/.
The batch editor should look similar to the Figure 14.11;
∗ In case there is no ‘SPM.mat’ file already available to use, create a new design by
selecting the option ‘Specify design’. Choose how many conditions you have, which
in this case are 8 conditions (corresponding to the 8 categories of images). The
4 http://www.mlnl.cs.ucl.ac.uk/pronto/prtdata.html

128

CHAPTER 14. MULTIPLE KERNEL LEARNING EXAMPLE

Figure 14.11: Data and design module in matlabbatch.
unit in which the onsets/durations are read is ‘Scans’. Write the names, onsets
and durations of each condition (Figure 14.12);

Figure 14.12: Data and design module. The ‘Specify design’ option.

• In the ‘Masks’ field, add a new modality and provide the same modality name, ‘fMRI’; and
select the ‘whole brain’ mask available in the masks directory of the Haxby dataset. The
name of the modality here has to be exactly the same as in ‘Modalities’, otherwise it will
not work;

14.2. BATCH ANALYSIS

129

• Leave the ‘HRF overlap’ and the ‘HRF delay’ options as default;
• In the ‘Review’ field, select ‘Yes’ if you would like to review your data and design in a
separate window. Otherwise, leave as it is, i.e. ‘No’.

14.2.2

Feature set / Kernel

• Click on the ’Feature set / Kernel’ option on PRoNTo’s matlabbatch menu (Figure 14.13);

Figure 14.13: Feature set/Kernel module in matlabbatch.
• With ‘Load PRT.mat’ field selected, click on the ’Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Data & Design’ step (Figure 14.14) or click on the
‘Select files’ button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the ‘Feature/kernel’ set, e.g. ‘HaxbyFeatures’;
• Add one modality and select the modality name with the ‘Dependency’ button5 (Data &
Design:Mod#1 name);
– In the ‘Scans/Conditions’ field , select ‘All scans’ option;
– In the ‘Voxels to include’ field, select ‘All voxels’ option, this means we are not entering
an additional second-level mask;
– In the ‘Detrend’ field, select ‘Polynomial detrend’ option with order 1;
– In the ‘Scale input scans’ field, select ‘No scaling’ option;
– In the ‘Use atlas to build ROI specific kernels’, select an atlas file AAL (named
‘aal 79x91x69) available in the PRoNTo directory (PRoNTo/atlas/).
• In the field ‘Use one kernel per modality’, select the ‘No’ option.
After all this steps, the batch editor should look similar to the one in Figure 14.15.

130

CHAPTER 14. MULTIPLE KERNEL LEARNING EXAMPLE

Figure 14.14: Feature set / Kernel module in matlabbatch. This window is called to establish a
dependency connection with the previous ‘Data and design’ module.

Figure 14.15: Feature set / Kernel module. Selected parameters in the Modality option.

14.2. BATCH ANALYSIS

131

Figure 14.16: Specify model module in matlabbatch.

14.2.3

Specify model

• Click on the ‘Specify model’ option on PRoNTo’s matlabbatch menu (Figure 14.16);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Feature set / Kernel’ step or click on the ‘Select
files’ button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the model, e.g. ‘mklFacesHouses’;
• Leave the ‘Use kernels’ field as it is, i.e. ‘Yes’;
• In the ‘Feature sets’ field, select the feature set name with the ‘Dependency’ button6 ;
• Select the ‘Classification’ model type:
– Create 2 new classes;
– For Class (1) write ‘Faces’ on the name field and add one group. Select the group
name from the ‘Data & Design’ module (‘Data & Design:Group#1 name’) with the
‘Dependency’ button7 . Similarly, for Class (2) write ‘Houses’ on the name field and
add the group created in the ‘Data & Design’ module, ‘G1’;
– In the ‘Subjects’ field, type ‘1’ (only subject one is selected);
– In the ‘Conditions / Scans’ field, select the ‘Specify Conditions’ option and add a new
condition. Provide a name for this condition, i.e. for Class (1) ‘Faces’ and for Class (2)
‘Houses’. Note that this name need to be exactly as specified in the ‘Data & Design’
module: if you simply loaded an ‘SPM.mat’ file for the design, you must know the
names of the conditions;
– After all this steps, the batch editor should look similar to the Figure 14.17
5 Or

type it in manually, ‘fMRI’, but the name needs to be exactly the same as the one specified in the ‘Data
& Design’ module.
6 or write it exactly as previously defined in the ‘Feature set / Kernel’ module (option ‘Feature set/Kernel:
Feature/kernel name’), here ‘HaxbyFeatures’.
7 Or write it exactly as previously defined in the Data & Design’ module, here ‘G1’

132

CHAPTER 14. MULTIPLE KERNEL LEARNING EXAMPLE

Figure 14.17: Specify model module. Selected parameters in the Class option.
• In the ‘Machine’ field:
– Select ‘L1 Multi-Kernel Learning’ option;
– In the ‘Optimize hyper-parameter’ field, select the ‘Yes’ option;
– In the ‘Arguments’ field, input the hyper-parameters that you wish to use for the
hyper-parameter optimization (e.g. [0.01, 0.1, 1, 10, 100, 1000]);
– In the ‘Cross validation type for hyper-parameter optimization’ (internal loop) field,
select the ‘k-folds CV on blocks’ option and on the field ‘k’ input the value 4;
• In the ‘Cross validation type’ (external loop) field, select ‘Leave one block out’ option;
• Leave the ‘Include all scans’ field as it is, i.e. ‘No’;
• In the ‘Data operations’ field:
– Leave the ‘Mean centre features’ field as it is, i.e. ‘Yes’;
– Click on the ‘Other Operations’ field and select the option ‘Select Operations’, then
add a new operation and select the ‘Normalize samples’ option;

14.2.4

Run model

• Click on the ’Run model’ option on PRoNTo’s matlabbatch menu (Figure 14.18);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Specify model’ step;
• Select the model name from the ‘Specify model’ module with the ‘Dependency’ button8 ;
• Leave the ‘Do permutation test?’ field as it is, i.e. ‘No permutation test’.

14.2. BATCH ANALYSIS

Figure 14.18: Run model module in matlabbatch.

Figure 14.19: Compute weights module in matlabbatch.

133

134

CHAPTER 14. MULTIPLE KERNEL LEARNING EXAMPLE

14.2.5

Compute weights (optional step)

• Click on the ‘Compute weights’ option on PRoNTo’s matlabbatch menu (Figure 14.19);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Run model’ step;
• Select the model name from the ‘Specify model’ module with the ‘Dependency’ button;
• It’s optional to define a name for the image;
• In the ‘Build weight images per ROI’, load the atlas that was used for building the feature
set, which can be found in PRoNTo/atlas/;
• Leave the ‘Build weights images for permutations’ field as it is, i.e. ‘No’;
Finally, save the batch (e.g. as batch run all.m) and click on the ‘Run Batch’ option, in the
‘File menu’. The batch file created can then be opened and edited for further analyses. The
results will be the same as those obtained using the GUI (see Section 14.1.6 of this tutorial).

8 Or

write it exactly as previously defined in the ‘Specify model’ module, here ‘mklFacesHouses’

Chapter 15

Removing confounds: a
classification example
Contents
15.1
15.2

Introduction . . . . . . . .
GUI analysis . . . . . . . .
15.2.1 Data & Design . . .
15.2.2 Prepare feature set
15.2.3 Specify model . . .
15.2.4 Display results . . .
15.3
Batch analysis . . . . . . .
15.3.1 Data & Design . . .
15.3.2 Feature set / Kernel
15.3.3 Specify model . . .
15.3.4 Run model . . . . .
15.3.5 Display weights . .

15.1

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

. 135
. 136
. 136
. 138
. 138
. 139
. 140
. 140
. 141
. 142
. 143
. 144

Introduction

This chapter will describe the steps necessary to remove confounding effects using PRoNTo.
Confounds (also known as nuissance variables) are included into predictive variables to improve
the statistical significance as well as classification performance. Age, gender or the site where the
medical imaging were acquired (in case of a multi-site study) are examples of what can be used
to assure that the results obtained are real and not due to other secondary variables.
The dataset used in this chapter can be found in OASIS’s website http://www.oasis-brains.
org (data set 1) and consists of a longitudinal collection of 150 subjects aged 60 to 96. Each
subject was scanned on two or more visits, separated by at least one year for a total of 373
imaging sessions. For each subject, 3 or 4 individual T1-weighted MRI scans obtained in single
scan sessions are included. The subjects are all right-handed and include both men and women.
72 of the subjects were characterized as nondemented throughout the study. 64 of the included
subjects were characterized as demented at the time of their initial visits and remained so for
subsequent scans, including 51 individuals with mild to moderate Alzheimers disease. Another 14
subjects were characterized as nondemented at the time of their initial visit and were subsequently
characterized as demented at a later visit. In this example, 95 subjects were selected from OASIS
dataset with 56 demented patients and 39 nondemented subjects. Age is considered as the
confounds.
Firstly, we will use PRoNTo to predict if the subject suffers from dementia or not based on
MRI scans. We will classify the whole brain images using Support Vector Machines, so the reader
is advised to complete the tutorial in Chapter 12 before moving on. As in previous Chapters, we
135

136

CHAPTER 15. REMOVING CONFOUNDS: A CLASSIFICATION EXAMPLE

will start analyzing the data with the PRoNTO’s GUI and then repeating the analysis using the
matlabbatch system. Please create a folder in your computer to store the results and type

15.2

GUI analysis

We will first analyse the data using PRoNTo’s GUI and then repeat the analysis using the
matlabbatch system. We now start up MATLAB and type ‘prt’ or ‘pronto’ in the MATLAB
prompt. This will open the main interface of PRoNTo (see Figure 12.1 in Chapter 12).

15.2.1

Data & Design

• In PRoNTo’s main window, click on ‘Data & Design’ and a new window will open, ‘Data
and design’ (Figure 12.2). Like in previous chapters, browse the directory in which to save
the PRT structure (saved as ‘PRT.mat’);
• In the panel Groups, click on Add and provide a name to the group, e.g. ‘Dem’. Since in
this example we have two groups, it is necessary to add one more called ‘Non-Dem’ after
inputting all information about the ’Dem’ group;
• All the images in the dataset correspond to different subjects; therefore, click on the Scans
tick box. This will lock the Subjects/Scans field, allowing you to skip to the third field. See
Chapter 12 of the manual for more information on this option;
• In the ‘Modalities’ panel, click on ‘Add’ and provide a name to the modality, e.g. ‘MRI’.
• Load all the image files belonging the first group. You can select all the files by using the
right mouse button and clicking on the option ‘Select All’. When all the images are selected,
click on the ‘Done’ button;
• In the ‘Covariates’ field, type the covariates you want to regress out. Or you can save all the
covariates in a matrix named R and input the full path of this matrix to Covariate. Each
variable is a column, hence if there is only one covariate, R is a column vector;if there are
multiple variables, R will be a matrix containing multiple columns where each column is a
covariate. Users can give this matrix any saving name, but after loading into MATLAB,
this matrix should only be R.mat. In this example, ‘Age’ is the confounding factor chosen
(Figure 15.1);

Figure 15.1: ‘Specify modality’ GUI allows one to enter the covariates to be regressed out.
• In the ‘Masks’ field, on the bottom left of the ‘Data and design window, select the ‘mergedsmwrc1 mask 01’
mask for the modality specified (Figure 15.2).

15.2. GUI ANALYSIS

137

Figure 15.2: This window is called when one clicks ‘Masks’.

• Finally, repeat the previous steps for ‘Non demented’ group, including the corresponding
images and covariates, and using the same modality and mask as for ‘Demented’ group. The
‘Data and design’ window should look similar to Figure 15.3. Click on the ‘Save’ button to
create ‘PRT.mat’ file with the structure containing the information that has been previously
specified. If no errors are shown in the MATLAB command, leave the ‘Data and design’
window by clicking ‘Quit’.

Figure 15.3: ‘Data and design’ GUI final configuration..

138

CHAPTER 15. REMOVING CONFOUNDS: A CLASSIFICATION EXAMPLE

15.2.2

Prepare feature set

• In PRoNTo’s main window, click on ‘Prepare feature set’ and a new window will open,
‘Prepare feature set’ (see Figure 12.9 in Chapter 12);
• Select the ‘PRT.mat’ file previously created in the ‘Data & Design’ step and another window
will open, ‘Specify modality to include’ (Figure 15.4);

Figure 15.4: ‘Specify modality to include’ GUI.
• In the ‘Prepare feature set’ window, provide a name for the feature set, e.g. ‘svmDemNonDem’;
• Click on ‘Build kernel/data matrix’ to build the feature set and kernel.

15.2.3

Specify model

• In PRoNTo’s main window, click on ‘Specify model’ and a new window will open, ‘Specify
model’ (see Figure 12.12 in Chapter 12);
• Select the ‘PRT.mat’ file and provide a name to the model, e.g. ‘svm’;
• Select one of the ‘Feature Set’ previously defined. In this case, there is only one: ‘svmDemNonDem’;
• Leave the option ‘Use kernels’ tick box as it is, i.e. ‘Yes’;
• Select the ‘Classification’ model type and click on ‘Define classes‘ button. A new window
will open, ‘Specify classes’ (Figure 15.5), to define the number of classes and a name for
each class. We will define 2 classes. For ‘Class 1’ select group ‘Dem’, and all subjects and,
similarly, for ‘Class 2’ select group ‘Non-Dem’ and all the subjects from this group.
• Select the ‘Binary support vector machine’ option, in the ‘Machine’ field;
• Leave the option ‘Optimize hyper-parameter’ tick box unchecked and ‘Cross-Validation
Scheme’ (internal loop) as it is;
• Select the ‘Leave One Subject Out’ cross-validation scheme (external loop);
• In the ‘Data operations’ box, select the ‘Regress out covariates (subject level)’ option, which
corresponds to the removal of the contribution of some external variables to the data, and
‘Mean centre features using training data’ option. Then, the ‘Specify model’ window should
look similar to Figure 16.5. Click on the ‘Specify and run model’ button;

15.2. GUI ANALYSIS

139

Figure 15.5: ‘Specify classes’ GUI.

Figure 15.6: ‘Specify model’ GUI final configuration.

15.2.4

Display results

• In PRoNTo’s main window, click on ‘Display results’ and select the ‘PRT.mat’ file. This
will open the main results window (Figure 15.7);
• In the ‘Results’ window, one can select a different plot in the ‘Plots’ list;

140

CHAPTER 15. REMOVING CONFOUNDS: A CLASSIFICATION EXAMPLE

Figure 15.7: ‘Results’ GUI.

15.3

Batch analysis

In this section, the previous experiment will be repeated using the ‘matlabbatch’ system. The
reader is advised to complete the tutorial in Section 12.2 before continuing, since the explanation
of each step will be less descriptive.
Once again, to analyse the data, create a new directory in which to save the results of the
analysis. On the main interface of PRoNTo click on the ‘Batch’ button to open the ‘matlabbatch’.
Alternatively, type ‘prt batch’ in the MATLAB prompt.

15.3.1

Data & Design

• Click on ‘Data & Design’ in the PRoNTo menu (see Figure 12.26 in Chapter 12);
• In the ‘Directory’ field, select a directory where the ‘PRT.mat’ file will be saved;
• In the ‘Groups’ field:
– Add two groups;
– In the field ‘Name’, provide a name without spaces to both groups, e.g. ‘Dem’ and
’Non-Dem’;
– In the field ‘Select by’, select the ‘Scans’ option and add a new modality. For more
information on the Scans option please consult Chapter 2;
– Add one modality for each subject and provide a name, e.g. ‘MRI’; and in the field
‘Scans’, select the proper images for each of the groups;
– Specify in the ‘Covariates’ the file with the confounds you want to remove. Remember
that this file should contain a variable ‘R’ with a matrix of covariates. The batch
editor should look similar to the Figure 15.8;

15.3. BATCH ANALYSIS

141

Figure 15.8: Data and design module in matlabbatch.
• In the ‘Masks’ field, add a new modality and provide the same modality name, ‘MRI’; and
select the ‘mergedsmwrc1 mask 01’ mask available of the Oasis dataset. The name of the
modality here has to be exactly the same as in ‘Modalities’, otherwise it will not work;
• Leave the ‘HRF overlap’ and the ‘HRF delay’ fields as default;
• In the ‘Review’ field, select ‘Yes’ if you would like to review your data and design in a
separate window. Otherwise, leave as it is, i.e. ‘No’.

15.3.2

Feature set / Kernel

• Click on the ‘Feature set / Kernel’ option on PRoNTo’s matlabbatch menu (see Figure
12.29 in Chapter 12);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Data & Design’ step (Figure 12.30) or click on the
‘Select files’ button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the ‘Feature/kernel’ set, e.g. ‘svmDem-NonDem’;
• Add one modality and select the modality name with the ‘Dependency’ button1 (Data &
Design:Mod#1 name);
– In the ‘Scans/Conditions’ field , select the ‘All scans’ option;
– In the ‘Voxels to include’ field, select ‘All voxels’ option, this means we are not entering
an additional second-level mask;
– In the ‘Detrend’ field, select ‘Polynomial detrend’ option with order 1;
– In the ‘Scale input scans’ field, select ‘No scaling’ option;
1 Or type it in manually, ‘MRI’, but the name needs to be exactly the same as the one specified in the ‘Data &
Design’ module.

142

CHAPTER 15. REMOVING CONFOUNDS: A CLASSIFICATION EXAMPLE

– Leave ‘Load Atlas’ as default. After all these steps, the batch editor should look similar
to the one in Figure 15.9;

Figure 15.9: Feature set / Kernel module. Selected parameters in the Modality option.
• Leave the ‘Generate Multiple Kernels’ and the ‘Use one kernel per modality’ fields as default.

15.3.3

Specify model

• Click on the ‘Specify model’ option on PRoNTo’s matlabbatch menu (see Figure 12.32 in
Chapter 12);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Feature set / Kernel’ step or click on the ‘Select
files’ button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the model, e.g. ‘SVM’;
• Leave the ‘Use kernels’ field as it is, i.e. ‘Yes’;
• Select the feature set name with the ‘Dependency’ button2 ;
• Select the ‘Classification’ model type:
– Add 2 new classes;
– For Class (1) write ‘Demented’ on the name field and add one group. Select the group
name from the ‘Data & Design’ module (‘Data & Design:Group#1 name’) with the
‘Dependency’ button3 . Similarly, for Class (2) write ‘Non-Demented’ on the name field
and add the group created in the ‘Data & Design’ module, ‘Non-Dem’;
– In the ‘Subjects’ field of the Group 1, type ‘1:56’. This will instruct the program to
use all the 56 scans, i.e. from scan 1 to scan 56. Similarly, in the ‘Subjects’ field of
the Group 2, type ‘1:39’.
2 or write it exactly as previously defined in the ‘Feature set / Kernel’ module (option ‘Feature set/Kernel:
Feature/kernel name’), here ‘svmDem-NonDem’.
3 Or write it exactly, as previously defined in the Data & Design’ module, here ‘Dem’

15.3. BATCH ANALYSIS

143

• In the ‘Machine’ field:
– Select the ‘SVM Classification’ option;
– Leave the ‘Optimize hyper-parameter’ field as it is, i.e. ‘No’;
– Leave the ‘Soft-margin hyper-parameter’ field as it is, i.e. ‘1’;
– Leave the ‘Cross validation type for hyper-parameter optimization’ field as it is, i.e.
‘Leave one subject out’;
• In the ‘Cross-validation type’ field, select ‘Leave one subject out’ option;
• Leave the ‘Include all scans’ field as it is, i.e. ‘No’;
• In the ‘Data operations’ field:
– Leave the ‘Mean centre features’ field as it is, i.e. ‘Yes’;
– Click on the ‘Other Operations’ button, choose ‘Select operations’, ‘New Operation’
and select the ‘Regress out covariates (subject level)’ option from the list. After all
these steps, the batch editor should look similar to the one in Figure 15.10;

Figure 15.10: Data and design module in matlabbatch.

15.3.4

Run model

• Click on the ‘Run model’ option on PRoNTo’s matlabbatch menu (see Figure 12.35 in
Chapter 12);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Specify model’ step;
• Select the model name from the ‘Specify model’ module with the ‘Dependency’ button4 ;
• In the field ‘Do permutation test?’, leave as it is, i.e. ‘No permutation test’
4 or

write it exactly as previously defined in the ‘Specify model’ module, here ‘SVM’

144

CHAPTER 15. REMOVING CONFOUNDS: A CLASSIFICATION EXAMPLE

15.3.5

Display weights

• In PRoNTo’s main window, click on ‘Display weights’ and select the ‘PRT.mat’ file. This
will open the ‘Model interpretation’ window.
• By clicking on ‘Model’, SVM, an image will appear in the ‘Weights map’ box; and to show
the ‘Anatomical img’ you have to load an anatomical image for reference. A template image
can be found in SPM’s canonical folder (‘single subj T1‘ file). The final window will look
similar to the one shown in the Figure 15.11.

Figure 15.11: Weights map

Chapter 16

Removing confounds: a regression
example
Contents
16.1
16.2

Introduction . . . . . . . . . . . .
GUI analysis . . . . . . . . . . . .
16.2.1 Data & Design . . . . . . .
16.2.2 Prepare feature set . . . .
16.2.3 Specify model . . . . . . .
16.2.4 Display results . . . . . . .
16.2.5 Compute weights (optional
16.3
Batch analysis . . . . . . . . . . .
16.3.1 Data & Design . . . . . . .
16.3.2 Feature set / Kernel . . . .
16.3.3 Specify model . . . . . . .
16.3.4 Run model . . . . . . . . .

16.1

. . .
. . .
. . .
. . .
. . .
. . .
step)
. . .
. . .
. . .
. . .
. . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

. 145
. 145
. 145
. 146
. 147
. 148
. 148
. 148
. 149
. 152
. 153
. 153

Introduction

This chapter will describe the steps necessary to remove confounding effects using PRoNTo in a
regression approach. These are similar to the ones used in the previous chapter, thus, the reader
is advised to complete the tutorial in Chapter 15 before moving on, since the explanation of some
steps will be less descriptive. The dataset used in this chapter can be found on PRoNTo’s website
http://www.mlnl.cs.ucl.ac.uk/pronto/prtdata.html (data set 3).
As in previous Chapters, we will start analyzing the data with the PRoNTo’s GUI and then
repeating the analysis using the matlabbatch system. Please create a folder in your computer to
store the results and type

16.2

GUI analysis

We will first analyse the data using PRoNTo’s GUI and then repeat the analysis using the
matlabbatch system, then start up MATLAB and type ‘prt or ‘pronto in the MATLAB prompt.
This will open the main interface of PRoNTo (see Figure 12.1 in Chapter 12).

16.2.1

Data & Design

• In PRoNTo’s main window, click on ‘Data & Design’ and a new window will open, ‘Data
and design’ (Figure 12.2). Like in previous chapters, browse the directory in which to save
the PRT structure (saved as ‘PRT.mat’);
145

146

CHAPTER 16. REMOVING CONFOUNDS: A REGRESSION EXAMPLE

• In the panel Groups, click on Add and provide a name to the group, e.g. ‘Aged’.
• All the images in the dataset correspond to different subjects; therefore, click on the Scans
tick box. This will lock the Subjects/Scans field, allowing you to skip to the third field. See
Chapter 12 of the manual for more information on this option;
• In the ‘Modalities’ panel, click on ‘Add’ and provide a name to the modality, e.g. ‘MRI’.
• Load all the image files available in the directory (IXIdata/aged/Guys/). You can select all
the files by using the right mouse button and clicking on the option ‘Select All’. When all
the images are selected, click on the ‘Done’ button;
• In the ‘Covariates’ field, type the covariates you want to regress out. In this example,
‘Gender’ is the one chosen. As in the previous chapter, covariates can be input to PRoNTo
by directly copying values into the Covariate if there is only one confounding factor, or by
typing the full path of the R matrix containing either one column or multiple columns;
• In the ‘Regression targets’ field, write (or paste) the list of target values available in the
‘Age old Guys’ file (IXIdata/aged/), or input the full path of a rt subj.mat file containing
the target values in a column. The final window should look like to the Figure 16.1. Press
‘OK’;

Figure 16.1: ‘Specify modality’ GUI allows one to enter the covariates to be regressed out.
• In the ‘Masks’ field, on the bottom left of the ‘Data and design’ window, select the
‘mergedsmwrc1 mask 01’ mask for the modality specified (Figure 12.6);
• The ‘Data and design’ window should look similar to Figure 16.2. Click on the ‘Save’
button to create ‘PRT.mat’ file with the structure containing the information that has been
previously specified. If no errors are shown in the MATLAB command, leave the ‘Data and
design’ window by clicking ‘Quit’.

16.2.2

Prepare feature set

• In PRoNTo’s main window, click on ‘Prepare feature set’ and a new window will open,
‘Prepare feature set’ (see Figure 12.9 in Chapter 12);
• Select the ‘PRT.mat’ file previously created in the ‘Data & Design’ step and another window
will open, ‘Specify modality to include’ (see Figure 15.4 in Chapter 15);
• In the ‘Prepare feature set’ window, provide a name for the feature set, e.g. ‘KRRAged’;
• Click on ‘Build kernel/data matrix’ to build the feature set and kernel.

16.2. GUI ANALYSIS

147

Figure 16.2: ‘Data and design’ GUI final configuration..

16.2.3

Specify model

• In PRoNTo’s main window, click on ‘Specify model’ and a new window will open, ‘Specify
model’ (see Figure 12.12 in Chapter 12);
• Select the ‘PRT.mat’ file and provide a name to the model, e.g. ‘KRR’;
• Select one of the ‘Feature Set’ previously defined. In this case, there is only one: ‘KRRAged’;
• Leave the option ‘Use kernels’ tick box as it is, i.e. ‘Yes’;
• Select the ‘Regression’ model type and click on the ‘Select subjects/scans‘ button. This
will open a new window, ‘Specify subjects/scans to regress’, click on the ‘Select all’ button
to use all the scans for the regression (Figure 16.3);

Figure 16.3: ‘Specify classes’ GUI.

148

CHAPTER 16. REMOVING CONFOUNDS: A REGRESSION EXAMPLE

• Select the ‘Kernel Ridge Regression’ option, in the Machine field;
• Check the option ‘Optimize hyper-parameter’ tick box (we will see later the importance of
optimizing the hyperparameter when the confounds are removed). Type ‘10.ˆ[-5:5]’ on the
right field. Then, select ‘k-fold CV on Subject-Out’ option in the ‘Cross-Validation Scheme’
field (internal loop). After that, type ‘5’ in the new window that will open (see Figure 16.4);

Figure 16.4: Number of folds selected in the internal loop.
• Select the ‘Leave One Subject Out’ cross-validation scheme (external loop);
• In the ‘Data operations’ box, select the ‘Regress out covariates (subject level)’ option, which
corresponds to the removal of the contribution of some external variables to the data, and
‘Mean centre features using training data’ option. Then, the ‘Specify model’ window should
look similar to Figure 16.5. Click on the ‘Specify and run model’ button;

16.2.4

Display results

• In PRoNTo’s main window, click on ‘Display results’ and select the ‘PRT.mat’ file. This
will open the main results window (Figure 16.6);
• In the ‘Results’ window, one can select the different regression models in the ‘Model’ list on
the upper right region. This will show the results obtained using each one of the regression
models.
• It is important to optimize the hyperparameter when we try to remove confounds. From our
experience, results without the hyperparameter optimization may be significantly different
from results with hyperparameter optimization. But this effect may be data dependent.
We therefore highly suggest users choose to perform this step.

16.2.5

Compute weights (optional step)

• In PRoNTo’s main window, click on ‘Compute weights’ and a new window will open, ‘Compute weights’ (Figure 16.7);
• Select the ‘PRT.mat’ file;
• Select the model from the list to ‘Models computed in PRT’, ‘KRR’ model;
• Leave the option ‘Compute average/kernel weight per region’ tick box unchecked;
• Click on ‘Compute weights’ button. Computations will be displayed on the MATLAB
command window.

16.3

Batch analysis

In this section, the previous experiment will be repeated using the ‘matlabbatch’ system. The
reader is advised to complete the tutorial in Section 12.2 before continuing, since the explanation
of each step will be less descriptive.

16.3. BATCH ANALYSIS

149

Figure 16.5: ‘Specify model’ GUI final configuration.
Once again, to analyse the data, create a new directory in which to save the results of the
analysis. On the main interface of PRoNTo click on the ‘Batch’ button to open the ‘matlabbatch’.
Alternatively, type ‘prt batch’ in the MATLAB prompt.

16.3.1

Data & Design

• Click on ‘Data & Design’ in the PRoNTo menu (see Figure 12.26 in Chapter 12);
• In the ‘Directory’ field, select a directory where the ‘PRT.mat’ file will be saved;
• In the ‘Groups’ field:
– Add one group;
– In the field ‘Name’, provide a name without spaces for this group, e.g. ‘Aged’;
– In the field ‘Select by’, select the ‘Scans’ option and add a new modality. For more
information on the Scans option please consult Chapter 2;

150

CHAPTER 16. REMOVING CONFOUNDS: A REGRESSION EXAMPLE

Figure 16.6: ‘Results’ GUI.
– Provide a name for this modality, e.g. ‘MRI’; select the image files available in the
‘aged/Guys’ directory of the IXI dataset and write (or paste) the regression targets1
in the ‘Regression targets (per scans)’ field;
– Specify in the ‘Covariates’ the file with the confounds you want to remove. Remember that this file should contain a variable ‘R’ with a matrix of covariates (Data &
Design:Mod#1 name) 2 . The batch editor should look similar to the Figure 16.8;
• In the ‘Masks’ field, add a new modality and provide the same modality name, ‘fMRI’; and
select the ‘whole brain’ mask available in the masks directory of the Haxby dataset. The
name of the modality here has to be exactly the same as in ‘Modalities’, otherwise it will
not work;
1 Available

in the ‘Age old Guys’ file (IXIdata/aged/)
is possible to regress out more than one covariate just adding to the variable ‘R’ as many columns as the
number of covariates
2 It

16.3. BATCH ANALYSIS

151

Figure 16.7: ‘Compute weights’ GUI.

Figure 16.8: Data and design module in matlabbatch.

152

CHAPTER 16. REMOVING CONFOUNDS: A REGRESSION EXAMPLE

• Leave the ‘HRF overlap’ and the ‘HRF delay’ fields as default;
• In the ‘Review’ field, select ‘Yes’ if you would like to review your data and design in a
separate window. Otherwise, leave as it is, i.e. ‘No’.

16.3.2

Feature set / Kernel

• Click on the ‘Feature set / Kernel’ option on PRoNTo’s matlabbatch menu (see Figure
12.29 in Chapter 12);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Data & Design’ step (Figure 12.30) or click on the
‘Select files’ button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the ‘Feature/kernel’ set, e.g. ‘KRRAged’;
• Add one modality and select the modality name with the ‘Dependency’ button3 (Data &
Design:Mod#1 name);
– In the ‘Scans/Conditions’ field , select the ‘All scans’ option;
– In the ‘Voxels to include’ field, select ‘All voxels’ option, this means we are not entering
an additional second-level mask;
– In the ‘Detrend’ field, select ‘Polynomial detrend’ option with order 1;
– In the ‘Scale input scans’ field, select ‘No scaling’ option;
– Leave ‘Load Atlas’ as default. After all these steps, the batch editor should look similar
to the one in Figure 16.9;

Figure 16.9: Feature set / Kernel module. Selected parameters in the Modality option.
• Leave the ‘Generate Multiple Kernels’ and the ‘Use one kernel per modality’ fields as default.
3 Or

type it in manually, ‘MRI’, but the name needs to be exactly the same as the one specified in the ‘Data &
Design’ module.

16.3. BATCH ANALYSIS

16.3.3

153

Specify model

• Click on the ‘Specify model’ option on PRoNTo’s matlabbatch menu (see Figure 12.32 in
Chapter 12);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Feature set / Kernel’ step or click on the ‘Select
files’ button to browse where ‘PRT.mat’ file was saved;
• Provide a name to the model, e.g. ‘KRR’;
• Leave the ‘Use kernels’ field as it is, i.e. ‘Yes’;
• Select the feature set name with the ‘Dependency’ button4 ;
• Select the ‘Regression’ model type:
– Add a new group and call it ‘Aged’;
– In the ‘Subjects’ field, type ‘1:102’. This will instruct the program to use all the 102
scans, i.e. from scan 1 to scan 102;
• In the ‘Machine’ field:
– Select the ‘Kernel Ridge Regression’ option:
– Change the ‘Optimize hyper-parameter’ field as to ‘Yes’; Type ‘10.ˆ[-5:5]’ on the
‘Regularization’ field. Then, select ‘k-fold CV on Subject-Out’ option in the ‘CrossValidation Scheme’ field (internal loop). After that, type ‘5’ in ‘k’ field;
• In the ‘Cross-validation type’ field, select ‘Leave one subject out’ option;
• Leave the ‘Include all scans’ field as it is, i.e. ‘No’;
• In the ‘Data operations’ field:
– Leave the ‘Mean centre features’ field as it is, i.e. ‘Yes’;
– Click on the ‘Other Operations’ button, choose ‘Select operations’, ‘New Operation’
and select the ‘Regress out covariates (subject level)’ option from the list. After all
these steps, the batch editor should look similar to the one in Figure 16.10;

16.3.4

Run model

• Click on the ‘Run model’ option on PRoNTo’s matlabbatch menu (see Figure 12.35 in
Chapter 12);
• With ‘Load PRT.mat’ field selected, click on the ‘Dependency’ button to associate the
‘PRT.mat’ file created in the previous ‘Specify model’ step;
• Select the model name from the ‘Specify model’ module with the ‘Dependency’ button5 ;
• In the field ‘Do permutation test?’, leave as it is, i.e. ‘No permutation test’

4 or write it exactly as previously defined in the ‘Feature set / Kernel’ module (option ‘Feature set/Kernel:
Feature/kernel name’), here ‘KRRAged’.
5 or write it exactly as previously defined in the ‘Specify model’ module, here ‘KRR’

154

CHAPTER 16. REMOVING CONFOUNDS: A REGRESSION EXAMPLE

Figure 16.10: Data and design module in matlabbatch.

Part IV

Advanced topics

155

Chapter 17

Developer’s guide
Contents

17.1

17.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

17.2

Code organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
17.2.1

User interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

158

17.2.2

Machine learning

. . . . . . . . . . . . . . . . . . . . . . . . . . . .

158

17.2.3

Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

158

Introduction

As described in the Introduction, PRoNTo was developed using a modular structure. There are
five main modules (‘Data and Design’, ‘Prepare Feature set’, ‘Specify model’, ‘Run model’ and
‘Compute weights’) and three reviewing and displaying modules (‘Review data’, ‘Review kernel
and CV’ and ‘Display results’). This structure not only facilitates the use of the toolbox but
also new contributions from developers. These modules have very few dependencies between each
other, and for most of them to work, one needs only to provide the ‘PRT.mat’ obtained from
the previous module and a few more module specific inputs. This means that the developer can
contribute with code for any of the modules without having to adapt the whole toolbox to the
changes. Developers can also work only on the module of interest and do not need to be familiar
with the functions and sub functions that comprise the rest of the toolbox. In this chapter, we
provide a brief description of how the code is organised and how one can contribute with new code,
in particular new GUIs and Batch functions. At the end of this chapter, we provide instructions
on how to integrate a new machine into PRoNTo. In the current version of PRoNTo, this is the
most straightforward extension that can be added.

17.2

Code organisation

Although from the user’s point of view there are five main modules, from the developer’s side one
can see PRoNTo’s functions as belonging to three categories, depending on what they deal with
(Figure 17.1). The first set of functions and sub-functions is responsible for creating PRoNTo’s
GUIs and matlabbatch menu. The second set of functions comprise all the core routines that
implement the machine learning methods (including extraction and preparation of the features,
specification and estimation of a model, cross-validation, etc.). The last set of functions correspond to the actual machine learning algorithms that PRoNTo uses for classification and regression. We call these functions the ‘machines’. The main difference between this last set of functions
and the rest, is that one does not need to be familiar with PRoNTo’s PRT.mat structure or the
rest of the code to be able to integrate a new machine. This can be done very easily, as shown
below.
157

158

CHAPTER 17. DEVELOPER’S GUIDE

Figure 17.1: Organisation of PRoNTo’s functions. From the developer’s perspective the code is
organised in functions and sub functions that deal with the user interface experience (such as
GUI and batch functions), the core functions that implement the machine learning approaches
(including extracting and preparing the features, specification and estimation of a model, crossvalidation routines, etc.) and the actual machine learning algorithms (also known as machines).

17.2.1

User interface

The functions that deal with creating and running PRoNTo’s main GUIs have the prefix ‘prt ui’.
All GUI functions (‘.m’ files) have a corresponding ‘.fig’ file. This file can be opened and edited
using MATLAB’s guide functionality. This way one can change the GUI and add extra options
to the available menus (for more information, please consult MATLAB’s documentation1 ).
PRoNTo’s matlabbatch functions have either the prefix ‘prt cfg’, to create a new menu on
the batch interface and the prefix ‘prt run’, to execute instructions using the variables from the
corresponding ‘prt cfg’ function. PRoNTo’s batch functionalities work exactly like SPM. For more
information on how to contribute with new matlabbatch functions please consult the developer’s
guide on the SPM8 manual2 .

17.2.2

Machine learning

The machine learning routines comprise the core of PRoNTo. These routines do most of the necessary instructions to run all machine learning procedures currently implemented in the toolbox.
They don’t have a specific prefix but from the name of the .m file it is easy to find out what the
function does (e.g. prt compute weights.m deals with creating new weights images). Most of the
functions take as input the PRT.mat structure. Therefore, knowledge of this structure (see next
chapter) is required in order to contribute with new code. In the future, we intend to make the
process of introducing new feature selection algorithms and other model functionalities as easy
as implementing a new machine, as described below.

17.2.3

Machines

Integrating a new machine algorithm, i.e. classifier or regressor, into the PRoNTo framework is
straightforward. PRoNTo provides a function called ‘prt machine.m’ that works as a wrapper
around the different machine learning algorithms (functions with prefix ‘prt machine’). In brief,
this wrapper translates PRoNTo’s structure and internal variables into the required inputs for
each machine. This way, the machine function needs only to read a simple input format that does
not depend on knowledge of how PRoNTo works or of the PRT.mat fields (Figure 17.2).
More specifically, to contribute with a new machine the developer needs to provide a matlab
function, which reads the following input data structure, d, and optional arguments, args, (all
1 http://www.mathworks.com/help/techdoc
2 http://www.fil.ion.ucl.ac.uk/spm/doc/manual.pdf

17.2. CODE ORGANISATION

159

Figure 17.2: Code organisation for integrating a new machine with PRoNTo. ‘prt machine.m’
works has a wrapper that translates PRoNTo’s inputs into the machine format inputs and performs extensive error checks on these variables. The machines (such as ‘prt machine svm bin.m’)
perform the classifier/regression algorithms and have inputs and outputs that are not dependent
on PRoNTo’s internal structures.

fields are mandatory except where otherwise stated):
• d - data structure with input to the classifier/regressor:
.train - training data (cell array of matrices of row vectors, each [Ntr x D]). Each matrix
contains one representation of the data. This is useful for approaches such as multiple kernel
learning. Ntr is the number of training examples and D is the dimension of the feature set
(e.g. number of voxels).
.test - testing data (cell array of matrices row vectors, each [Nte x D]). Nte is the number
of testing examples and D the dimension of the feature set.
.tr targets - training labels (for classification) or values (for regression) (column vector, [Ntr
x 1]).
.use kernel - flag: is data in form of kernel matrices (true) of in form of features (false)?
• args (optional) - anything else that is specific to the algorithm (e.g. LIBSVM arguments).
In addition, the outputs of the function need to have the following format, so that they can
be read by the wrapper and translated back to the PRoNTo framework (all fields are mandatory
except where otherwise stated) (Figure 17.2):
• output - output structure with fields:
.predictions - predictions of classification or regression [Nte x D]. Nte is the number of test
examples and D the dimension of the feature set.
.func val (optional) - value of the decision function (if it applies).
.type (optional) - type of machine (string: ‘classifier’ or ‘regressor’).
The rest of the function can be designed entirely as the developer wishes. The last thing to
have in mind is the name of the function itself. It needs to have the prefix ‘prt machine’ (e.g.
‘prt machine svm bin’ is a function that implements support vector machine binary classification
by calling the LIBSVM library). Importantly, the cross-validation and performance measures are
performed outside in the main PRoNTo framework, and therefore the machine function should
provide only the necessary instructions to implement the classifier/regressor algorithm.

160

CHAPTER 17. DEVELOPER’S GUIDE

Figure 17.3: ‘prt machine .m’ function help. Example of a PRoNTo machine for 2-class SVM
classification.

Finally, the new machine is easily integrated with PRoNTo by including the name of the file in
the corresponding GUI and Batch functions (prt ui model.m and prt cfg model.m, respectively).
The same procedure applies to the weights functions. PRoNTo provides a wrapper function
called ‘prt weights’. The procedure for integrating a new weights function is exactly the same as
for a new machine.
Both wrapper functions, ‘prt machine’ and ‘prt weights’, perform extensive tests to make
sure the machines and weights code complies to the specific inputs and outputs required by the
framework.

Chapter 18

PRT structure
This is how the main PRT structure is organised.
PRT
• group
• gr name
• subject
• subj name()
• modality()
• mod name
• detrend
• covar
• rt subj
• scans
• design
• conds
•
•
•
•
•
•
•
•

cond name()
onsets()
durations()
rt trial()
scans()
blocks()
discardedscans()
hrfdiscardedscans()

•
•
•
•
•
•
•
•

overlap
goodscans
discscans
meanovl
stdovl
mgoodovl
sgoodovl
goodovl

• stats

• TR
• unit
• covar
• hrfoverlap
161

162

CHAPTER 18. PRT STRUCTURE

• hrfdelay
• masks
• mod name
• fname
• fs
• fs name
• k file
• id col names
• fas
• im
• ifa
• modality
•
•
•
•
•
•

mod name
detrend
param dt
mode
idfeat fas
normalise
• type
• scaling

• id mat
• fas
• mod name
• dat
• detrend
• param dt
• hdr
•
•
•
•
•
•
•
•

fname
dim
mat
pinfo
dt
n
descrip
private

• idfeat img
• model
• model name()
• input()
• use kernel
• type
• machine
• function
• args

163

• class
• class name()
• group()
• gr name
• subj
• num()
• modality()
• fs
•
•
•
•
•
•
•

• fs name
samp idx
include allscans
targets
targ allscans
cv mat
operations
cv type

• output()
• fold
• targets()
• predictions()
• stats()
•
•
•
•
•
•
•
• func val()
• type()
• alpha()
• b()
• totalSV()
• stats
•
•
•
•
•
•
•

con mat
acc
c acc
b acc
c pv
acc lb
acc ub

con mat
acc
c acc
b acc
c pv
acc lb
acc ub

164

CHAPTER 18. PRT STRUCTURE

Chapter 19

List of PRoNTo functions
Contents
19.1
19.2
19.3
19.4
19.5
19.6
19.7
19.8
19.9
19.10
19.11
19.12
19.13
19.14
19.15
19.16
19.17
19.18
19.19
19.20
19.21
19.22
19.23
19.24
19.25
19.26
19.27
19.28
19.29
19.30
19.31
19.32
19.33
19.34
19.35
19.36
19.37

pronto.m . . . . . . . . . . .
prt.m . . . . . . . . . . . . .
prt apply operation.m . . .
prt batch.m . . . . . . . . .
prt build region weights.m .
prt check.m . . . . . . . . .
prt check design.m . . . . .
prt check flag.m . . . . . . .
prt comp ranking dist.m . .
prt compute cv mat.m . . .
prt compute weights.m . . .
prt compute weights class.m
prt compute weights regre.m
prt cv fold.m . . . . . . . . .
prt cv model.m . . . . . . .
prt cv opt param.m . . . . .
prt data conditions.m . . . .
prt data modality.m . . . . .
prt data review.m . . . . . .
prt defaults.m . . . . . . . .
prt fs.m . . . . . . . . . . .
prt fs modality.m . . . . . .
prt func2html.m . . . . . . .
prt getKernelModel.m . . .
prt get defaults.m . . . . . .
prt get filename.m . . . . . .
prt init fs.m . . . . . . . . .
prt init model.m . . . . . . .
prt latex.m . . . . . . . . . .
prt load.m . . . . . . . . . .
prt load blocks.m . . . . . .
prt model.m . . . . . . . . .
prt nested cv.m . . . . . . .
prt permutation.m . . . . .
prt plot ROC.m . . . . . . .
prt plot confusion matrix.m
prt plot histograms.m . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

165

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

167
167
167
168
168
168
169
169
170
170
170
170
171
171
172
172
173
173
173
174
174
175
175
176
176
176
177
178
178
179
179
179
180
180
181
181
181

166

CHAPTER 19. LIST OF PRONTO FUNCTIONS

19.38
19.39
19.40
19.41
19.42
19.43
19.44
19.45
19.46
19.47
19.48
19.49
19.50
19.51
19.52
19.53
19.54
19.55
19.56
19.57
19.58
19.59
19.60
19.61
19.62
19.63
19.64
19.65
19.66
19.67
19.68
19.69
19.70

prt plot nested cv.m . . . . . . . . . . . . .
prt plot prediction.m . . . . . . . . . . . . .
prt plot prediction reg bar.m . . . . . . . . .
prt plot prediction reg line.m . . . . . . . .
prt plot prediction reg scatter.m . . . . . . .
prt region histogram.m . . . . . . . . . . . .
prt remove confounds.m . . . . . . . . . . .
prt stats.m . . . . . . . . . . . . . . . . . . .
prt struct.m . . . . . . . . . . . . . . . . . .
prt struct2latex.m . . . . . . . . . . . . . . .
prt text input.m . . . . . . . . . . . . . . . .
prt ui compute weights.m . . . . . . . . . .
prt ui custom CV.m . . . . . . . . . . . . .
prt ui cv model.m . . . . . . . . . . . . . . .
prt ui design.m . . . . . . . . . . . . . . . .
prt ui disp weights.m . . . . . . . . . . . . .
prt ui kernel construction.m . . . . . . . . .
prt ui main.m . . . . . . . . . . . . . . . . .
prt ui model.m . . . . . . . . . . . . . . . .
prt ui prepare data.m . . . . . . . . . . . . .
prt ui prepare datamod.m . . . . . . . . . .
prt ui results.m . . . . . . . . . . . . . . . .
prt ui results ROI.m . . . . . . . . . . . . .
prt ui results help.m . . . . . . . . . . . . .
prt ui results stats.m . . . . . . . . . . . . .
prt ui reviewCV.m . . . . . . . . . . . . . .
prt ui reviewmodel.m . . . . . . . . . . . . .
prt ui select class.m . . . . . . . . . . . . . .
prt ui select reg.m . . . . . . . . . . . . . . .
prt ui specify CV basis.m . . . . . . . . . .
prt ui stats.m . . . . . . . . . . . . . . . . .
prt ui sure.m . . . . . . . . . . . . . . . . . .
machines . . . . . . . . . . . . . . . . . . . .
19.70.1 machines\prt KRR.m . . . . . . . . .
19.70.2 machines\prt machine.m . . . . . . .
19.70.3 machines\prt machine RT bin.m . . .
19.70.4 machines\prt machine gpclap.m . . .
19.70.5 machines\prt machine gpml.m . . . .
19.70.6 machines\prt machine gpr.m . . . . .
19.70.7 machines\prt machine krr.m . . . . .
19.70.8 machines\prt machine rvr.m . . . . .
19.70.9 machines\prt machine sMKL cla.m .
19.70.10 machines\prt machine sMKL reg.m .
19.70.11 machines\prt machine svm bin.m . .
19.70.12 machines\prt rvr.m . . . . . . . . . .
19.70.13 machines\prt weights.m . . . . . . . .
19.70.14 machines\prt weights bin linkernel.m
19.70.15 machines\prt weights gpclap.m . . .
19.70.16 machines\prt weights sMKL cla.m . .
19.70.17 machines\prt weights sMKL reg.m .
19.70.18 machines\prt weights svm bin.m . . .
19.71 utils . . . . . . . . . . . . . . . . . . . . . . .
19.71.1 utils\prt centre kernel.m . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

182
182
183
183
183
183
183
184
184
184
184
185
185
186
186
187
187
188
188
189
189
190
190
190
191
191
192
192
193
193
193
194
194
194
194
195
195
196
197
198
198
199
199
200
200
201
202
202
202
202
203
203
203

19.1. PRONTO.M

19.71.2 utils\prt checkAlphaNumUnder.m . . . . . . . . . . . . . . . . . . .
19.71.3 utils\prt normalise kernel.m . . . . . . . . . . . . . . . . . . . . . .
19.71.4 utils\prt utils update mask.m . . . . . . . . . . . . . . . . . . . . .

167

203
203
203

This is the list of PRoNTo functions, including the subdirectories: machines and utils.

19.1

pronto.m

Function launching PRoNTo (Pattern Recognition for Neuroimaging Toolbox),
see prt.m for more details

19.2

prt.m

Pattern Recognition for Neuroimaging Toolbox, PRoNTo.
This function initializes things for PRoNTo and provides some low level
functionalities

19.3

prt apply operation.m

function to apply a data operation to the training, test and
in.train:
training data
in.tr id:
id matrix for training data
in.use kernel: are the data in kernelised form
in.tr targets: training targets (optional field)
in.pred type: ’classification’ or ’regression’ (required for tr targets)
A test set may
in.test:
in.testcov:
in.te targets:
in.te id:
opid
1
2
3
4
5

also be specified, which require the following fields:
test data
test covariance (only if use kernel = true)
test targets
id matrix for test data

specifies the operation to apply, where:
= Temporal Compression
= Sample averaging (average samples for each subject/condition)
= Mean centre features over subjects
= Divide data vectors by their norm
= Perform a GLM (fMRI only)

N.B: - all operations are applied independently to training and test
partitions
- see Chu et. al (2011) for mathematical descriptions of operations
1 and 2 and Shawe-Taylor and Cristianini (2004) for a description
of operation 3.

168

CHAPTER 19. LIST OF PRONTO FUNCTIONS

References:
Chu, C et al. (2011) Utilizing temporal information in fMRI decoding:
classifier using kernel regression methods. Neuroimage. 58(2):560-71.
Shawe-Taylor, J. and Cristianini, N. (2004). Kernel methods for Pattern
analysis. Cambridge University Press.

19.4

prt batch.m

Pattern Recognition for Neuroimaging Toolbox, PRoNTo.
This function prepares and launches the batch system.
This builds the whole tree for the various tools and their GUI at the
first call to this script.

19.5

prt build region weights.m

19.6

prt check.m

Function to automatically test PRoNTo’s integrity
The goal is to have PRoNTo run through typical analysis and check if the
calculations proceed smoothly.
This relies on pre-specified
- organisation of data in subdirectories
- batches with all the operations, in a .mat file with known location
Data sets considered, in this *specific order*:
1. "Haxby" - Haxby data, single subject, fmri
2. "IXI"
- IXI data, multi subject, divergence & momentum maps
3. "Faces" - SPM’s famous-vs-nonfamous faces data, multi subject.
See the subfunctions for a detailed description of the tests performed.
FORMAT ok = prt check(list check,dir root)
INPUT
list check
dir root

OUTPUT:
ok

- list of data sets to use, [1 2 3] by default
- root directory of data sets (you’d better set this for
your own HD organization!)

- vector of output (1=’ok’, 0=’failed’, -1=’not tested’)

NOTE:
- For a more automatic testing on your own system, then up date the
default ’dir root’ variable with the path to the ’PRoNTo data’
directory on your system.
- This will close all Matlab windows before relaunching PRoNTo and the
matlabbatch system.
WARNING:

19.7. PRT CHECK DESIGN.M

This version was developped for and is running on **SPM12**

19.7

prt check design.m

FORMAT [conds] = prt check design(cond,tr,units,hrfoverlap)
Check the design and discards scans which are either overlapping between
conditions or which do not respect a minimum time interval between
conditions (due to the width of the HRF function).
INPUT
- cond
-

:

structure containing the names, durations and onsets of the
conditions
tr
:
interscan interval (TR)
units :
1 for seconds, 0 for scans
hrfoverlap : value to correct for BOLD overlap (in seconds)
hrfdelay
: value to correct for BOLD delay (in seconds)

OUTPUT
the same cond structure containing supplementary fields:
- scans :
scans retained for further classification
- discardedscans:
scans discarded because they overlapped between
conditions
- hrfdiscardedscans: scans discarded because they didn’t respect the
minimum time interval between conditions
- blocks:
represents the grouping of the stimuli (for
cross-validation)
- stats:
struct containing the original time intervals, the
time interval with only the ’good’ scans, their
means and standard deviation

19.8

prt check flag.m

FORMAT flags = prt check flag(flags o,flags)
Function to automatically check the content of a "flag" structure, using
a "default flag structure", adding the missing fields and putting in the
default value if none was provided.
INPUT:
flags o
flags

OUPUT:
flags

default or reference structure
input flag/option structure that need to be filled for missing
fields with default values

filled flag/option structure

NOTE:
This function was originally named ’crc check flag’ and was distributed
with the FASST toolbox:
http://www.montefiore.ulg.ac.be/~phillips/FASST.html

169

170

19.9

CHAPTER 19. LIST OF PRONTO FUNCTIONS

prt comp ranking dist.m

Function to compute the distance between two ranking vectors, as detailed
in Lampel and Moran, 2005 (in Information Retrieval, 8, 245-264).
INPUT : two ranking vectors of the same size
OUTPUT: their distance

19.10

prt compute cv mat.m

19.11

prt compute weights.m

FORMAT prt compute weights(PRT,in)
This function calls prt weights to compute weights
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
in
- structure with specific information to create
weights
.model name - model name (string)
.img name
- (optional) name of the file to be created
(string)
.pathdir
- directory path where to save weights (same as the
one for PRT.mat) (string)
.atl name
- name of the atlas for post-hoc local averages of
flag
- set to 1 to compute the weight images for each
permutation (default: 0)
flag2
- set to 1 to build image of weight per ROI
weights according to atlas
Output:
img name
- name of the .img file created
+ image file created on disk

19.12

prt compute weights class.m

FORMAT prt compute weights class(PRT,in,model idx)
This function calls prt weights to compute weights
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
in
- structure with specific information to create
weights
.model name - model name (string)
- (optional) name of the file to be created
.img name
(string)
.pathdir
- directory path where to save weights (same as the
one for PRT.mat) (string)
model idx
- model index (integer)

19.13. PRT COMPUTE WEIGHTS REGRE.M

flag
ibe
flag2

- compute weight images for each permutation if 1
- which beta to use for MKL and multiple modalities
- build image of weights per region

Output:
- name of the .img file created
img name
+ image file created on disk

19.13

prt compute weights regre.m

FORMAT prt compute weights regre(PRT,in,model idx)
This function calls prt weights to compute weights
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
in
- structure with specific information to create
weights
.model name - model name (string)
.img name
- (optional) name of the file to be created
(string)
.pathdir
- directory path where to save weights (same as the
one for PRT.mat) (string)
model idx
- model index (integer)
flag
- compute weight images for each permutation if 1
ibe
- which beta to use for MKL and multiple modalities
flag2
- build image of weights per region
Output:
- name of the .img file created
img name
+ image file created on disk

19.14

prt cv fold.m

Function to run a single cross-validation fold
Inputs:
------PRT:
in.mid:
in.ID:
in.CV:
in.Phi all:
in.t

data structure
index to the model we are working on
ID matrix
Cross-validation matrix (current fold only)
Cell array of data matri(ces) (training and test)
prediction targets

Outputs:
-------model:
the model returned by the machine
targets.train: training targets
targets.test: test targets
Notes:
-----The training and test targets output byt this function are not
necessarily equivalent to the targets that are supplied to the function.
e.g. some data operations can modify the number of samples (e.g. sample

171

172

CHAPTER 19. LIST OF PRONTO FUNCTIONS

averaging). In such cases size(targets.train) ~= size(in.t)

19.15

prt cv model.m

Function to run a cross-validation structure on a given model
Inputs:
------PRT:
in.fname:
in.model name:

data structure
filename for PRT.mat (string)
name for this model (string)

Outputs:
-------Writes the following fields in the PRT data structure:
PRT.model(m).output.fold(i).targets:
targets for fold(i)
PRT.model(m).output.fold(i).predictions: predictions for fold(i)
PRT.model(m).output.fold(i).stats:
statistics for fold(i)
PRT.model(m).output.fold(i).custom:
optional fields
Notes:
-----The PRT.model(m).input fields are set by prt init model, not by
this function

19.16

prt cv opt param.m

Function to pass optional (advanced) parameters into the classifier.
This is primarily used for prediction methods that need to know something
about the experimental design that is normally not accessible to ordinary
(i.e. generic) prediction functions (e.g. task onsets or TR). Examples of
this kind of method include multi-class classifier using kernel
regression (MCKR) and the machine that implements nested cross-validation
Inputs:
------PRT:
data structure
ID:
id matrix for the current cross-validation fold
model id: which model are we working on?
Outputs:
-------param.id fold:
param.model id:
param.PRT:

the id matrix for this fold
id for the model being computed
PRT data structure

Notes:
-------The outputs (param.xxx) are provided for use by the classifier

19.17. PRT DATA CONDITIONS.M

19.17

173

prt data conditions.m

PRT DATA CONDITIONS M-file for prt data conditions.fig
PRT DATA CONDITIONS, by itself, creates a new PRT DATA CONDITIONS or
raises the existing singleton*.
H = PRT DATA CONDITIONS returns the handle to a new PRT DATA CONDITIONS
or the handle to the existing singleton*.
PRT DATA CONDITIONS(’CALLBACK’,hObject,eventData,handles,...) calls the
local function named CALLBACK in PRT DATA CONDITIONS.M with the given
input arguments.
PRT DATA CONDITIONS(’Property’,’Value’,...) creates a new
PRT DATA CONDITIONS or raises the existing singleton*. Starting from the
left, property value pairs are applied to the GUI before
prt data conditions OpeningFcn gets called. An unrecognized property name
or invalid value makes property application stop. All inputs are passed
to prt data conditions OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.18

prt data modality.m

PRT DATA MODALITY M-file for prt data modality.fig
PRT DATA MODALITY, by itself, creates a new PRT DATA MODALITY or raises
the existing singleton*.
H = PRT DATA MODALITY returns the handle to a new PRT DATA MODALITY or
the handle to the existing singleton*.
PRT DATA MODALITY(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT DATA MODALITY.M with the given input arguments.
PRT DATA MODALITY(’Property’,’Value’,...) creates a new PRT DATA MODALITY
or raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt data modality OpeningFcn gets called.
An unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt data modality OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.19

prt data review.m

PRT DATA REVIEW M-file for prt data review.fig

174

CHAPTER 19. LIST OF PRONTO FUNCTIONS

PRT DATA REVIEW, by itself, creates a new PRT DATA REVIEW or raises the
existing singleton*.
H = PRT DATA REVIEW returns the handle to a new PRT DATA REVIEW or the
handle to the existing singleton*.
PRT DATA REVIEW(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT DATA REVIEW.M with the given input arguments.
PRT DATA REVIEW(’Property’,’Value’,...) creates a new PRT DATA REVIEW or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt data review OpeningFcn gets
called. An unrecognized property name or invalid value makes property
application stop. All inputs are passed to prt data review OpeningFcn
via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.20

prt defaults.m

Sets the defaults which are used by the Pattern Recognition for
Neuroimaging Toolbox, aka. PRoNTo.
FORMAT prt defaults

This file can be customised to any the site/person own setup.
Individual users can make copies which can be stored on their own
matlab path. Make sure your ’prt defaults’ is the first one found in the
path. See matlab documentation for details on setting path.
Care must be taken when modifying this file!
The structure and content of this file are largely inspired by SPM:
http://www.fil.ion.ucl.ac.uk/spm

19.21

prt fs.m

Function to build file arrays containing the (linearly detrended) data
and compute a linear (dot product) kernel from them
Inputs:
------in.fname:
in.fs name:

filename for the PRT.mat (string)
name of fs and relative path filename for the kernel matrix

in.mod(m).mod name:
in.mod(m).detrend:
in.mod(m).param dt:
in.mod(m).mode:

name of modality to include in this kernel (string)
detrend (scalar: 0 = none, 1 = linear)
parameters for the kernel detrend (e.g. DCT bases)
’all cond’ or ’all scans’ (string)

19.22. PRT FS MODALITY.M

in.mod(m).mask:
in.mod(m).normalise:
in.mod(m).matnorm:
in.mod(m).multroi
in.mod(m).atlasroi

mask file used to create the kernel
0 = none, 1 = normalise kernel, 2 = scale modality
filename for scaling matrix
1 if one kernel per region required
name of the atlas to build one kernel per region

Perform multi-kernel learning (1) or not (0)? If yes, the
in.flag mm:
kernel is saved as a cell vector, with one kernel per modality
Outputs:
-------Calls prt init fs to populate basic fields in PRT.fs(f)...
Writes PRT.mat
Writes the kernel matrix to the path indicated by in.fs name

19.22

prt fs modality.m

Function to build file arrays containing the (linearly detrended) data
and compute a linear (dot product) kernel from them
Inputs:
------in.fname:
in.fs name:

filename for the PRT.mat (string)
name of fs and relative path filename for the kernel matrix

in.mod(m).mod name:
in.mod(m).detrend:
in.mod(m).param dt:
in.mod(m).mode:
in.mod(m).mask:
in.mod(m).normalise:
in.mod(m).matnorm:

name of modality to include in this kernel (string)
detrend (scalar: 0 = none, 1 = linear)
parameters for the kernel detrend (e.g. DCT bases)
’all cond’ or ’all scans’ (string)
mask file used to create the kernel
0 = none, 1 = normalise kernel, 2 = scale modality
filename for scaling matrix

in.fid:
index of feature set to create
in.tocomp:
vector of booleans indicating whether to build the feature set
in.precmask: cell array containing the names of the second-level mask for
each modality to build
flag:
addin:

set to 1 to compute one kernel per region as labelled in atlas
additional inputs for this operation to optimize computation

Outputs:
-------Writes the kernel matrix to the path indicated by in.fs name and the
feature set in a file array if it needs to be computed

19.23

prt func2html.m

Script to generate the list of .m functions into html files
which can be browsed around with your favourite browser.
Note that this script relies on the M2HTML package which is *NOT*
distributed with PRoNTo!

175

176

CHAPTER 19. LIST OF PRONTO FUNCTIONS

For more information, please read the M2HTML tutorial and FAQ at:
$<$http://www.artefact.tk/software/matlab/m2html/$>$

19.24

prt getKernelModel.m

Function to load the kernels according to the samples considered in a
given model
Inputs:
------PRT:
prt dir:
mid :

data structure
path for PRT.mat (string)
index of model in the data structure/ PRT.mat

Output:
-------Phi all : cell array with one kernel per cell (in case of
multiple kernels) or a single cell with the samples considered in the
specified model, as defined by the class selection.
ID :
the ID matrix for the considered samples
fid :
index of feature set in data structure / PRT.mat

19.25

prt get defaults.m

Get/set the defaults values associated with an identifier
FORMAT defaults = prt get defaults
Return the global "defaults" variable defined in prt defaults.m.
FORMAT defval = prt get defaults(defstr)
Return the defaults value associated with identifier "defstr".
Currently, this is a ’.’ subscript reference into the global
"prt def" variable defined in prt defaults.m.
FORMAT prt get defaults(defstr, defval)
Sets the defaults value associated with identifier "defstr". The new
defaults value applies immediately to:
* new modules in batch jobs
* modules in batch jobs that have not been saved yet
This value will not be saved for future sessions of PRoNTo. To make
persistent changes, edit prt defaults.m.
The structure and content of this file are largely inspired by SPM &
Matlabbatch.
http://www.fil.ion.ucl.ac.uk/spm
http://sourceforge.net/projects/matlabbatch/

19.26

prt get filename.m

out = prt get filename(ids)

19.27. PRT INIT FS.M

19.27

prt init fs.m

function to initialise the kernel data structure
-----------------------------------------------FORMAT: Two modes are possible:
fid = prt init fs(PRT, in)
[fid, PRT, tocomp] = prt init fs(PRT, in)
USAGE 1:
------------------------------------------------------------------------function will return the id of a feature set or an error if it doesn’t
exist in PRT.mat
Input:
-----in.fs name: name for the feature set (string)
Output:
------fid : is the identifier for the feature set in PRT.mat
USAGE 2:
------------------------------------------------------------------------function will create the feature set in PRT.mat and overwrite it if it
already exists.
Input:
-----in.fs name: name for the feature set (string)
in.fname:
name of PRT.mat
in.mod(m).mod name: name of the modality
in.mod(m).detrend:
type of detrending
in.mod(m).mode:
’all scans’ or ’all cond’
in.mod(m).mask:
mask used to create the feature set
in.mod(m).param dt: parameters used for detrending (if any)
in.mod(m).normalise: scale the input scans or not
in.mod(m).matnorm:
mat file used to scale the input scans
Output:
------fid : is the identifier for the model constructed in PRT.mat
Populates the following fields in PRT.mat (copied from above):
PRT.fs(f).fs name
PRT.fs(f).fas
PRT.fs(f).k file
Also computes the following fields:
Identifier matrix (useful later)
PRT.fs(f).id mat:
PRT.fs(f).id col names: Columns in the id matrix
Note: this function does not write PRT.mat. That should be done by the
calling function

177

178

19.28

CHAPTER 19. LIST OF PRONTO FUNCTIONS

prt init model.m

function to initialise the model data structure
FORMAT: Two modes are possible:
mid = prt init model(PRT, in)
[mid, PRT] = prt init model(PRT, in)
USAGE 1:
-----------------------------------------------------------------------function will return the id of a model or an error if it doesn’t
exist in PRT.mat
Input:
-----in.model name: name of the model (string)
Output:
------mid : is the identifier for the model in PRT.mat
USAGE 2:
------------------------------------------------------------------------function will create the model in PRT.mat and overwrite it if it
already exists.
Input:
-----in.model name:
in.use kernel:
in.machine:
in.type:

name of the model to be created (string)
use kernel or basis functions for this model (boolean)
prediction machine to use for this model (struct)
’classification’ or ’regression’

Output:
------Populates the following fields in PRT.mat (copied from above):
PRT.model(m).input.model name
PRT.model(m).input.type
PRT.model(m).input.use kernel
PRT.model(m).input.machine
Note: this function does not write PRT.mat. That should be done by the
calling function

19.29

prt latex.m

Extract information from the toolbox m-files and output them as usable
.tex files which can be directly included in the manual.
There are 2 types of m2tex operations:
1. converting the job configuration tree, i.e. * cfg * files defining the
batching interface into a series of .tex files.
NOTE: Only generate .tex files for each exec branch of prt batch.
2. converting the help header of the functions into .tex files.

19.30. PRT LOAD.M

179

These files are then included in a manually written prt manual.tex file,
which also includes chapter/sections written manually.
FORMAT prt latex(opt)
INPUT
opt: option structure
.tex cfg : turn the config files help into a tex file (1), or not (0)
.tex fct : turn the functions help into a tex file (1), or not (0)
NOTE:
File derived from that of the SPM8 distribution.
http://www.fil.ion.ucl.ac.uk/spm

19.30

prt load.m

Function to load the PRT.mat and check its integrity regarding the
kernels and feature sets that it is supposed to contain. Updates the
feature name if needed.
input

set

: name of the PRT.mat, path included

output : PRT structure updated

19.31

prt load blocks.m

Load one or more blocks of data.
This script is a effectively a wrapper function that for the routines
that actually do the work (SPM nifti routines)
The syntax is either:
img = prt load blocks(filenames, block size, block range) just to specify
continuous blocks of data
or
img = prt load blocks(filenames, voxel index) to access non continuous
blocks

19.32

prt model.m

Function to configure and build the PRT.model data structure
Input:
-----PRT fields:
model.fs(f).fs name:
feature set(s) this CV approach is defined for
model.fs(f).fs features: feature selection mode (’all’ or ’mask’)
model.fs(f).mask file:
mask for this feature set (fs features=’mask’)
in.fname:

filename for PRT.mat

180

CHAPTER 19. LIST OF PRONTO FUNCTIONS

in.model name:
in.type:
in.use kernel:
in.operations:

name for this cross-validation structure
’classification’ or ’regression’
does this model use kernels or features?
operations to apply before prediction

in.fs(f).fs name:

feature set(s) this CV approach is defined for

in.class(c).class name
in.class(c).group(g).subj(s).num
in.class(c).group(g).subj(s).modality(m).mod name
EITHER: in.class(c).group(g).subj(s).modality(m).conds(c).cond name
OR:
in.class(c).group(g).subj(s).modality(m).all scans
OR:
in.class(c).group(g).subj(s).modality(m).all cond
in.cv.type:
type of cross-validation (’loso’,’losgo’,’custom’)
in.cv.mat file: file specifying CV matrix (if type=’custom’);
Output:
------This function performs the following functions:
1. populates basic fields in PRT.model(m).input
2. computes PRT.model(m).input.targets based on in.class(c)...
3. computes PRT.model(m).input.samp idx based on targets
4. computes PRT.model(m).input.cv mat based on the labels and CV spec

19.33

prt nested cv.m

Function to perform the nested CV
Inputs:
------in.nc:
in.ID:
in.mid:
in.CV:
in.Phi all:

number of classes
ID matrix
model id
cross-validation matrix
Kernel

Outputs:
-------out.opt param:

optimal hyper-parameter choosen using the stats from
the inner CVs
out.vary param: stats values associated with all the hyper-parameters

19.34

prt permutation.m

Function to compute permutation test
Inputs:
------PRT:
n perm:
modelid:
path:

PRT structured including model
number of permutations
model ID
path

19.35. PRT PLOT ROC.M

flag:

181

boolean variable. set to 1 to save the weights for each
permutation. default: 0

Outputs:
-------for classification
permutation.c acc:
permutation.b acc:
permutation.pvalue b acc:
permutation.pvalue c acc:

Permuted accuracy per class
Permuted balanced accuracy
p-value for c acc
p-value for b acc

for regression
permutation.corr: Permuted correlation
permutation.mse: Permuted mean square error
permutation.pval corr: p-value for corr
permutation.pval r2: p-value for r2;
permutation.pval mse: p-value for mse
permutation.pval nmse: p-value for nmse

19.35

prt plot ROC.m

FORMAT prt plot ROC(PRT, model, fold, axes handle)
This function plots the ROC plot that appears on prt ui results
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
model
- the number of the model that will be ploted
fold
- the number of the fold
axes handle
- (Optional) axes where the plot will be displayed
Output:
None

19.36

prt plot confusion matrix.m

FORMAT prt plot confusion matrix(PRT, model, fold, axes handle)
This function plots the confusion matrix that appears on prt ui results
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
model
- the number of the model that will be ploted
fold
- the number of the fold
axes handle
- (Optional) axes where the plot will be displayed
Output:
None

19.37

prt plot histograms.m

182

CHAPTER 19. LIST OF PRONTO FUNCTIONS

FORMAT prt plot histograms(PRT, model, fold, axes handle)
This function plots the histogram that appears on prt ui results.
The maximum number of classes that can be ploted is 7. However, this can
be increased by editing the function. Just add more colours to the
colourList variable.
Inputs:
PRT
model
fold
axes handle

- data/design/model structure (it needs to contain
at least one estimated model).
- the number of the model that will be ploted
- the number of the fold
- (Optional) axes where the plot will be displayed

Output:
None

19.38

prt plot nested cv.m

FORMAT prt plot nested cv(PRT, model, fold, axes handle)
Plots the results of the nested cv that appear on prt ui results.

Inputs:
PRT
model
fold
axes handle

- data/design/model structure (it needs to contain
at least one estimated model).
- the number of the model that will be ploted
- the number of the fold
- (Optional) axes where the plot will be displayed

Output:
None

19.39

prt plot prediction.m

FORMAT prt plot prediction(PRT, model, fold, marker size, axes handle)
This function plots the prediction plot that appears on prt ui results
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
model
- the number of the model that will be ploted
fold
- the number of the fold
marker size
- (Optional) the size of the markers in the plot,
the default is 7
axes handle
- (Optional) axes where the plot will be displayed
Output:
None

19.40. PRT PLOT PREDICTION REG BAR.M

19.40

prt plot prediction reg bar.m

FORMAT prt plot prediction reg bar(PRT, model, axes handle)
This function plots the bar plot that appears on prt ui results
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
model
- the number of the model that will be ploted
- (Optional) axes where the plot will be displayed
axes handle
Output:
None

19.41

prt plot prediction reg line.m

FORMAT prt plot prediction reg line(PRT, model, axes handle)
This function plots the line plot that appears on prt ui results
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
model
- the number of the model that will be ploted
axes handle
- (Optional) axes where the plot will be displayed
Output:
None

19.42

prt plot prediction reg scatter.m

FORMAT prt plot prediction reg scatter(PRT, model, axes handle)
This function plots the scatter plot that appears on prt ui results
Inputs:
PRT
- data/design/model structure (it needs to contain
at least one estimated model).
model
- the number of the model that will be ploted
axes handle
- (Optional) axes where the plot will be displayed
Output:
None

19.43

prt region histogram.m

19.44

prt remove confounds.m

[Kr, R] = prt remove confounds(K,C)
Function to remove confounds from kernel.

183

184

19.45

CHAPTER 19. LIST OF PRONTO FUNCTIONS

prt stats.m

Function to compute predictions machine performance statistcs statistics
Inputs:
---------------model.predictions: predictions derived from the predictive model
model.type:
what type of prediction machine (e.g. ’classifier’,’regression’)
tte: true targets (test set)
nk: number of classes if classification (empty otherwise)
flag: ’fold’ for statistics in each fold
’model’ for statistics in each model
Outputs:
------------------Classification:
stats.con mat: Confusion matrix (nClasses x nClasses matrix, pred x true)
stats.acc:
Accuracy (scalar)
stats.b acc:
Balanced accuracy (nClasses x 1 vector)
stats.c acc:
Accuracy by class (nClasses x 1 vector)
stats.c pv:
Predictive value for each class (nClasses x 1 vector)
Regression:
stats.mse:
stats.corr:
stats.r2:

19.46

Mean square error between test and prediction
Correlation between test and prediction
Squared correlation

prt struct.m

fields that it is supposed to contain. Updates the PRT if needed.
input

: PRT structure to check

output : PRT structure updated

19.47

prt struct2latex.m

Function that takes in a structure S and writes down the latex code
describing the whole structure and substructures recursively.
The routine specifically generates the ’adv PRTstruct.tex’ file that is
included, in the prt manual.
Bits of the code are copied/inspired by spm latex.m from the SPM8
distribution: http://www.fil.ion.ucl.ac.uk/spm

19.48

prt text input.m

PRT TEXT INPUT M-file for prt text input.fig
PRT TEXT INPUT, by itself, creates a new PRT TEXT INPUT or raises the

19.49. PRT UI COMPUTE WEIGHTS.M

185

existing singleton*.
H = PRT TEXT INPUT returns the handle to a new PRT TEXT INPUT or the
handle to the existing singleton*.
PRT TEXT INPUT(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT TEXT INPUT.M with the given input arguments.
PRT TEXT INPUT(’Property’,’Value’,...) creates a new PRT TEXT INPUT or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt text input OpeningFcn gets called.
An unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt text input OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.49

prt ui compute weights.m

PRT UI COMPUTE WEIGHTS M-file for prt ui compute weights.fig
PRT UI COMPUTE WEIGHTS, by itself, creates a new PRT UI COMPUTE WEIGHTS
or raises the existing singleton*.
H = PRT UI COMPUTE WEIGHTS returns the handle to a new PRT UI COMPUTE WEIGHTS
or the handle to the existing singleton*.
PRT UI COMPUTE WEIGHTS(’CALLBACK’,hObject,eventData,handles,...) calls
the local function named CALLBACK in PRT UI COMPUTE WEIGHTS.M with the
given input arguments.
PRT UI COMPUTE WEIGHTS(’Property’,’Value’,...) creates a new PRT UI COMPUTE WEIGHTS
or raises the existing singleton*. Starting from the left, property
value pairs are applied to the GUI before prt ui compute weights OpeningFcn
gets called. An unrecognized property name or invalid value makes
property application stop. All inputs are passed to prt ui compute weights OpeningFcn
via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.50

prt ui custom CV.m

PRT UI CUSTOM CV M-file for prt ui custom CV.fig
PRT UI CUSTOM CV, by itself, creates a new PRT UI CUSTOM CV or
raises the existing singleton*.
H = PRT UI CUSTOM CV returns the handle to a new PRT UI CUSTOM CV
or the handle to the existing singleton*.

186

CHAPTER 19. LIST OF PRONTO FUNCTIONS

PRT UI CUSTOM CV(’CALLBACK’,hObject,eventData,handles,...) calls the
local function named CALLBACK in PRT UI CUSTOM CV.M with the given
input arguments.
PRT UI CUSTOM CV(’Property’,’Value’,...) creates a new PRT UI CUSTOM CV
or raises the existing singleton*. Starting from the left, property
value pairs are applied to the GUI before prt ui custom CV OpeningFcn
gets called. An unrecognized property name or invalid value makes
property application stop. All inputs are passed to prt ui custom CV OpeningFcn
via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.51

prt ui cv model.m

PRT UI CV MODEL M-file for prt ui cv model.fig
PRT UI CV MODEL, by itself, creates a new PRT UI CV MODEL or raises the
existing singleton*.
H = PRT UI CV MODEL returns the handle to a new PRT UI CV MODEL or the
handle to the existing singleton*.
PRT UI CV MODEL(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI CV MODEL.M with the given input
arguments.
PRT UI CV MODEL(’Property’,’Value’,...) creates a new PRT UI CV MODEL or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt ui cv model OpeningFcn gets
called. An unrecognized property name or invalid value makes property
application stop. All inputs are passed to prt ui cv model OpeningFcn
via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.52

prt ui design.m

PRT UI DESIGN M-file for prt ui design.fig
PRT UI DESIGN, by itself, creates a new PRT UI DESIGN or raises the
existing singleton*.
H = PRT UI DESIGN returns the handle to a new PRT UI DESIGN or the handle
to the existing singleton*.
PRT UI DESIGN(’CALLBACK’,hObject,eventData,handles,...) calls the local

19.53. PRT UI DISP WEIGHTS.M

187

function named CALLBACK in PRT UI DESIGN.M with the given input arguments.
PRT UI DESIGN(’Property’,’Value’,...) creates a new PRT UI DESIGN or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt ui design OpeningFcn gets called.
An unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt ui design OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.53

prt ui disp weights.m

PRT UI DISP WEIGHTS MATLAB code for prt ui disp weights.fig
PRT UI DISP WEIGHTS, by itself, creates a new PRT UI DISP WEIGHTS or raises the
existing singleton*.
H = PRT UI DISP WEIGHTS returns the handle to a new PRT UI DISP WEIGHTS or the
handle to the existing singleton*.
PRT UI DISP WEIGHTS(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI DISP WEIGHTS.M with the given input arguments.
PRT UI DISP WEIGHTS(’Property’,’Value’,...) creates a new PRT UI DISP WEIGHTS or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt ui disp weights OpeningFcn gets called.
An unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt ui disp weights OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.54

prt ui kernel construction.m

PRT UI KERNEL MATLAB code for prt ui kernel.fig
PRT UI KERNEL, by itself, creates a new PRT UI KERNEL or raises the
existing singleton*.
H = PRT UI KERNEL returns the handle to a new PRT UI KERNEL or the handle
to the existing singleton*.
PRT UI KERNEL(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI KERNEL.M with the given input arguments.
PRT UI KERNEL(’Property’,’Value’,...) creates a new PRT UI KERNEL or raises
the existing singleton*. Starting from the left, property value pairs are
applied to the GUI before prt ui kernel OpeningFcn gets called. An
unrecognized property name or invalid value makes property application

188

stop.

CHAPTER 19. LIST OF PRONTO FUNCTIONS

All inputs are passed to prt ui kernel OpeningFcn via varargin.

*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.55

prt ui main.m

PRT UI MAIN M-file for prt ui main.fig
PRT UI MAIN, by itself, creates a new PRT UI MAIN or raises the existing
singleton*.
H = PRT UI MAIN returns the handle to a new PRT UI MAIN or the handle to
the existing singleton*.
PRT UI MAIN(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI MAIN.M with the given input arguments.
PRT UI MAIN(’Property’,’Value’,...) creates a new PRT UI MAIN or raises
the existing singleton*. Starting from the left, property value pairs are
applied to the GUI before prt ui main OpeningFcn gets called. An
unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt ui main OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.56

prt ui model.m

PRT UI KERNEL CONSTRUCTION M-file for prt ui kernel construction.fig
PRT UI KERNEL CONSTRUCTION, by itself, creates a new
PRT UI KERNEL CONSTRUCTION or raises the existing singleton*.
H = PRT UI KERNEL CONSTRUCTION returns the handle to a new
PRT UI KERNEL CONSTRUCTION or the handle to the existing singleton*.
PRT UI KERNEL CONSTRUCTION(’CALLBACK’,hObject,eventData,handles,...)
calls the local function named CALLBACK in PRT UI KERNEL CONSTRUCTION.M
with the given input arguments.
PRT UI KERNEL CONSTRUCTION(’Property’,’Value’,...) creates a new
PRT UI KERNEL CONSTRUCTION or raises the existing singleton*. Starting
from the left, property value pairs are applied to the GUI before
prt ui kernel construction OpeningFcn gets called. An unrecognized
property name or invalid value makes property application stop. All
inputs are passed to prt ui kernel construction OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

19.57. PRT UI PREPARE DATA.M

189

See also: GUIDE, GUIDATA, GUIHANDLES

19.57

prt ui prepare data.m

PRT UI KERNEL MATLAB code for prt ui kernel.fig
PRT UI KERNEL, by itself, creates a new PRT UI KERNEL or raises the
existing singleton*.
H = PRT UI KERNEL returns the handle to a new PRT UI KERNEL or the handle
to the existing singleton*.
PRT UI KERNEL(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI KERNEL.M with the given input arguments.
PRT UI KERNEL(’Property’,’Value’,...) creates a new PRT UI KERNEL or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt ui kernel OpeningFcn gets called.
An unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt ui kernel OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.58

prt ui prepare datamod.m

PRT UI KERNEL MODALITY M-file for prt ui kernel modality.fig
PRT UI KERNEL MODALITY, by itself, creates a new PRT UI KERNEL MODALITY
or raises the existing singleton*.
H = PRT UI KERNEL MODALITY returns the handle to a new
PRT UI KERNEL MODALITY or the handle to the existing singleton*.
PRT UI KERNEL MODALITY(’CALLBACK’,hObject,eventData,handles,...) calls
the local function named CALLBACK in PRT UI KERNEL MODALITY.M with the
given input arguments.
PRT UI KERNEL MODALITY(’Property’,’Value’,...) creates a new
PRT UI KERNEL MODALITY or raises the existing singleton*. Starting from
the left, property value pairs are applied to the GUI before
prt ui kernel modality OpeningFcn gets called. An unrecognized property
name or invalid value makes property application stop. All inputs are
passed to prt ui kernel modality OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".
See also: GUIDE, GUIDATA, GUIHANDLES

Choose "GUI allows only one

190

19.59

CHAPTER 19. LIST OF PRONTO FUNCTIONS

prt ui results.m

PRT UI RESULTS MATLAB code for prt ui results.fig
PRT UI RESULTS, by itself, creates a new PRT UI RESULTS or raises the
existing singleton*.
H = PRT UI RESULTS returns the handle to a new PRT UI RESULTS or the
handle to the existing singleton*.
PRT UI RESULTS(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI RESULTS.M with the given input arguments.
PRT UI RESULTS(’Property’,’Value’,...) creates a new PRT UI RESULTS or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt ui results OpeningFcn gets called.
An unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt ui results OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.60

prt ui results ROI.m

PRT UI RESULTS ROI M-file for prt ui results ROI.fig
PRT UI RESULTS ROI, by itself, creates a new PRT UI RESULTS ROI or
raises the existing singleton*.
H = PRT UI RESULTS ROI returns the handle to a new PRT UI RESULTS ROI
or the handle to the existing singleton*.
PRT UI RESULTS ROI(’CALLBACK’,hObject,eventData,handles,...) calls the
local function named CALLBACK in PRT UI RESULTS ROI.M with the given
input arguments.
PRT UI RESULTS ROI(’Property’,’Value’,...) creates a new
PRT UI RESULTS ROI or raises the existing singleton*. Starting from the
left, property value pairs are applied to the GUI before
prt ui results ROI OpeningFcn gets called. An unrecognized property name
or invalid value makes property application stop. All inputs are passed
to prt ui results ROI OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.61

prt ui results help.m

PRT UI RESULTS HELP MATLAB code for prt ui results help.fig

19.62. PRT UI RESULTS STATS.M

191

PRT UI RESULTS HELP, by itself, creates a new PRT UI RESULTS HELP or
raises the existing singleton*.
H = PRT UI RESULTS HELP returns the handle to a new PRT UI RESULTS HELP
or the handle to the existing singleton*.
PRT UI RESULTS HELP(’CALLBACK’,hObject,eventData,handles,...) calls the
local function named CALLBACK in PRT UI RESULTS HELP.M with the given
input arguments.
PRT UI RESULTS HELP(’Property’,’Value’,...) creates a new
PRT UI RESULTS HELP or raises the existing singleton*. Starting from the
left, property value pairs are applied to the GUI before
prt ui results help OpeningFcn gets called. An unrecognized property
name or invalid value makes property application stop. All inputs are
passed to prt ui results help OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.62

prt ui results stats.m

PRT UI RESULTS STATS MATLAB code for prt ui results stats.fig
PRT UI RESULTS STATS, by itself, creates a new PRT UI RESULTS STATS or raises the
existing singleton*.
H = PRT UI RESULTS STATS returns the handle to a new PRT UI RESULTS STATS or the
handle to the existing singleton*.
PRT UI RESULTS STATS(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI RESULTS STATS.M with the given input arguments.
PRT UI RESULTS STATS(’Property’,’Value’,...) creates a new PRT UI RESULTS STATS or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt ui results stats OpeningFcn gets called.
An unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt ui results stats OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.63

prt ui reviewCV.m

PRT UI REVIEWCV M-file for prt ui reviewCV.fig
PRT UI REVIEWCV, by itself, creates a new PRT UI REVIEWCV or raises the
existing singleton*.

192

CHAPTER 19. LIST OF PRONTO FUNCTIONS

H = PRT UI REVIEWCV returns the handle to a new PRT UI REVIEWCV or the
handle to the existing singleton*.
PRT UI REVIEWCV(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI REVIEWCV.M with the given input
arguments.
PRT UI REVIEWCV(’Property’,’Value’,...) creates a new PRT UI REVIEWCV or
raises the existing singleton*. Starting from the left, property value
pairs are applied to the GUI before prt ui reviewCV OpeningFcn gets
called. An unrecognized property name or invalid value makes property
application stop. All inputs are passed to prt ui reviewCV OpeningFcn
via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.64

prt ui reviewmodel.m

PRT UI REVIEWMODEL M-file for prt ui reviewmodel.fig
PRT UI REVIEWMODEL, by itself, creates a new PRT UI REVIEWMODEL or raises
the existing singleton*.
H = PRT UI REVIEWMODEL returns the handle to a new PRT UI REVIEWMODEL or
the handle to the existing singleton*.
PRT UI REVIEWMODEL(’CALLBACK’,hObject,eventData,handles,...) calls the
local function named CALLBACK in PRT UI REVIEWMODEL.M with the given
input arguments.
PRT UI REVIEWMODEL(’Property’,’Value’,...) creates a new PRT UI REVIEWMODEL
or raises the existing singleton*. Starting from the left, property
value pairs are applied to the GUI before prt ui reviewmodel OpeningFcn
gets called. An unrecognized property name or invalid value makes
property application stop. All inputs are passed to prt ui reviewmodel OpeningFcn
via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.65

prt ui select class.m

PRT UI SELECT CLASS M-file for prt ui select class.fig
PRT UI SELECT CLASS, by itself, creates a new PRT UI SELECT CLASS or
raises the existing singleton*.
H = PRT UI SELECT CLASS returns the handle to a new PRT UI SELECT CLASS
or the handle to the existing singleton*.

19.66. PRT UI SELECT REG.M

193

PRT UI SELECT CLASS(’CALLBACK’,hObject,eventData,handles,...) calls the
local function named CALLBACK in PRT UI SELECT CLASS.M with the given
input arguments.
PRT UI SELECT CLASS(’Property’,’Value’,...) creates a new PRT UI SELECT CLASS
or raises the existing singleton*. Starting from the left, property
value pairs are applied to the GUI before prt ui select class OpeningFcn
gets called. An unrecognized property name or invalid value makes
property application stop. All inputs are passed to prt ui select class OpeningFcn
via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.66

prt ui select reg.m

PRT UI SELECT REG M-file for prt ui select reg.fig
PRT UI SELECT REG, by itself, creates a new PRT UI SELECT REG or raises
the existing singleton*.
H = PRT UI SELECT REG returns the handle to a new PRT UI SELECT REG or
the handle to the existing singleton*.
PRT UI SELECT REG(’CALLBACK’,hObject,eventData,handles,...) calls the
local function named CALLBACK in PRT UI SELECT REG.M with the given input
arguments.
PRT UI SELECT REG(’Property’,’Value’,...) creates a new PRT UI SELECT REG
or raises the existing singleton*. Starting from the left, property
value pairs are applied to the GUI before prt ui select reg OpeningFcn
gets called. An unrecognized property name or invalid value makes
property application stop. All inputs are passed to prt ui select reg OpeningFcn
via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.67

prt ui specify CV basis.m

19.68

prt ui stats.m

PRT UI STATS MATLAB code for prt ui stats.fig
PRT UI STATS, by itself, creates a new PRT UI STATS or raises the
existing singleton*.

194

CHAPTER 19. LIST OF PRONTO FUNCTIONS

H = PRT UI STATS returns the handle to a new PRT UI STATS or the handle
to the existing singleton*.
PRT UI STATS(’CALLBACK’,hObject,eventData,handles,...) calls the local
function named CALLBACK in PRT UI STATS.M with the given input arguments.
PRT UI STATS(’Property’,’Value’,...) creates a new PRT UI STATS or raises
the existing singleton*. Starting from the left, property value pairs
are applied to the GUI before prt ui stats OpeningFcn gets called. An
unrecognized property name or invalid value makes property application
stop. All inputs are passed to prt ui stats OpeningFcn via varargin.
*See GUI Options on GUIDE’s Tools menu.
instance to run (singleton)".

Choose "GUI allows only one

See also: GUIDE, GUIDATA, GUIHANDLES

19.69

prt ui sure.m

19.70

machines

19.70.1

machines\prt KRR.m

w = prt KRR(K,t,reg)

19.70.2

machines\prt machine.m

Run machine function for classification or regression
FORMAT output = prt machine(d,m)
Inputs:
d
- structure with information about the data, with fields:
Mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) or in
form of features (false)
Optional fields: the machine is respnsible for dealing with this
optional fields (e.g. d.testcov)
m
- structure with information about the classification or
regression machine to use, with fields:
.function - function for classification or regression (string)
.args
- function arguments (either a string, a matrix, or a
struct). This is specific to each machine, e.g. for
an L2-norm linear SVM this could be the C parameter
Output:

19.70. MACHINES

output
- output of machine (struct).
Mandatory fields:
.predictions - predictions of classification or regression
[Nte x D]
Optional fields: the machine is responsible for returning
parameters of interest. For exemple for an SVM this could be the
number of support vector used in the hyperplane weights computation

19.70.3

machines\prt machine RT bin.m

Run binary Ensemble of Regression Tree - wrapper for Pierre Geurt’s
RT code
FORMAT output = prt machine RT bin(d,args)
Inputs:
d
- structure with data information, with mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) of in
form of features (false)
args
- vector of RT arguments
args(1) - number of trees (default: 501)
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.func val - value of the decision function
.type
- which type of machine this is (here, ’classifier’)

19.70.4

machines\prt machine gpclap.m

Run multiclass Gaussian process classification (Laplace approximation)
FORMAT output = prt machine gpclap(d,args)
Inputs:
d
- structure with data information, with mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.testcov
- testing covariance (cell array of matrices row vectors,
each [Nte x Nte])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) or in
form of features (false)
args
- argument string, where
-h
- optimise hyperparameters (otherwise don’t)

195

196

CHAPTER 19. LIST OF PRONTO FUNCTIONS

-c covfun

- covariance function:
’covLINkcell’ - simple dot product
’covLINglm’
- construct a GLM
experimental args (use at your own risk):
-p
- use priors for the hyperparameters. If specified, this
indicates that a maximum a posteriori (MAP) approach
will be used to set covariance function
hyperparameters. The priors are obtained
by calling prt gp priors(’covFuncName’)
N.B.: for the arguments specifying functions, pass in a string, not
a function handle. This script will generate a function handle
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.type
- which type of machine this is (here, ’classifier’)
.func val - predictive probabilties
.loghyper - log hyperparameters
.nlml
- negative log marginal likelihood
.mu
- test latent means
.s2
- test latent variances
.alpha
- GP weighting coefficients

19.70.5

machines\prt machine gpml.m

Run Gaussian process model - wrapper for gpml toolbox
FORMAT output = prt machine gpml(d,args)
Inputs:
d
- structure with data information, with mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.testcov
- testing covariance (cell array of matrices row vectors,
each [Nte x Nte])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) or in
form of features (false)
args
- argument string, where
-h
- optimise hyperparameters (otherwise don’t)
-f iter
- max # iterations for optimiser (ignored if -h not set)
-l likfun - likelihood function:
’likErf’ - erf/probit likelihood (binary only)
-c covfun - covariance function:
’covLINkcell’ - simple dot product
’covLINglm’
- construct a GLM
-m meanfun - mean function:
’meanConstcell’ - suitable for dot product
’meanConstglm’ - suitable for GLM
-i inffun - inference function:

19.70. MACHINES

’prt infEP’ - Expectation Propagation
experimental args (use at your own risk):
-p
- use priors for the hyperparameters. If specified, this
indicates that a maximum a posteriori (MAP) approach
will be used to set covariance function
hyperparameters. The priors are obtained by calling
prt gp priors(’covFuncName’)
N.B.: for the arguments specifying functions, pass in a string, not
a function handle. This script will generate a function handle
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.type
- which type of machine this is (here, ’classifier’)
.func val - predictive probabilties
.mu
- test latent means
.s2
- test latent variances
.loghyper - log hyperparameters
.nlml
- negative log marginal likelihood
.alpha
- GP weighting coefficients
.sW
- likelihood matrix (see Rasmussen & Williams, 2006)
.L
- Cholesky factor

19.70.6

machines\prt machine gpr.m

Run Gaussian process regression - meta-wrapper for regression with gpml
FORMAT output = prt machine gpml(d,args)
Inputs:
d
- structure with data information, with mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.testcov
- testing covariance (cell array of matrices row vectors,
each [Nte x Nte])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) or in
form of features (false)
args
- argument string, where
-h
- optimise hyperparameters (otherwise don’t)
-f iter
- max # iterations for optimiser (ignored if -h not set)
-l likfun - likelihood function:
’likErf’ - erf/probit likelihood (binary only)
-c covfun - covariance function:
’covLINkcell’ - simple dot product
’covLINglm’
- construct a GLM
-m meanfun - mean function:
’meanConstcell’ - suitable for dot product
’meanConstglm’ - suitable for GLM
-i inffun - inference function:

197

198

CHAPTER 19. LIST OF PRONTO FUNCTIONS

’prt infEP’ - Expectation Propagation
experimental args (use at your own risk):
-p
- use priors for the hyperparameters. If specified, this
indicates that a maximum a posteriori (MAP) approach
will be used to set covariance function
hyperparameters. The priors are obtained by calling
prt gp priors(’covFuncName’)
N.B.: for the arguments specifying functions, pass in a string, not
a function handle. This script will generate a function handle
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.type
- which type of machine this is (here, ’classifier’)
.func val - predictive probabilties
.mu
- test latent means
.s2
- test latent variances
.loghyper - log hyperparameters
.nlml
- negative log marginal likelihood
.alpha
- GP weighting coefficients
.sW
- likelihood matrix (see Rasmussen & Williams, 2006)
.L
- Cholesky factor

19.70.7

machines\prt machine krr.m

Kernel ridge regression
FORMAT output = prt machine svm bin(d,args)
Inputs:
d
- structure with data information, with mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) of in
form of features (false)
args
- libSVM arguments
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.func val - value of the decision function
.type
- which type of machine this is (here, ’classifier’)

19.70.8

machines\prt machine rvr.m

Relevance vector regression (training and testing)
FORMAT output = prt machine svm bin(d,args)

19.70. MACHINES

Inputs:
d
.train

- structure with data information, with mandatory fields:
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) of in
form of features (false)
args
- libSVM arguments
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.func val - value of the decision function
.type
- which type of machine this is (here, ’classifier’)

19.70.9

machines\prt machine sMKL cla.m

Run L1-norm MKL - wrapper for simpleMKL
FORMAT output = prt machine sMKL cla(d,args)
Inputs:
d
- structure with data information, with mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) of in
form of features (false)
args
- simpleMKL arguments
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.func val - value of the decision function
.type
- which type of machine this is (here, ’classifier’)
.

19.70.10

machines\prt machine sMKL reg.m

Run L1-norm MKL - wrapper for simpleMKL
FORMAT output = prt machine sMKL reg(d,args)
Inputs:
d
- structure with data information, with mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation

199

200

CHAPTER 19. LIST OF PRONTO FUNCTIONS

of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) of in
form of features (false)
args
- simpleMKL arguments
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.func val - value of the decision function
.type
- which type of machine this is (here, ’classifier’)
.

19.70.11

machines\prt machine svm bin.m

Run binary SVM - wrapper for libSVM
FORMAT output = prt machine svm bin(d,args)
Inputs:
d
- structure with data information, with mandatory fields:
.train
- training data (cell array of matrices of row vectors,
each [Ntr x D]). each matrix contains one representation
of the data. This is useful for approaches such as
multiple kernel learning.
.test
- testing data (cell array of matrices row vectors, each
[Nte x D])
.tr targets - training labels (for classification) or values (for
regression) (column vector, [Ntr x 1])
.use kernel - flag, is data in form of kernel matrices (true) of in
form of features (false)
args
- libSVM arguments
Output:
output - output of machine (struct).
* Mandatory fields:
.predictions - predictions of classification or regression [Nte x D]
* Optional fields:
.func val - value of the decision function
.type
- which type of machine this is (here, ’classifier’)

19.70.12

machines\prt rvr.m

Optimisation for Relevance Vector Regression
[w,alpha,beta,ll] = prt rvr(Phi,t)
Phi
- MxM matrix derived from kernel function of vector pairs
t
- the values to be matched
w
- weights
alpha - 1/variance for the prior part of the model
beta - 1/variance for the likelihood part of the model
ll
- the negative log-likelihood.

19.70. MACHINES

[w,alpha,beta,nu,ll]=spm rvr(K,t,opt)
K
- a cell-array of MxM dot-product matrices.
t
- the values to be matched
opt
- either ’Linear’ or ’Gaussian RBF’
’Linear’
is for linear regression models, where
the optimal kernel is generated by
[nu(1)*K1 + nu(1)*K2... ones(size(K1,1),1)]
’Gaussian RBF’ is for regression using Gaussian radial basis
functions. The kernel is generated from
P1 = nu(1)*K1 + nu(1)*K2 ... ;
P2 = repmat(diag(P1) ,1,size(P1,2)) +...
repmat(diag(P1)’,size(P1,1),1) - 2*P1;
Phi = exp([-0.5*P2 ones(size(P1,1),1)]);
w
- weights
alpha - 1/variance for the prior part of the model
beta - 1/variance for the likelihood part of the model
nu
- parameters that convert the dot-product matrices into
a kernel matrix (Phi).
ll
- the negative log-likelihood.
The first way of calling the routine simply optimises the
weights. This involves estimating a restricted maximum
likelihood (REML) solution, which maximises P(alpha,beta$|$t,Phi).
Note that REML is also known as Type II Maximum Likelihood
(ML-II). The ML-II solution tends towards infinite weights for
some the regularisation terms (i.e. 1/alpha(i) approaches 0).
The appropriate columns are removed from the model when
this happens.
The second way of calling the routine also estimates additional
input scale parameters as described in Appendix C of Tipping (2001).
This method is much slower, as a full optimisation for the scale
parameters is done after each update of the alphas and beta.
see: http://research.microsoft.com/mlp/RVM/relevance.htm
Refs:
The Relevance Vector Machine.
In S. A. Solla, T. K. Leen, and K.-R. Mller (Eds.),
Advances in Neural Information Processing Systems 12,
pp. 652-658. Cambridge, Mass: MIT Press.
Michael E. Tipping
Sparse Bayesian Learning and the Relevance Vector Machine
Journal of Machine Learning Research 1 (2001) 211-244

19.70.13

machines\prt weights.m

Run function to compute weights
FORMAT weights = prt weights(d,m)
Inputs:
d
- data structure
(fields of .d can vary depending on weights function)
m
- machine structure
.function - function to compute weights (string)
.args
- function arguments

201

202

CHAPTER 19. LIST OF PRONTO FUNCTIONS

Output:
weights - weights vector [Nfeatures x 1]

19.70.14

machines\prt weights bin linkernel.m

Run function to compute weights for linear kernel binary classifiers
FORMAT weights = prt weights bin linkernel (d,args)
Inputs:
d
- data structure
.datamat
- data matrix [Nfeatures x Nexamples]
.coeffs
- coefficients vector [Nexamples x 1]
args
- function arguments (can be empty)
Output:
weights
- vector with weights [Nfeatures x 1]

19.70.15

machines\prt weights gpclap.m

Run function to compute weights for linear multiclass classifiers
FORMAT weights = prt weights gpclap (d,args)
Inputs:
d
- data structure
.datamat
- data matrix [Nfeatures x Nexamples]
.coeffs
- coefficients vector [Nexamples x 1]
args
- function arguments (can be empty)
Output:
weights
- vector with weights Nclass[Nfeatures x 1]

19.70.16

machines\prt weights sMKL cla.m

Run function to compute weights for binary MKL
FORMAT weights = prt weights sMKL (d,args)
Inputs:
d
- data structure
.datamat
- data matrix [Nfeatures x Nexamples]
.coeffs
- coefficients vector [Nexamples x 1]
.betas
- kernel weights
.idfeat img - cell with indece
args
- function arguments (can be left empty)
Output:
weights
- vector with weights [Nfeatures x 1]

19.70.17

machines\prt weights sMKL reg.m

Run function to compute weights for binary MKL
FORMAT weights = prt weights sMKL (d,args)
Inputs:
d
- data structure
.datamat
- data matrix [Nfeatures x Nexamples]
.coeffs
- coefficients vector [Nexamples x 1]
.betas
- kernel weights
.idfeat img - cell with indece
args
- function arguments (can be left empty)
Output:

19.71. UTILS

weights

19.70.18

203

- vector with weights [Nfeatures x 1]

machines\prt weights svm bin.m

Run function to compute weights for binary SVM
FORMAT weights = prt weights svm bin (d,args)
Inputs:
d
- data structure
.datamat
- data matrix [Nfeatures x Nexamples]
.coeffs
- coefficients vector [Nexamples x 1]
args
- function arguments (can be left empty)
Output:
weights
- vector with weights [Nfeatures x 1]

19.71

utils

19.71.1

utils\prt centre kernel.m

This function centres the kernel matrix, respecting the independence of
training and test partitions. See Shawe-Taylor and Cristianini for
background on this approach.
Shawe-Taylor, J. and Cristianini, N. (2004). Kernel methods for Pattern
analysis. Cambridge University Press.

19.71.2

utils\prt checkAlphaNumUnder.m

check whether a given string is alphanumerical or underscore
FORMAT out = prt checkAlphaNumUnder(s)
Inputs:
s - a string of arbitrary length to check
Output:
out - logical 1 if the all chars in the string are alphanumerical
logical 0 otherwise
Based on isalpha num in the identification toolbox

19.71.3

utils\prt normalise kernel.m

FORMAT K normalised = prt normalise kernel(K)
This function normalises the kernel matrix such that each entry is
divided by the product of the std deviations, i.e.
K new(x,y) = K(x,y) / sqrt(var(x)*var(y))

19.71.4

utils\prt utils update mask.m

204

CHAPTER 19. LIST OF PRONTO FUNCTIONS

Part V

Bibliography

205

Bibliography
[1] M. Aizerman, E. Braverman, and L. Rozonoer. Theoretical foundations of the potential
function method in pattern recognition learning. Automation and Remote Control, 25:821–
837, 1964.
[2] Francis R. Bach, Gert R. G. Lanckriet, and Michael I. Jordan. Multiple kernel learning, conic
duality, and the smo algorithm. In Proceedings of the Twenty-first International Conference
on Machine Learning, ICML ’04, pages 6–, New York, NY, USA, 2004. ACM.
[3] Christopher M. Bishop. Pattern Recognition and Machine learning. Springer, 2006.
[4] Nello Cristianini and John Shawe-Taylor. An introduction to support Vector Machines: and
other kernel-based learning methods. Cambridge University Press, New York, NY, USA,
2000.
[5] Volkmar Glauche. MATLAB batch system. http://sourceforge.net/projects/matlabbatch/.
[6] J. V. Haxby, M. I. Gobbini, M. L. Furey, A. Ishai, J. L. Schouten, and P. Pietrini. Distributed
and overlapping representations of faces and objects in ventral temporal cortex. Science,
293(5539):2425–2430, 2001.
[7] J. D. Haynes and G. Rees. Decoding mental states from brain activity in humans. Nat. Rev.
Neurosci., 7:523–534, 2006.
[8] Thomas Hofmann, Bernhard Scholkopf, and Alexander J. Smola. Kernel methods in machine
learning. Annals of Statistics, 36(3):1171–1220, 2008.
[9] A. Marquand, M. Howard, M. Brammer, C. Chu, S. Coen, and J. Mourao-Miranda. Quantitative prediction of subjective pain intensity from whole-brain fMRI data using Gaussian
processes. Neuroimage, 49:2178–2189, 2010.
[10] Members and collaborators of the Wellcome Trust Centre for Neuroimaging. Statistical Parametric Mapping, SPM8. http://www.fil.ion.ucl.ac.uk/spm, 2008. Wellcome Trust Centre for
Neuroimaging, University College London, UK.
[11] Janaina Mourao-Miranda, Karl J Friston, and Michael Brammer. Dynamic discrimination
analysis: a spatial-temporal svm. Neuroimage, 36(1):88–99, 2007.
[12] K. A. Norman, S. M. Polyn, G. J. Detre, and J. V. Haxby. Beyond mind-reading: multi-voxel
pattern analysis of fMRI data. Trends Cogn. Sci. (Regul. Ed.), 10:424–430, 2006.
[13] F. Pereira, T. Mitchell, and M. Botvinick. Machine learning classifiers and fMRI: a tutorial
overview. Neuroimage, 45:199–209, 2009.
[14] Alain Rakotomamonjy, Francis Bach, Stephane Canu, and Yves Grandvalet. Simplemkl.
Journal of Machine Learning Research, 9:2491–2521, 2008.
[15] Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian Processes for Machine
Learning. Adaptive Computation and Machine Learning. the MIT Press, 2006.
[16] J. Schrouff, M. J. Rosa, J. M. Rodina, A. F. Marquand, C. Chu, J. Ashburner, C. Phillips,
J. Richiardi, and J. Mourao-Miranda. PRoNTo: Pattern Recognition for Neuroimaging
Toolbox. Neuroinformatics, 11(3):319–337, 2013.
207

208

BIBLIOGRAPHY

[17] Jessica Schrouff, Julien Cremers, Gaetan Garraux, L Baldassarre, Janaina Mourao-Miranda,
and Christophe Phillips. Localizing and comparing weight maps generated from linear kernel
machine learning models. In Pattern Recognition in Neuroimaging (PRNI), 2013 International Workshop on, pages 124–127. IEEE, 2013.
[18] Jessica Schrouff, Joao Monteiro, Maria Rosa, Liana Portugal, Christophe Phillips, and
Janaina Mourao-Miranda. Can we interpret linear kernel machine learning models using
anatomically labelled regions? In Organization for Human Brain Mapping, 2014.
[19] John Shawe-Taylor and Nello Cristianini. Kernel Methods for Pattern Analysis. Cambridge
University Press, 2004.
[20] Michael E. Tipping. Sparse bayesian learning and the relevance vector machine. Journal of
Machine Learning Research, 1:211–244, 2001.

Source Exif Data:

File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.5
Linearized                      : No
Page Count                      : 208
Page Mode                       : UseOutlines
Author                          : The PRoNTo Team
Title                           : PRoNTo Manual
Subject                         : Pattern Recognition in Neuroimaging Toolbox
Creator                         : LaTeX with hyperref package
Producer                        : pdfTeX-1.40.18
Keywords                        : neuroimaging, data mining, multivariate analysis, pattern recognition
Create Date                     : 2018:03:26 10:21:11+01:00
Modify Date                     : 2018:03:26 10:21:11+01:00
Trapped                         : False
PTEX Fullbanner                 : This is MiKTeX-pdfTeX 2.9.6499 (1.40.18)

EXIF Metadata provided by EXIF.tools

PRoNTo Manual Prt

prt_manual

Navigation menu

Versions of this User Manual:

Views

Navigation