Comparison of Designs Towards a Subject-Independent Brain-Computer Interface based on Motor Imagery F. Lotte, C.T. Guan, K.K. Ang Brain-Computer Interfaces laboratory Institute for Infocomm Research (I2R) Singapore EMBC’09, 2-6 September 2009, Minneapolis, USA
1
Introduction
BCI are promising interaction devices… … but are not ready yet for practical use by the general public A main limitation: BCI are Subject-Specific (SS)
Need from a lot of EEG data from the target user Long and inconvenient process
An ideal BCI should be Subject-Independent (SI)
No training Immediate use, from the very first time EMBC’09, 2-6 September 2009, Minneapolis, USA
2
Objectives
Start exploring the concept of SubjectIndependent BCI
Compare several designs of SI-BCI
Different feature extraction methods Different classifiers
EMBC’09, 2-6 September 2009, Minneapolis, USA
3
Subject-independent BCI design Training EEG signals: Calibrate the BCI
Testing EEG signals: Use the BCI
A
A
Subject Specific BCI
A Subject Independent BCI
B
E
F
G
…
Training EMBC’09, on the pooled data set 2-6 September 2009, Minneapolis, USA
4
Features explored
Band Power (BP) Features [Pfurtscheller01]
AutoRegressive (AR) modeling (order 4)
Power in 8-13 Hz (mu) + 13-30 Hz (beta) bands Common Average Reference (CAR) Band pass filtering in 8-30 Hz Surface Laplacian
Power Spectral Densities (PSD)
In 8-30 Hz – 2Hz resolution CAR 40 features selected EMBC’09, 2-6 September 2009, Minneapolis, USA
5
Features explored
Common Spatial Patterns (CSP) [Ramoser00]
Spatial filter to maximize the class separability Pre-filtering in 8-30 Hz (mu+beta) 3 pairs of features selected
Filter Bank Common Spatial Patterns (FBCSP) [Ang08]
Principle
CSP for multiple frequency bands BCI competition IV winning algorithm on all EEG data sets
9 bands of 4 Hz (from 4 to 40 Hz) 4 pairs of features selected EMBC’09, 2-6 September 2009, Minneapolis, USA
6
Classifiers explored
Linear Discriminant Analysis (LDA)
Quadratic Discriminant Analysis (QDA)
Linear classifier Quadratic classifier
Gaussian Mixture Model (GMM)
Non linear classifier Here, 3 Gaussians to model each class
EMBC’09, 2-6 September 2009, Minneapolis, USA
7
Other: a-priori based BCI
BCI based only on motor imagery knowledge [Pfurtscheller01] Preprocessing
2 Features
Common Average Reference Average Band Power in 8-30 Hz over the left (5 electrodes) and right motor cortices (5 electrodes)
Classification
Sign of difference between the two activities EMBC’09, 2-6 September 2009, Minneapolis, USA
8
EEG data set
BCI competition IV, data set 2a
4 class motor imagery (left hand, right hand, foot, tongue)
Only the left hand and right hand classes are kept
9 subjects 22 electrodes 72 trials per class
EMBC’09, 2-6 September 2009, Minneapolis, USA
9
Study 1: Comparisons between features and classifiers Classifier
Type
CSP
FBCSP
BP
AR
PSD
Mean
Specific
74.76
81.10
73.77
69.52
68.98
73.63±4.89
Independent 65.12
62.19
67.59
63.19
65.05
64.63±2.07
73.23
79.94
64.97
63.27
64.97
69.28±7.12
Independent 61.81
61.27
63.73
57.64
56.56
60.2±3
71.45
77.24
50.62
61.65
51.31
62.44±11.85
Independent 57.56
61.11
62.11
60.73
55.48
60.03±3.52
LDA
Specific QDA
Specific GMM
Accuracy of the a-priori based BCI : 64.2 % EMBC’09, 2-6 September 2009, Minneapolis, USA
10
Study 2: Impact of frequency bands in FBCSP
FBCSP is the best for SS-BCI but the worst for SI-BCI
Why? Difference in frequency decomposition?
Exploring different frequency decompositions
Default: 9 bands of 4 Hz Medium bands: 6 bands of 5 Hz Large bands: 5 bands of 6 Hz Larger bands: 4 bands of 7 Hz Multi-Resolution: 8-30 Hz + standard rhythms (4-7 Hz, 8-13 Hz, 13-30 Hz, 30-40 Hz) + 5 bands of 6 Hz EMBC’09, 2-6 September 2009, Minneapolis, USA
11
Results Default
4 bands of 7 Hz
5 bands of 6 Hz
6 bands of 5 Hz
Multi resolution 12 features
Multi resolution 20 features
Multi resolution 40 features
Mean
Specific
81.10
79.94
81.40
79.94
80.17
80.48
79.94
80.4
Independent
62.19
65.35
67.67
66.59
68.52
70.07
70.99
67.34
Specific
79.94
79.32
81.56
78.47
80.4
80.94
78.94
79.94
Independent
61.27
64.04
67.05
68.52
69.29
66.74
68.06
66.42
Specific
77.16
77.24
80.09
80.09
79.63
76.08
77.70
75.5
Independent
61.11
64.04
64.27
62.11
63.58
66.82
65.28
63.89
LDA
QDA
GMM
EMBC’09, 2-6 September 2009, Minneapolis, USA
12
Conclusion
Subject-Independent BCI design comparisons
SI-BCI design is possible! (71% accuracy on average) Linear classifiers seem more efficient than non-linear ones for this problem Multi-resolution FBCSP is the best
Future works
Methods combination [Fazli09] Larger data base of subjects Subject-independent design + adaptation [Vidaurre05, Vidaurre08] EMBC’09, 2-6 September 2009, Minneapolis, USA
13
