Spatially Regularized Common Spatial Patterns for EEG Classification Fabien LOTTE, Cuntai GUAN, Institute for Infocomm Research (I2R), A*STAR, Singapore Introduction

Method: CSP & SRCSP

• Brain-Computer Interfaces (BCI) are communication systems that enable users to send commands to computer by using only their brain activity

CSP Goal: extremizing

• Common Spatial Patterns (CSP) is a famous BCI feature extraction method to learn spatial filters maximizing the discriminability of two classes [1] BUT • • •

wC1w T wC2 w Ci: EEG spatial covariance matrix for class i

• In this paper, we propose a Spatially Regularized CSP (SRCSP) to address these issues. SRCSP extends and improves CSP by including spatial priors in the learning process, under the form of a regularization term.

Evaluation • EEG data sets used • BCI competition III, data set IVa • 5 subjects, 118 EEG electrodes • Right hand and foot motor imagery • Training sets ranging from large (~250 trials) to small (28 trials) • BCI competition IV, data set IIa • 9 subjects, 22 EEG electrodes • Left hand and right hand motor imagery • 72 trials per class for training and testing • EEG signals contaminated by eye movements artifacts

• Classifier • Linear Discriminant Analysis

wC w wC w + αP(w)

T

It is highly sensitive to noise [2] It is prone to overfitting [3] It learns spatial filters but ignores the electrodes spatial location

• Spatial prior • Neighboring neurons are responsible for SRCSP similar brain functions • (close) neighboring electrodes should measure similar brain signals Goal: maximizing • (close) neighboring electrodes should have T similar contributions in the spatial filters 1 • Regularization term T • Penalizes spatial filters that are not 2 spatially smooth [4], i.e., whose neighboring electrodes have very different weights Regularization term and

w:spatial filter

P ( w) = ∑ G (i, j )( wi − w j )

(spatial prior)

T

wC2 w T wC1w + αP(w)

i, j proximity of two electrodes

,

weight difference between electrodes

Results Test set classification accuracies (%): BCI competition III, data set IVa Subject CSP

A1

A2

A3

A4

A5

BCI competition IV, data set IIa B1

B2

B3

B4

B5

B6

B7

B8

B9

66.07 96.43 47.45 71.88 49.6 88.89 51.39 96.53 70.14 54.86 71.53 81.25 93.75 93.75

SRCSP 72.32 96.43

60.2

2

77.68 86.5 88.89 63.19 96.53 66.67 63.19 63.89 78.47 95.83 92.36

Mean 73.1 78.7

Spatial filters (electrode weights) obtained:

Discussion and conclusion • SRCSP is an extension of CSP exploiting spatial priors • SRCSP is substantially more efficient than CSP • SRCSP is particularly efficient for subjects with poor initial performances (e.g., A3, A5, B2, B5: accuracy gain of 10% or more) • SRCSP leads to smoother and more interpretable spatial filters

References [1] Blankertz et al, IEEE Sig. Proc. Magazine, 2008 [2] Grosse-Wentrup et al, IEEE TBME, 2008

[3] Reuderink et al, TR, 2009 [4] Xiang et al, NIPS 2009