Outliers resistant methods for motor imagery classification

"Common spatial patterns analysis (CSP) and linear discriminant analysis (LDA) are widely used techniques for spatial filtering and classifying in motor imagery (MI). However, CSP is very sensitive to noise and artifacts. A method to detect and eliminate anomalous electroencephalogram (EGG) signals before applying CSP is presented. An outlier score of the signal is obtained by calculating the similarities with the other signals of the sample through the Bounded Coordinate System (BCS). Besides, it is proposed to replace the usual estimators of mean, covariance and scale, used in the algorithms, by Olive and Hawkins estimators to get robust versions of BCS and CSP. The assumption done in LDA that the covariance of each of the classes in MI are identical may not be true; if it is not satisfied, it is better to use quadratic discrimination. Tests to verify this hypothesis and decide which discriminant function must be used are considered. The performances of the methods are evaluated and compared on EGG data from BCI competition datasets; results show that robust methods outperformed classical techniques, especially for subjects with poor classification accuracy."