EEG classification for multiclass motor imagery BCI

This paper describes the method for classifying multiclass motor imagery EEG signals of brain-computer interfaces (BCIs) according to the phenomena of event-related desynchronization and synchronization (ERD/ERS). The method of one-versus-one common spatial pattern (CSP) for multiclass feature extraction was employed. And we extended two different kinds of classifiers: 1) support vector machines (SVM) based on maximal average decision value; 2) k-nearest neighbor (KNN) rule for multiclass classification. In order to testify the performance of each classifier, dataset IIa of BCI Competition IV (2008) which involved nine subjects in a four-class motor imagery (MI) based BCI experiment were used. And the final classification results showed that our extended SVM classification method based on decision value is much better than the majority voting rule, and the extended KNN performed the best.