Classification of Three-Class Motor Imagery EEG Data by Combining Wavelet Packet Decomposition and Common Spatial Pattern

Low information transfer rate is inherent in a binary brain-computer interface (BCI) and largely limits its practical application. To increase information transfer speed, it is necessary to put emphasis on the research of multi-task BCIs. This paper proposes a new algorithm for classifying single-trial motor imagery EEG data in a three-task BCI. Wavelet packet decomposition (WPD) and common spatial pattern (CSP) are respectively applied to lowpass (0-64Hz) and bandpass (8-30 Hz) filtered data to extract discriminative features. The two feature vectors are reduced to two dimensions by Fisher discriminant analysis (FDA) that is followed by a support vector machine (SVM) for classification. The algorithm was applied to three datasets recorded during BCI experiments of three-class motor imagery tasks. The classification accuracies for these three datasets range from 95.6% to 88.1% and their mean is 90.6%. The results verify the feasibility and validity of the algorithm.