Adaptive Common Spatial Pattern for single-trial EEG classification in multisubject BCI

Abstract

Common Spatial Patterns (CSP) is a widely used spatial filtering method for electroencephalogram (EEG)-based brain computer interface (BCI). It is a supervised technique that needs subject specific training data. Due to the non-stationary nature of EEG, EEG signal may exhibit significant inter- and intra-subject variation. Consequently, spatial filters learned from one subject may not perform well for EEG data acquired from another subject performing a same task, or even from the same subject at a different time. Various methods have been developed to improve CSP's multisubject performance by adding regularizing terms into the learning process. Most of these methods include target subjects' training data in the CSP learning, and the trained spatial filters are fixed when applied to classification. In this work, an adaptive CSP method was proposed to classify single trial EEG data from multiple subjects. The method does not require training data from target subjects, and updates spatial filters based on target subjects' data during the classification. Three different methods were proposed to adapt the CSP learning to target subjects. Experimental results on motor imagery data indicate that the proposed method can efficiently integrate target subjects' information into the CSP learning, and provide better discrimination performance (about 20% increase in overall classification accuracy) than the standard CSP method for multisubject BCI.