Wavelet Common Spatial Pattern in asynchronous offline brain computer interfaces

Abstract

Common Spatial Pattern (CSP) is the most popular technique used in the brain computer interfacing (BCI) context. It aims to decorrelate electroencephalography (EEG) signals obtained from different electrodes. This increases the accuracy of synchronous systems wherein users are only able to send signals during certain predefined window lengths. However CSP has poor performance and accuracy in true asynchronous BCI systems wherein users send commands at any time during the interaction with various durations of imaginations. This is more desirable and flexible compared to the synchronous version. CSP's poor performance is mainly due to the nonstationary properties of EEG signals and the need to incorporate flexible control timing. A new Wavelet Common Spatial Pattern (WCSP) technique is introduced in this paper in which EEG signals are decomposed using wavelet packets. For each packet discriminatory features are extracted and discriminative packets are selected for each subject using fuzzy logic. Classification results indicate WCSP outperforms CSP for the true asynchronous BCI system with an average Kappa increase of 0.4. The apparent poor performance of CSP in this study compared to other published articles was a consequence of using a true asynchronous system, i.e. having variable imagination durations.