ACSP: Adaptive CSP filter for BCI applications

Brain Computer Interface (BCI) provides a communication channel via computer between mind and environment. Extracting suitable and discriminant features is one of the most important stages in BCI Applications. Common spatial patterns (CSP) is a well-known feature extraction method; however, due to the non-stationary nature of EEG signals CSP should be updated through time. This paper proposes a novel recursive adaptation method inspired from extended-Kalman-filter equations for CSP feature elicitation and classification. In this method, CSP filters are updated with each new EEG data. The proposed method was compared with a standard CSP method and an extended version of it, which uses incremental covariance matrices (ICM). These methods were applied to dataset `a' of BCI competition-III containing two- and multi-task imagery movements. Results demonstrate a considerable improvement in terms of classification accuracy by the proposed method in comparison with standard CSP, also the proposed method performed better or as well as CSP method with ICM in most cases.