Enhancing feature extraction with sparse component analysis for brain-computer interface.

Feature extraction is very important to EEG-based brain computer interfaces (BCI) in helping achieve high classification accuracy. Preprocessing of EEG signals plays an important role, because an effective preprocessing method will help enhance the efficiency of the feature extraction. In this paper, sparse component analysis (SCA) is employed as a preprocessing method for EEG based BCI. A combined feature vector is constructed. This feature vector consists of a dynamical power feature and a dynamical common spatial pattern (CSP) feature. The dynamical power feature is extracted from selected SCA components, while the dynamical CSP feature is extracted from raw EEG data. Using the presented preprocessing and feature extraction method, we analyze the data for a cursor control BCI carried out at Wadsworth Center. Our results show that SCA preprocessing is the most effective in extracting a component which reflects the subject's intention, and demonstrate the validity of SCA preprocessing for the enhancement of feature extraction.