Shrinkage Common Spatial Pattern for Feature Extraction in Brain-Computer Interface

Common spatial pattern (CSP) has been one of the most popular methods for EEG feature extraction in brain-computer interface (BCI) application. Although the CSP usually provides good discriminant features for classification, it is also known to be sensitive to overfitting and noise. This study introduces a shrinkage technique to regularize estimation of the covariance matrices in the CSP and hence a novel shrinkage CSP (SCSP) method, which could effectively alleviate the effects of small training sample size and unbalanced data on classification. The proposed SCSP is validated on feature extraction of P300 that has been widely adopted for BCI development. Classification accuracies are evaluated by using linear discriminant analysis (LDA) with experimental EEG data from seven subjects. The results indicate that the proposed SCSP extracts more effective features that yield higher classification accuracy than that by the traditional CSP.