Mixtures of common spatial patterns for feature extraction of EEG signals

The method of common spatial patterns (CSP) is an effective feature extractor for representing electroencephalogram (EEG) signals with the purpose of classification in brain computer interfaces (BCIs). However, it has two apparent demerits mainly about the estimation of covariance matrices which sacrifice its performance. First, the estimator of sample covariance is non-robust; Second, the possible high dimension of the covariance matrices makes exact parameter estimation difficult. In this paper, we propose the approach of mixtures of CSP (MCSP) to conquer this problem. MCSP constructs multiple CSP feature extractors by the bootstrap sampling method, and has the potential to improve the performance of one single feature extractor. The classification result of a new EEG sample is obtained by combining several classifier outputs corresponding to the multiple feature extractors. Experimental comparisons with the state-of-the-art method for classifying real EEG signals of motor imagery tasks show the effectiveness of the proposed MCSP method.