CSP based extraction and F-score based optimization of time-frequency power features for EEG mental task classification

Aiming the mental tasks recognition tasks in brain computer interface (BCI), this paper proposes an Electroencephalography (EEG) feature extraction method which makes use of the discriminative common spatial patterns (CSP). Apart from CSP analysis's traditional application in time domain, it is extended to consider frequency domain information of EEG signal. After an artifact removal through the independent component analysis (ICA), the multichannel EEG signals are decomposed into a set of spatial patterns by CSP analysis, and the logarithmic frequency domain and time domain power distributions are calculated. Time-frequency power features are extracted on these distributions and optimized by a F-score method. Comparing with the traditional CSP methods, the proposed method not only retained the time domain variance features, but also induced the frequency band power features. Since F-score is easy and fast to calculate, and the F-score based method can quickly select more effective features from high dimensional data, depending on the importance and the discriminative ability of each data pattern. In our method, the Fscore algorithm is also used to solve the traditional CSP problems such as the definition of common pattern number. The proposed method was tested on a five-task cognitive state analysis problem, and a recognition accuracy of 89.4% was achieved, that well approved the effectiveness and versatility of the proposed method.