Combination of amplitude and phase features under a uniform framework with EMD in EEG-based Brain-Computer Interface.

In a Brain-Computer Interface (BCI) system, the variations of the amplitude and the phase in EEG signal convey subjects' movement intention and underpin the differentiation of the various mental tasks. Combining these two kinds of information under a uniform feature extraction framework can better reflect the brain states and potentially contribute to BCI classification. Here the Common Spatial Pattern (CSP) and the Phase Locking Value (PLV) were used to capture the amplitude and the phase information. To integrate these two feature extraction procedures, the Empirical Mode Decomposition (EMD) is introduced in preprocessing which behaved as filter bank to optimize bands selection automatically for CSP and exactly calculate the instantaneous phase for PLV. The most discriminative features were selected from the feature pool by the sequential floating forward feature selection method (SFFS). The proposed method was applied to both public and recorded datasets (each n=4). Compared with the traditional CSP, the average increment of classification accuracy is 5.4% (2.0% for public and 8.7% for recorded datasets), which both manifests statistically significances (p<0.05). Moreover, we preliminarily investigate the possibility of the online realization of this method and it shows a comparable result with the offline result.