A bispectrum approach to feature extraction for a motor imagery based brain-computer interfacing system

Existing feature extraction techniques for BCI systems are developed based on traditional signal processing techniques assuming that the signal is Gaussian and has linear characteristics. But the motor imagery (MI) related EEG is highly non-Gaussian, non-stationary and non-linear. This paper proposes an advanced, robust but simple feature extraction procedure for MI based BCI system. This novel approach uses higher order statistics technique, the bispectrum, and extracts the non-linear features from EEG. Along with a linear classifier (LDA), the proposed technique has been applied to an MI based BCI system. The performance (classification accuracy, mutual information and Cohens kappa) of the system is evaluated and compared with the power spectrum based BCI. It is observed that the proposed technique extracts more pragmatic information resulting in better and consistent cross-session detection accuracy and Cohens kappa. It is concluded that the bispectrum based feature extraction is a promising technique for detecting different brain states.