Novel feature generation and classification for a 2-state Self-paced Brain Computer Interface system

A two-state Self-paced Brain Computer Interface (SBCI) system enables users to control external devices at any time they desire by intentionally switching their mental states. In order to accurately control devices, distinguishing features representing different brain states must be present in the EEG signals. This paper introduces a novel feature generation method for EEG motor imagery data, based on Multivariate Empirical Mode Decomposition (MEMD), the Hilbert Transform and a phase synchronization index. Novelties of our approach are (1) MEMD is applied for decomposing an EEG signal into its narrow-band frequency components, from which features of the different brain states are calculated, and (2) a phase synchronization index is calculated without averaging over trials or time. We used a simple and fast classification scheme that employed an empirical threshold value obtained from the Receiver Operating Characteristic (ROC) curve of the training data. Applying the proposed method on only two mono-polar EEG channels, the SBCI system with the selected features yields a 93% True Positive rate, and a 5.8% False Positive rate using the BCI competition III data set Iva.