Spatio-temporal discrepancy feature for classification of motor imageries

Abstract

Abstract Brain-computer interface (BCI) is an alternative pathway for communication between the brain and the outside world. Electroencephalography (EEG) records electrical signal that can reveal the mental state of the brain. EEG-based motor imagery classification is an important branch of BCI research. ERS (event-related synchronization) and ERD (event-related desynchronization) based features have been widely employed to classify motor imageries. Meanwhile, various methods derived from ERP (event-related potential) have been developed. Previous studies have demonstrated that ERP and ERD could provide complementary information about brain activity, so approach combining them is expected to give better performance for motor imagery classification. In this study, a novel variant of ERP called spatio-temporal discrepancy feature (STDF) is proposed, which evaluates the difference of the EEG signals from the left and the right sensorimotor area. With STDF, the noise which affects the left brain signal and the right brain signal simultaneously could be suppressed and the signal difference between the left and the right sensorimotor areas could be enhanced, which will certainly benefit the left vs. right motor imagery classification. STDF as a temporal feature is then combined respectively with three kinds of frequency features describing the ERD/ERS phenomenon to further improve the classification performance. Experiments on BCI competition IV dataset 2a and 2b have been conducted. Different features and state-of-the-art methods have been compared and the proposed STDF based method has obtained the best performance.