Improvement of Error-Related Potential Monitoring in Brain–Computer Interface Based on Optimal Feature Dimensionality Selection

The burgeoning advancement of brain-computer interface (BCI) continues to enrich its applications in human-computer interaction (HCI), e.g., assistive robotics, game design, and so on. Notwithstanding this progress, the efficacy of BCI remains impeded by classification errors which cause erroneous behavior for controlled devices. In this article, we proposed an error detection method via error-related potential (ErrP) monitoring based on optimal feature dimensionality selection. During the experiment, subjects observed the movement of cursor, and the electroencephalogram (EEG) signals were recorded synchronously. ErrP would occur when they found the cursor was moving to a wrong target, which provided the interface with biological intelligence. We extracted mean average value (MAV) as time domain feature and Welch power spectrum as frequency domain feature and combined them. Least squares support vector machine (LSSVM) was adapted as the classifier and a model of the time-frequency domain features and the event categories were directly mapped afterward. For redundant feature elimination, feature dimensionality reduction (FDR) was conducted via mutual information (MutInf) criterion. The optimal feature dimensionality was selected to form the feature subsets, and the model of the time-frequency domain features and the event categories was optimized. Our average classification accuracy is up to 82.8%, which is conductive to timely error detection. Practically, our work provides a potential way in aborting unexpected error operation and adjusting the next operation continuously.