Obviating Session-to-Session Variability in a Rapid Serial Visual Presentation-Based Brain-Computer Interface

Informative patterns of neural data obtained from electroencephalography (EEG) can be classified by machine learning techniques to improve performance of human-computer interaction. A rapid serial visual presentation (RSVP)-based brain-computer interface (BCI) system relies on single-trial classification of event-related potentials (ERP) to categorize target and non-target images. The system works well in well-controlled laboratory settings; however, transitioning this approach into more dynamic, unconstrained environments poses several significant challenges. One major challenge is how to address the problem of session-to-session variability in EEG decoding. For a new session, a time-consuming and laborious calibration procedure is usually required to collect sufficient individual data for training a new classifier. This paper employed a subspace decomposition algorithm, Signal-to-noise ratio Maximizer for event-related potentials (SIM), to improve the session-to-session transfer performance of the RSVP-based BCI system. EEG data were collected from 17 subjects, each of whom performed two task sessions on two different days. Compared with the standard hierarchical discriminant component analysis (HDCA) algorithm, the classification performance was significantly improved by combining the SIM algorithm with the HDCA algorithm. The mean area under the receiver operating characteristic curve (AUC) across all subjects was improved from 0.7242 to 0.8546. The results suggest that the proposed approach efficiently obviates the session-to-session variability in task-related ERP signals and thereby facilitates system calibration of the RSVP-based BCIs.