A Cross-Subject SSVEP-BCI Based on Task Related Component Analysis

Abstract

SSVEP-BCIs have attracted extensive attention because of high information transfer rate. High-speed BCIs need to collect sufficient user’s own data to train optimal subject-specific parameters. However, one of the challenges which limits the real-life application of BCIs is the timeconsuming and tiring calibration process. This study developed two cross-subject frameworks. One of them uses data from all training subjects to train task-related component analysis based spatial filters (all-to-one, A2O), and the other uses data from each training subject to train task-related component analysis based spatial filters (one-to-one, O2O). Both of them do not need calibration process for a new user. The study further proposed O2O with threshold (O2O-Thr) to increase the reliability of recognition process. The proposed strategies can exploit information from existing subjects’ SSVEP data and transfer it to new users. The performance of these methods was compared using an 8-class SSVEP dataset recorded from 10 subjects. O2O-Thr achieves average accuracy of 94.6% with data length of 1.5 seconds. The proposed methods have great potential for building subject-independent BCI that do not require any calibration data from new users, which make BCI more practical and user-friendly.