Online and Offline Implementation of Time-frequency Template Matching Method for Classifying Motor Imagery in Brain Computer Interface

The purpose of this study was to develop and online evaluate an effective feature-extraction method for distinguishing signal patterns of scalp EEG during imageries of left and right hand movements. Time-frequency template based on Morlet wavelet decomposition and event-related (de) synchronization patterns was extracted from two electrodes. The template matching index was calculated based on correlation coefficient with the template, which was further classified by linear discriminant analysis classifier to distinguish the imagined movements. The 9-subject left/right motor imagery data from NIPS2001 (Neural Information Processing System) competition were used to test the algorithm offline. The tenfold cross-validation averaged accuracy was 79.3%, better than previously reported accuracy. A pilot experiment was subsequently performed online on four subjects. Three of the four subjects achieved an accuracy of more than 70% at the end of the 15-minute training session.