Comparison of three methods for adapting LDA classifiers with BCI applications

Due to the non-stationarity of electroencephalogram (EEG) signals, online training and adaptation is essential to EEG based brain-computer interface (BCI) systems. Three methods were used to adapt linear discriminant analysis (LDA) classifiers during simulated online training for a comparative study. One method generates a new classifier based on updated means and variances of the BCI data of different classes, and the other two are Kalman filter and extended Kalman filter based methods that adapt LDA's parameters directly. Cue-based motor imagery BCI experiments were carried out with 9 naive subjects. Results show that all methods returned comparable improvement during online training, but the mean-variance updating based method is much simpler than the other two methods.