Unsupervised short-term covariate shift minimization for self-paced BCI

A major challenge for Brain Computer Interface systems (BCIs) is dealing with non-stationarity in the EEG signal. There are two types of EEG non-stationarities 1) long-term changes related to fatigue, changes in recording conditions or effects of feedback training which is addressed in classification step and 2) short-term changes related to different mental activities and drifts in slow cortical potentials which can be addressed in the feature extraction step. In this paper we use a covariate shift minimization (CSM) method to alleviate the short-term (single trial) effects of EEG non-stationarity to improve the performance of self-paced BCIs in detecting foot movement from the continuous EEG signal. The results of applying this unsupervised covariate shift minimization with two different classifiers, linear discriminant analysis (LDA) and probabilistic classification vector machines (PCVMs) along with two different filtering methods (constant bandwidth and constant-Q filters) show the considerable improvement in system performance.