An empirical bayesian framework for brain-computer interfaces

Abstract

Current brain-computer interface (BCI) systems suffer from high complex feature selectors in comparison to simple classifiers. Meanwhile, neurophysiological and experimental information are hard to be included in these two separate phases. In this paper, based on the hierarchical observation model, we proposed an empirical Bayesian linear discriminant analysis (BLDA), in which the neurophysiological and experimental priors are considered simultaneously; the feature selection, weighted differently, and classification are performed jointly, thus it provides a novel systematic algorithm framework which can utilize priors related to feature and trial in the classifier design in a BCI. BLDA was comparatively evaluated by two simulations of a two-class and a four-class problem, and then it was applied to two real four-class motor imagery BCI datasets. The results confirmed that BLDA is superior in accuracy and robustness to LDA, regularized LDA, and SVM.