Comparison of decision rules for automatic EEG classification

Research in automated EEG classification has frequently focused on parametric linear discriminant analysis. In this paper we discuss eight classification rules, four based on parametric Gaussian assumptions and four based on nonparametric k-nearest neighbor density estimation, which were tested on human EEG samples representing seven forms of mental activity. With a set of primary EEG features, the k-NN rules, as a class, were significantly more effective than the parametric classifiers; best results were obtained with an optimized version of the generalized k-NN rule. With a reduced set of secondary features, the two types performed approximately equally, but below the best k-NN performances in the original space.