Optimal set of EEG features in infant sleep stage classification

Abstract

This paper evaluates six classification algorithms to assess the importance of individual EEG rhythms in the context of automatic classification of infant sleep. EEG features were obtained by Fourier transform and by a novel technique based on the empirical mode decomposition and generalized zero crossing method. Of six evaluated classification algorithms, the best classification results were obtained with the support vector machine for the combination of all presented features from four EEG channels. Three methods of attribute ranking were assessed: relief, principal component analysis, and wrapper-based optimized attribute weights. The outcomes revealed that the optimal selection of features requires one feature from every significant frequency band, either a spectral feature or a frequency dynamic feature. This means that reducing the number of features will have a minimal impact on the classification accuracy.