Ensemble learning based on overlapping clusters of subjects to predict microsleep states from EEG

Microsleeps are brief and involuntary instances of complete loss of sleep-related consciousness. We present a novel approach of creating overlapping clusters of subjects and training of an ensemble classifier to enhance the prediction of microsleep states from EEG. Overlapping clusters are created using Kullback-Leibler divergence between responsive state features of each pair of training subjects. Highly correlated features within each overlapping cluster are discarded. The remaining features are selected via Fisher score based ranking followed by an average of 5-fold cross-validation areas under the curves of receiver operating characteristics (AUCRoc) of a linear discriminant analysis (LDA) classifier. The decisions of LDA classifiers on overlapping clusters are fused using weighted average. We evaluated this new approach on 16-channel EEG data from 8 subjects who had performed a 1-D visuomotor task for two l-h sessions. Joint entropy features were extracted from a 5-s window of EEG with steps of 0.25 s Test performances were evaluated using leave-one-subject-out cross-validation. Our ensemble of overlapping clusters of subjects achieved a mean prediction performance, phi, of 0.42 compared with 0.39 for a single LDA classifier and 0.37 for generalized stacking.