A two-steps sleep/wake stages classifier taking into account artefacts in the polysomnographic signals

This paper focuses on the development of an automatic system for sleep analysis. The system proposed in this paper combines two phases needed in sleep analysis. In a first step, an artefact detection system selects the polysomnographic signals (EEG, EOG, EMG) that are not corrupted by artefacts. In a second step, relevant features are extracted from the selected signals and classified using a neural network chosen among a bank of four neural networks. The four classifiers differ one from the others by the signals used for the classification. They were learnt using information provided by different combination of signals (EEG, EEG+EOG, EEG+EMG, EEG+EOG+EMG). Thus, the complete system enables the classification to be performed using relevant features computed from artefact-free signals, without losing too many data. The performance reached by the two-steps system is 85% of accuracy, calculated on 47 night sleep recordings.