Automated sleep–wake staging combining robust feature extraction, artificial neural network classification, and flexible decision rules

The classification of sleep–wake stages suffers from poor standardization in scoring criteria and heterogeneous conditioning of polysomnographic signals. To improve applicability of fully automated sleep staging, we have designed a formal classification framework to rigorously (1) select robust candidate features, (2) emulate artificial neural network classifiers, and (3) assign sleep–wake stages using flexible decision rules. An extensive database of 48 PSG records scored in 20 s epochs by two independent clinicians was used. A small subset of 2 s elementary epochs representative of each stages with unequivocal expert scores was selected to form a limited set of learning exemplars. From 16 statistical, spectral and non-linear candidate features extracted in 2 s epochs from EEG and EMG signals, a sequential forward search selected an optimal set of five features with a 22% error rate. Multiple layer perceptrons were trained from this optimal feature set while classification accuracy was assessed using the unequivocal instance subset. A simple majority vote among 10 consecutive classifier outputs ensured a final scoring resolution comparable to that of the experts. Poor classification performance was obtained for movement time, wakefulness, and intermediate sleep stages with a 36±15% error rate (Cohen's kappa 0.48±0.18). In contrast, deep and paradoxical sleep was classified with an 82% accuracy not far from inter-expert expert agreement (83±3%). Significant improvements should be expected using a larger learning set compensating for a high inter-individual variability, and decision rules incorporating more domain-knowledge. Copyright © 2009 John Wiley & Sons, Ltd.