Sleep staging using semi-unsupervised clustering of EEG: Application to REM sleep behavior disorder

Abstract

ObjectiveTo develop a semi-unsupervised automatic sleep staging method capable of detecting novel sleep patterns beyond standard sleep stages in the EEG, as may be seen in clinical populations. MethodsWe employed a two-step approach that utilized prior knowledge extracted from labeled data to cluster unlabeled data into standard sleep stages and potential emergent EEG patterns. In the first step, 62 standard EEG features per channel from 30 s labelled epochs are obtained. Subsequently, the subset of features that best discriminate each standard sleep stage from all other stages (the most discriminative features -- MDF) are determined. In the second step, applied to unlabeled data, iterative binary clustering performed in the stage-specific MDFs, with the staging order and initial cluster centers that are obtained in the first step. In the first experiment, we tested the performance of the method on EEG data from 20 healthy subjects and compared its performance with state-of-the-art methods on the same dataset. In the second experiment, we applied the method to the data from 20 subjects with rapid eye movement (REM) sleep behavior disorder (RBD) utilizing the prior knowledge derived from 9 healthy subjects. ResultsResults of the first experiment showed that the proposed method could provide comparable performance to other semi-supervised methods across all sleep stages. Results of the second experiment showed that the prior knowledge inferred from healthy participants were transferable to RBD populations with a minimal performance drop. In addition, the step-wise binary clustering beyond standard sleep stages resulted in the discovery of a novel EEG characteristic in subjects with RBD. This was predominately seen in NREM2-3 stages and was characterized by significantly lower power in the delta band and significantly higher power in alpha, beta, theta, and sigma bands compared to normal NREM2-3. ConclusionOur results suggest that the proposed approach may fill an important gap in the situation where labels of the target data are not readily available for a fully supervised approach, but some prior knowledge is still available from related data. SignificanceLabels from healthy data can be used to still allow for investigation of clinical populations, with possible discovery of novel sleep patterns.