Automated Sleep Staging on Wearable EEG Enables Sleep Analysis at Scale

This study presents automated sleep staging on a large number of sleep electroencephalography (EEG) recordings collected using the Muse S headband. Two recent deep learning models; a single-channel Deep Sleep Net (DSN) and a multi-channel Muse Net (MNet) were evaluated on a 5-class sleep stage classification task on 200 expert-labelled overnight sleep EEG recordings. The learned representations of the models were visualized using uniform manifold approximation projection (UMAP). Moreover, a large scale analysis of the relationship between sleep stage distribution of non-rapid eye movement (NREM) and rapid eye movement (REM) sleep with age was performed on 1020 unlabelled EEG recordings. The results showed that the proposed models achieved high accuracy (DSN: 85.2%, MNet: 86.3%) and Cohen's Kappa (DSN: 0.77, MNet: 0.79) indicating substantial agreement with human expert sleep scoring. Furthermore, the features learned by the deep neural networks showed a sleep continuum beyond the traditionally used sleep stages. Hypnogram analysis revealed a decrease in percentage of NREM 3 and REM sleep with increasing age, and an increase in percentage of NREM 2 sleep with increasing age. The results suggested that a 4-channel wearable EEG headband provides low-cost and powerful means to automatically score and analyze sleep at a large scale.