Unconstrained deep learning-based sleep stage classification using cardiorespiratory and body movement activities in adults with suspected sleep apnea.

This study assessed the feasibility of unconstrained deep-learning-based sleep stage classification using cardiorespiratory and body movement activities derived from piezoelectric sensors installed under a bed mattress. Heart and respiratory rates and their respective variabilities, cardiorespiratory coupling index, and body movement were simultaneously acquired through polysomnography (PSG) for 106 untreated participants with suspected sleep apnea. We used a bidirectional long short-term memory network to predict the five sleep stages using five different input feature combinations. The best performance was achieved with a model comprising six parameters, including cardiorespiratory variability features, with a balanced accuracy of 0.70 ± 0.05, Cohen's κ of 0.40 ± 0.12, and an F1 score of 0.62 ± 0.08. Deming regression and Bland-Altman analyses of the six major sleep parameters estimated by the model and those determined by PSG showed significant correlations (r = 0.426-0.695) with a low bias. These results demonstrate the effectiveness of the proposed approach and its potential to expand opportunities for in-home sleep monitoring.