Longitudinal cardiorespiratory wearable sleep staging in the home

Introduction There is a growing interest in performing automated, longitudinal tracking of sleep in the home using wearables and machine learning (ML). Wearables such as smart watches or chest patches can be comfortably worn for long periods, and cardiorespiratory waveforms measured by these wearables combined with ML models to estimate sleep state. However, the performance of these ML models is typically assessed using retrospective data from polysomnography, which is traditionally performed in a sleep lab. Importantly, polysomnography involves monitoring cardiorespiratory waveforms with bulky, specialized equipment, typically chest bands placed around the circumference of the upper toro and diaphragm, rather than with wearables. The performance of ML models evaluated on cardiorespiratory data from polysomnography may therefore not be representative of model performance when the cardiorespiratory input signals are acquired using modern wearable devices, where signal quality and available signal modalities may be more limited. Further work is needed to validate these ML models in their intended scenario of use: longitudinal, wearable sleep monitoring in the home. Methods This paper establishes and validates a pipeline for longitudinal cardiorespiratory sleep monitoring in the home using data from the RESTORE study. In RESTORE, 17 participants with a sleep-related condition (insomnia and depressive symptoms) underwent a sleep-related clinical intervention (sleep restriction therapy). Participants simultaneously wore a low-density home electroencephalogram device, allowing for expert, manual sleep staging using brain activity, as well as a wearable peel-and-stick chest patch, allowing for wearable monitoring of cardiorespiratory waveforms. Both devices were worn by participants for 10 nights while undergoing treatment at home. A state-of-the-art cardiorespiratory sleep staging model, combining transformer and convolutional neural networks, was then tuned and tested on the wearable data using leave-one-subject-out-cross-validation. Results After transfer learning, the cardiorespiratory sleep staging model had an accuracy of 77.1% and Cohen's Kappa of 0.679 for four-class sleep staging. Further, the model was able to accurately track sleep and sleep-derived metrics longitudinally while participants underwent sleep restriction therapy. Discussion These results represent one of the first direct demonstrations of the potential for wearable, cardiorespiratory sleep staging to track longitudinal, clinically relevant changes in sleep in individuals undergoing a sleep-related intervention in the home.