Ultra-wideband Radar-based Sleep Stage Classification in Smartphone using an End-to-end Deep Learning

As an increasing number of people suffer from sleep disorders, such as insomnia or sleep apnea, sleep monitoring and management using consumer devices have gained increasing attention from research communities. As sleep quality is closely related to sleep structure based on hypnograms, the classification of sleep stages over the course of the night is important for accurate sleep monitoring. We present sleep stage classification using a smartphone equipped with ultra-wideband (UWB) radar. We focused on the development of easily accessible sleep monitoring system for the general population by placing the smartphone on a table near a bed, which is commonly used during sleep. We collected 509 nights of UWB radar and nocturnal in-laboratory polysomnography (PSG) data from various participants, including patients with apnea, using a customized Samsung Galaxy smartphone with a UWB radar chip placed on a table near the bed. A combination of 1D convolutional neural network and transformer architecture was proposed in this study, and a domain adaptation technique was applied to train the model with both large-scale respiratory signals from open database PSGs and UWB radar data to boost the performance by overcoming the lack of UWB radar data. With 5-fold validation, an epoch-by-epoch comparison between the predicted and expert-annotated four sleep stages (Wake, REM sleep, light sleep, and deep sleep) resulted in 0.76 of accuracy and 0.64 of Cohen's kappa. This study demonstrated that sleep stages can be monitored with substantial accuracy by simply placing a smartphone on a bedtable, making it highly usable and reliable in real use cases.