AcceleRest: A Physiology-Aware Masked Autoencoder for Wrist Accelerometer-based Sleep Staging and Apnea Evaluation

Sleep is essential for physical and mental health, yet large-scale assessment of sleep stages and sleep apnea is limited by the cost and burden of clinical polysomnography. Wrist accelerometry provides a scalable lower-fidelity alternative, but its usefulness is highly reliant on modeling choices. Generative self-supervised learning has remained unexplored for this purpose. To address this, we developed and pretrained physiology-aware masked autoencoders to capture pulse-and respiration-related motion using [~]700,000 days of wrist accelerometry from 108,904 recordings in the UK Biobank cohort. This effort culminated in AcceleRest, a transformer model pretrained using a respiratory amplification objective. Performance was validated against polysomnography across 478 recordings from six cohorts and devices including independent external test cohorts. AcceleRest feature vectors enabled linear wake-NREM-REM sleep staging with a macro F1 score of 0.69 and respiratory event detection with a macro F1 of 0.56. The combined model outputs enabled sleep apnea severity evaluation with a 67% sensitivity and 96% specificity for severe apnea. Overall agreement between polysomnography and AcceleRest showed a bias of 0.8 min for total sleep duration, with 95% limits of agreement (LoA) of -101.6 to 103.2 min, and 32.5 min for REM sleep duration and 95% LoA of -68.7 to 133.6 min. These findings demonstrate that physiology-aware pretraining can enable robust and clinically meaningful sleep phenotyping from wrist accelerometers, supporting scalable screening and longitudinal monitoring of sleep health. To the best of our knowledge, AcceleRest represents the first wrist accelerometry model for joint sleep stage and apnea evaluation. All code and models will be made available upon final peer-reviewed publication.