Automatic Detection of Transitions in Sleep Stage Using Single Overnight Triaxial Accelerometry and Heart Rate

Accurate detection of sleep stage transitions is crucial for evaluating brain health, memory consolidation, and physiological stability, but current detection methods are resource-intensive, which makes it inaccessible to the general population. In this study, we develop an algorithm to identify sleep stage transitions between wake, REM, and non-REM (NREM) sleep with high temporal accuracy in a publicly available dataset of thirty healthy participants who wore an Apple watch while undergoing an overnight PSG sleep study. Specifically, we train a random forest model to use the heart rate and triaxial acceleration from the Apple watch to predict the times of sleep stage transitions within a 5-minute tolerance. When employing 10-fold cross-validation, our model achieved an average Jaccard index of 0.54 ([0.47-0.60]) with a 5-minute tolerance. This study demonstrates that wearable sensor data can effectively predict sleep stage transitions with reasonable temporal precision.Clinical Relevance- Detecting sleep transitions with a non-invasive wearable device enables more accessible measurement of sleep continuity and quality, leading to improved patient-centered diagnosis and treatment of sleep disorders.