Automated detection of sleep-boundary times using wrist-worn accelerometry

Abstract

ObjectiveCurrent polysomnography-validated measures of sleep status from wrist-worn accelerometers cannot be used in fully automated analysis as they rely on self-reported sleep-onset and -end (sleep-boundary) information. We set out to develop an automated, data-driven approach to sleep-boundary detection from wrist-worn accelerometer data.\n\nMethodsOn three separate occasions, participants were asked to wear a GENEActiv(R) wrist-worn accelerometer for nine days and concurrently complete sleep diaries with lights-off, asleep and wake-up information. We developed and evaluated three data-driven methods for sleep-boundary detection: a change-point detection based method, a thresholding method and a random forest classifier based method. Mean absolute errors between automatically-derived and self-reported sleep-onset and wake-up times were recorded in addition to kappa statistics for the minute-by-minute performance of each of the methods.\n\nResults46 participants provided 972 days of accelerometer recordings with corresponding self-reported sleep information. The three sleep-boundary detection methods resulted in mean absolute errors in sleep-onset and wake-up times per individual of 36 min, 34 min and 33 min and kappa statistics of 0.87, 0.89 and 0.89, respectively.\n\nConclusionOur methods provide a data-driven approach to detect sleep-onset and -end times without the need for self-reported sleep-boundary information. The methods are likely to be of particular use for large-scale studies where the collection of self-reported sleep diaries is impractical.\n\nSignificanceObjective measures of sleep are needed to reliably detect associations with health outcomes. This work lays the foundation for studies of objectively measured sleep duration and its health consequences in large studies.