Abstract
Wearable actimeters can improve our understanding of sleep in the natural environments. Current algorithms may produce inaccuracies in specific individuals and circumstances, such as quiet wakefulness. New hardware allows data collection at higher frequencies enabling sophisticated analytical methods. We have developed a novel statistical algorithm, the Wasserstein Algorithm for Classifying Sleep and Wakefulness (WACSAW), to identify behavioral states from recordings of everyday movement. WACSAW employs optimal transport techniques to identify segments with differing activity variability. Functions characterizing the segments' movement distributions were clustered into two groups using a k-nearest neighbors and labeled as sleep or wake based on their proximity to an idealized sleep distribution. It returned >95% overall accuracy validated against participant logs in the test data and performed ~10% better than a clinically validated actimetry system. We present the methodology describing how WACSAW results in a novel, individually-tuned, statistical approach to actimetry that improves sleep/wakefulness classification and provides auxiliary information as part of the calculations that can be further related to sleep-relevant outcomes.