Device agnostic sleep-wake segment classification from wrist-worn accelerometry

Abstract

Robustness for sleep-wake segment classification from accelerometry is critical when considering deployment in clinical studies. Deployed devices may change between or within studies, which alters critical clinical endpoints and negatively impact paired analyses if inter-device robustness is not assured. Here we present a neural network algorithm, deep learning sleep (DLS), for the classification of sleep-wake segments and show its robustness to different wrist-worn devices. Our results show that DLS delivers high accuracy when predicting sleep-wake segments in inter-device cross-validation experiments.