An Efficient Segmentation Algorithm to Estimate Sleep Duration from Actigraphy Data

Sleep duration is a recognized determinant of mental health, obesity and cardiovascular disease, cognition, and memory across the lifespan. Due to convenience and cost, sleep duration is often measured through self-report; yet, self-reported sleep duration can be highly biased. Actigraphy is a viable alternative as an objective measure of sleep. To analyze this actigraphy data, various sleep evaluation algorithms have been developed using regression methods, with coefficients constructed on minute-by-minute data measured at a specific device placement (wrist or hip). Because activity counts per minute may be affected by various factors in the study (e.g., type of device, sampling frequencies), regression-based algorithms developed within specific populations may not be generalizable to wider use. To address these concerns, we propose a new learning method to obtain robust and consistent sleep duration estimates. First, we identify temporal segments via pruned dynamic programming; then, we develop a calling algorithm with individual-specific thresholds and capture sleep periods. Our proposed method is motivated by and demonstrated in the Multi-Ethnic Study of Atherosclerosis (MESA) Sleep study and the Early Life Exposure in Mexico to ENvironmental Toxicants (ELEMENT) study.