Nighttime Sleep Duration Prediction for Inpatient Rehabilitation using Similar Actigraphy Sequences

Abstract

Actigraphs are wearable sensors used to collect activity and sleep time series data from healthy and unhealthy populations. Unhealthy populations, such as individuals undergoing inpatient rehabilitation, typically exhibit abnormal daytime physical activity and nighttime sleeping patterns due to their injury and drastic changes in their activities of daily living. Consequently, Actigraph data collected from patients attending inpatient rehabilitation are often noisy and can be difficult to reliably draw conclusions from. In this paper, we apply machine learning to analyze such highly variable Actigraph data. We collected 24-hour, minute-by-minute Actigraph data from 17 patients receiving inpatient therapy post-stroke or post-traumatic brain injury. Our approach utilizes similarities among historical sequences of data to train machine learning algorithms to predict nighttime sleep duration. By tuning parameters related to our regression algorithm, we obtained a normalized root mean square error of 14.40%. Our approach is suitable for point of care and remote monitoring to detect changes in sleep for individuals recovering from stroke and traumatic brain injuries.