Statistical Learning Methods to Identify Nonwear Periods From Accelerometer Data

Abstract

Background: Accelerometers are used to objectively measure movement in free-living individuals. Distinguishing nonwear from sleep and sedentary behavior is important to derive accurate measures of physical activity, sedentary behavior, and sleep. We applied statistical learning approaches to examine their promise in detecting nonwear time and compared the results with commonly used wear time (WT) algorithms. Methods: Fifteen children, aged 4–17, wore an ActiGraph wGT3X-BT monitor on their hip during overnight polysomnography. We applied Hidden Markov Models (HMM) and Gaussian Mixture Models (GMM) to classify states of nonwear and wear in triaxial acceleration data. Performance of methods was compared with WT algorithms across two conditions with differing amounts of consecutive nonwear. Clinical scoring of polysomnography served as the gold standard. Results: When the length of nonwear was less than or equal to WT algorithms’ predefined thresholds for consecutive nonwear time, GMM methods yielded improved classification error, specificity, positive predictive value, and negative predictive value over commonly used algorithms. HMM was superior to one algorithm for sensitivity and negative predictive value. When the length of nonwear was longer, results were mixed, with the commonly used algorithms performing better on some parameters but GMM with the greatest specificity. However, all approached the upper limits of performance for almost all metrics. Conclusions: GMM and HMM demonstrated robust, consistently strong performance across multiple conditions, surpassing or remaining competitive with commonly used WT algorithms which had marked inaccuracy when nonwear time periods were shorter. Of the two statistical learning algorithms, GMM was superior to HMM.