Statistical Learning Methods to Identify Nonwear Periods From Accelerometer Data

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