Abstract:
|
The measurement of physical activity in individuals is important for understanding obesity causes, prevention and control. Wrist and hip-worn accelerometers have become standard tools for monitoring movement over extended time periods, offering the possibility to determine long-term patterns of activity that are related to body fatness and fitness. Meaningful analysis of the real world data depends on the identification of intervals when the device is not worn. Lack of true knowledge of non-wear status together with the similarity among non-wear, sedentary behavior, and sleep makes this task difficult. We introduce an ensemble for non-wear detection that leverages, in an unsupervised way, the predictions of existing wear/non-wear/sleep classification methods. Since field data are inevitably different from the data on which the predictors were trained or evaluated, their true accuracy is unknown. We derive accuracy estimators, which we obtain based on experts' consistency of predictions by solving a linear system. We then use the accuracy estimates in a weighting scheme for the ensemble classification, in order to obtain superior predictions compared to the original classifiers.
|