127 – Statistical Applications in Epidemiology
Change Point Problems for the Detection of Activity Bouts in Accelerometer Data
Jung Ae Lee
University of Arkansas
Andrew Bartlett
Southern Illinois University Edwardsville
An accelerometer, a wearable motion sensor on hip or wrist, is a promising research tool to perform an objective experiment of human's behavior(s). An important aspect of analyzing the output of this experiment is how to classify the physical activity types regarding intensity and duration of activities, e.g., how long the moderate-to-vigorous physical activity (MVPA) last. Our approach to this is to use the change point analysis in activity series. The accelerometer output of a day is a sequence of activity counts measured at successive points in time (1 minute or 30 sec epoch). Our method justifies a change point in time at which unknown quantities of the distribution abruptly change, such as mean shift. A few well-known change point methods are applied for the detection of the exercise duration, like brisk walking or running, and compared with a threshold-based approach that detects the bouts by consecutive minutes over a threshold with the allowance of two-minutes deviation. We found that the length and the location of bouts are different across methods. This paper provides discussion between a change point approach and a threshold-based approach in terms of mathematical justification and practical implications. The National Health and Nutrition Examination Survey (NHANES) data is used to demonstrate the methods.