Keywords: accelerometry, physical activity, prediction, semi-parametric, wearable
Wearable devices, such as accelerometers and heart rate monitors, can now provide objective and continuous measurements of human activity. Such devices have been widely deployed in large observational and clinical studies because they are expected to produce objective measurements that could improve or replace current self-reported activity measuring practices. Accelerometry data were usually obtained in a very high sampling frequency (micro-level), and could be subsequently reduced to count data (macro-level) with one measurement per minute for easier usage. Different statistical methods are needed to analyze the accelerometry signals of different level. We first introduce a movement recognition method based on “movelets”, to predict the type of physical activity at the sub-second level using the micro-level accelerometry data. Then, we discuss a two-stage model for the macro-level data to describe the inactive/active and activity intensity dynamics of the circadian rhythm of physical activity.