Comparison of two-stage FPCA and bayesian mixture models applied to physical activity accelerometer data : pattern identification and group classification, A

Abstract

Collectively, society has begun to borderline obsess with ability to monitor our physical activity. A plethora of devices now exist to monitor steps, heart rate, sleep pattern, and many other metrics, giving individuals more and more information about their overall health. Research into physical activity and physical activity patterns has drastically increased as accelerometers’ ability to accurately capture information has grown. Often, physical activity data, collected as finely as every 0.25 seconds, is reduced into daily averages for individuals. This results in a huge loss of information and prevents answering any complicated questions about physical activity patterns during the course of a day. Two statistical methods to identify physical activity patterns and cluster individuals with similar patterns were compared. We used a two-stage functional principle components analysis (fPCA) model, which used fPCA results in a Gaussian mixture model, as well as a Bayesian mixture model that incorporates the estimation and clustering into a single model estimated with Markov chain Monte Carlo. We demonstrate these two methods on multiple simulation studies, to evaluate their ability to accurately classify like individuals and identify their activity patterns when true patterns and clusters are known. We then apply these methods to data from an Energy Expenditure study performed at CU Anschutz, to see how they perform with a real data set.