Statistical modeling of physical activity based on accelerometer data

Abstract

This thesis focuses on the objective measurement of physical activity (PA), recorded by accelerometers. Chapter 2 describes the objective measurement of PA using accelerometers in contrast to subjective measurements like PA questionnaires. Chapter 3 presents the basic assumption on PA. Contrary to the cutpoint method, it is more realistic to assume that human activity behavior consists of a sequence of non-overlapping, distinguishable activities that can be represented by a mean intensity level. The recorded accelerometer counts scatter around this mean level. In Chapter 4, two novel approaches to better capture PA are developed and implemented. The Hidden Markov models are stochastic models that allow fitting a Markov chain with a predefined number of activities to the data. Expectile regression utilizing the Whittaker smoother with an L0-penalty is introduced as a second innovative approach. Expectile regression is compared to HMMs and the cutpoint method in a simulation study. Chapter 5 presents the results of four studies on PA. Chapter 6 summarizes and discusses the findings of the previous chapters and ends with an outlook on future research.