Mining Activity Tracker Data To Analyse Physical Activity Behaviours And Provide Personalised Feedback In Health Education Programmes

Abstract

The use of activity trackers in health education offers great potential to objectively assess healthy behaviour learning. However, the current models and techniques used to exploit and analyse their data mainly compare highly aggregated amounts of physical activity, thus diluting the fine physical activity patterns contained in the granular tracker data. The extraction of these latent patterns is valuable for health education because they may help to better understand how interventions affect behaviours and how they change over time. Physical activity patterns offer a more detailed view of how participants learn and offer the possibility to encourage healthy behaviours when detected on time. This thesis proposes data mining models and techniques to extract these finer physical activity patterns that can then be used to analyse physical activity behaviours. This allows performing elaborate intervention assessments and generating timely personalised feedback during interventions. In Chapter 1, we introduce the research questions and objectives. In Chapter 2, we present a systematic literature review of the current state of data mining techniques that use physical activity sensor data in health education to detect behaviour changes. We discuss common challenges and opportunities to guide future work. In Chapter 3, we propose a data mining method that highlights the nature and timing of behaviour changes for a more insightful assessment of health interventions. In Chapter 4, we describe U-BEHAVED, an unsupervised machine learning technique to detect significant physical activity behaviour changes and to determine whether they become habitual as the health intervention unfolds. In Chapter 5, we model physical activity behaviour changes to provide personalised feedback. Finally, we discuss our work and describe how it answered the research questions.