A novel statistical approach for wearable data analysis in the context of physical activity

Abstract

Physical activity plays a crucial role in the treatment, management, and prevention of various health conditions. In recent years, wearable devices have emerged as a more objective and accurate technology for monitoring physical activity compared to traditional self-reported methods. These devices generate temporal data due to the sequential way it is collected, where each data point is recorded at a specific time and is not independent from other records but, rather, a correlation exists between adjacent points. These characteristics present major challenges in the statistical analysis of this type of data as traditional aggregation methods assume the observed datapoints to be independent, ignoring correlation structures, and failing to identify temporal patterns in the data (e.g. cycles/seasonality). Thus, new approaches are required to analyse longitudinal physical activity data.

The development of new tools to analyse werable device data could be valuable in medical research for the assessment of interventions, and to examine the interaction between longer term activity trends and the pathogenesis or progression of chronic diseases. Such tools could also be useful in clinical practice, allowing clinicians to monitor patients’ activity levels over time, and to give them tailored advice as part of primary and/or secondary prevention strategies. Additionally, patients themselves could gain an improved understanding of their physical activity over time, promoting adherence to interventions and programmes.

Our aim is to develop novel statistical methods for analysing high-resolution longitudinal physical activity data. Our approach uses Non-Stationary Time Series Analysis, a set of statistical models designed to account for the temporal dependency in continuous ordered data. In particular, we propose using the Trend Locally Stationary Wavelet model and introduce time-dependent metrics to assess and characterise physical activity performance over time, moving beyond traditional pre-post mean.

Date of Award	25 Jun 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Dylan Thompson (Supervisor) & Matthew Nunes (Supervisor)