Abstract
Physical inactivity is a critical public health challenge, contributing to rising rates of chronic diseases. Despite widespread promotion, generic physical activity (PA) guidelines have had limited success in sustaining longterm behavior change. Adapting PA interventions to individuals’ daily routines offers a promising strategy to promote lasting PA habits. Developing these intervention approaches requires a deep understanding of daily PA routines—sustained patterns characterized by regularity (consistency of behavior), stability (endurance of activity patterns), and low variability (degree of fluctuation). Current methods do not fully leverage the potential of wearable sensor data due to reliance on simple, aggregated measurements and unvalidated, unsupervised models.
This project addresses these gaps by developing novel time-series and machine learning methods to advance routine PA pattern assessment. Leveraging data from the TIME study (U01HL146327), which includes continues accelerometer and daily ecological momentary assessment (EMA) data collected over 12 months, the project aims to: (1) Develop advanced analytical methods and metrics (e.g., matrix profile, entropy) to comprehensively characterize routine PA patterns from accelerometer data, capturing key dimensions such as regularity, intensity, and stability; (2) Build adaptive and interpretable machine learning models (e.g., random forest, CNNs, LSTMs) to classify routine and non-routine PA days, integrating subjective labels from EMA to improve model precision and real-world applicability; and (3) Create an open-source R package to provide accessible tools for public health researchers and practitioners to assess PA patterns across diverse populations and datasets.
The tools and algorithms developed will be further tested in an R01 application, extending their use to public longitudinal motion sensor datasets (e.g., NIH’s All of Us cohort) to predict cardiometabolic health outcomes (HbA1c, blood pressure, cholesterol). The tools for routine detection will also have broader translational applications, including the development of next-generation adaptive interventions for sleep, nutrition, and tobacco use.