Characterizing daily physical activity patterns with unsupervised learning via functional mixture models.

Area of Science:

• Public Health
• Behavioral Science
• Biostatistics

Background:

• Physical inactivity is a major public health issue.
• Understanding variations in physical activity (PA) trends is crucial for designing effective interventions.
• Identifying distinct daily PA patterns can offer insights into population behaviors.

Purpose of the Study:

• To identify latent profiles, or "phenotypes," of daily physical activity trends in adults.
• To analyze accelerometry data to uncover distinct patterns of physical activity.
• To inform the design of targeted public health interventions.

Main Methods:

• Secondary analysis of 724 person-day accelerometry data from 133 urban adults.
• Utilized Actigraph accelerometers and Actilife software for data collection and processing.
• Employed a probabilistic clustering technique (functional mixture models) to identify PA phenotypes based on step counts, moderate-intensity PA (MOD), total activity, and sedentary minutes.

Main Results:

• A 4-cluster solution provided the best model fit, with moderate-intensity PA (MOD) showing the most significant between-cluster differences.
• Phenotype 1 (N=61) exhibited a morning PA peak.
• Phenotype 2 (N=18) showed later bedtimes and highest daily PA volume with a morning peak.
• Phenotype 3 (N=29) had the lowest overall PA levels.
• Phenotype 4 displayed evenly distributed PA throughout the day with later waking/bedtimes.

Conclusions:

• Distinct and interpretable physical activity phenotypes based on temporal patterns were identified.
• Functional clustering of PA data offers actionable insights for tailoring behavioral interventions.
• These findings can guide personalized approaches to address physical inactivity.