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

Physical inactivity is a significant public health concern. Consideration of inter-individual variations in physical activity (PA) trends can provide additional information about the groups under study to aid intervention design. This study aims to identify latent profiles (“phenotypes”) based on daily PA trends among adults living in. This was a secondary analysis of 724 person-level days of accelerometry data from 133 urban-dwelling adults (89% Latinx, age = 19–77 years). We used Actigraph accelerometers and the Actilife software to collect and process 24-hour PA data. We implemented a probabilistic clustering technique based on functional mixture models. Multiple days of data per person were averaged for entry into the models. We evaluated step counts, moderate-intensity PA (MOD), total activity and sedentary minutes as potential model variables. Bayesian Information Criterion (BIC) index was used to select the model that provided the best fit for the data. A 4-cluster resolution provided the best fit for the data (i.e., BIC=-3257, improvements of Δ = 13 and Δ = 7 from 3- and 5-cluster models, respectively). MOD provided the greatest between-cluster discrimination. Phenotype 1 (N = 61) was characterized by a morning peak in PA that declined until bedtime. Later bedtimes and the highest daily PA volume were distinct for phenotype 2 (N = 18), along with a similar peak pattern. Phenotype 3 (N = 29) membership was associated with the lowest PA levels throughout the day. Phenotype 4 was characterized by a more evenly distributed PA during the day, and later waking/bedtimes. Our findings point to distinct, interpretable PA phenotypes based on temporal patterns. Functional clustering of PA data could provide additional actionable points for tailoring behavioral interventions.