Behavior
Unsupervised Machine Learning Identifies Hidden Movement Behavior Patterns Linked to Metabolic Syndrome in Multiracial/Ethnic College Students: The 24h-MESYN Study Marcus Vinicius Nascimento Ferreira* Marcus Vinicius Nascimento Ferreira Antonio Gibran de Almeida Cardoso Augusto César F. De Moraes Marcia Ferreira Sales Ethan T. Hunt Gabriela Berg Tiago Almeida de Oliveira Francisco Leonardo Torres-Leal Heráclito Barbosa de Carvalho Shirley Cunha Feuerstein
Aim: We used unsupervised machine learning to identify 24-hour movement behavior patterns in multiracial/ethnic college students and examined their associations with metabolic syndrome.
Methods: Our sample included 518 students (60.1% aged ≤20 years; 68.7% female; 72.0% non-White) from two Brazilian cities (Gini indices ≤0.56). We analyzed self-reported 24-hour movement behaviors, standardized as minutes/day, and assessed guideline adherence. Metabolic syndrome, the outcome, was evaluated in 375 students based on three of five risk factors: high abdominal circumference, blood pressure, triglycerides, fasting glucose, or low HDL. Clusters were identified using unsupervised methods (Calinski-Harabasz index and k-median) and described with compositional data analysis (CoDA). Associations between clusters and outcomes were examined using multilevel logistic regression, adjusted for biological sex, age, residence, work status, and degree program and shift.
Results: We found 56.6% met physical activity guidelines (60 min/day until 17, >30 min/day for ≥18), 43.8% met sedentary behavior guidelines (<2 h/day until 17, <8 h/day for ≥18), and 39.2% met sleep duration guidelines (8–10 h/night until 17, 7–9 h/night for ≥18), with a metabolic syndrome prevalence of 24.5%. Unsupervised modeling identified four behavior clusters, but only cluster #4 (22.2% of participants, with daily distribution of 4.7%±9.5 for physical activity, 60.0%±14.5 for sedentary behavior, and 35.3%±14.2 for sleep) was associated with the outcome in bivariate analysis (χ² = 3.97; p = 0.046). In adjusted models, no significant associations were found for meeting physical activity (OR: 1.61 [95% CI: 0.86–2.99]), sedentary behavior (OR: 1.40 [95% CI: 0.76–2.58]), or sleep duration (OR: 0.90 [95% CI: 0.49–1.65]) guidelines, while cluster #4 was significant (OR: 2.35 [95% CI: 1.07–5.15]).
Conclusions: Unsupervised machine learning offers a sensitivity alternative for assessing 24-hour movement behaviors, revealing hidden risk patterns associated with metabolic syndrome.