Predicting physical inactivity risk in middle-aged and older adults: a machine learning approach using longitudinal data

<p dir="ltr">Physical inactivity risk (PIR) in middle-aged and older adults arises from a complex mix of individual, interpersonal, household, and societal factors, yet integrated analyses remain limited. This study employs a two-step machine learning approach grounded in a social-ecological framework. Least absolute shrinkage and selection operator (Lasso) reduces 64 candidate predictors across individual, interpersonal, and family domains to 38 key variables, minimizing bias from prior assumptions. Subsequently, a rolling Extreme Gradient Boosting (XGBoost) classifier identifies the top 20 factors most strongly associated with PIR, with SHapley Additive exPlanations (SHAP) used to interpret predictor contributions. Findings reveal PIR is influenced by various demographic, socioeconomic, health, behavioral factors across individual, interpersonal, and family levels; nonetheless, the decline in social interaction associated with reduced working hours stands out as a significant contributor. This study demonstrates that recent advances in ML can uncover complex, non-linear predictors of PIR that conventional variable selection methods may overlook. The integration of Lasso and rolling XGBoost provides a robust data-driven framework for identifying key risk factors, offering valuable insights for targeted interventions in aging and urbanizing populations</p>