Journal of Exercise Science and Fitness · Nov 2026
Explainable Machine Learning Analyses Uncover Nonlinear Associations Between 24-Hour Movement Behavior Composition and Cardiometabolic Risk Among Korean Adults
By Seol J, Cho J, Watso JC
Background/Objective To examine associations between 24-hour movement behavior composition and cardiometabolic risk and explore nonlinear associations and favorable compositions using explainable machine learning. Methods This cross-sectional study included 41,288 adults from the Korea National Health and Nutrition Examination Survey (2016–2024). Sleep, sedentary behavior, moderate-to-vigorous physical activity (MVPA), and other activities formed a 24-hour composition. Cardiometabolic risk was assessed using a cardiometabolic risk score (CmRS). Survey-weighted compositional data analysis examined associations and time reallocation; extreme gradient boosting with SHapley Additive exPlanations explored nonlinear associations. Grid-search identified favorable model-derived compositions. Results The 24-hour composition was associated with CmRS (p < 0.001). Greater relative sleep (β=−0.073, 95% CI −0.095 to −0.051) and MVPA (β=−0.017, 95% CI −0.020 to −0.013) were associated with lower CmRS, whereas greater relative sedentary behavior was associated with higher CmRS (β=0.086, 95% CI 0.070 to 0.102). Reallocating 30 min/day toward MVPA showed the largest reduction in CmRS (−0.0221, 95% CI −0.0264 to −0.0178). The machine-learning model achieved an R² of 0.321, with MVPA leading among movement behaviors. Unrestricted grid-search identified model-derived favorable compositions characterized by approximately 450 min/day of sleep, 480 min/day of sedentary behavior, and 80 min/day of MVPA. Constrained analyses favored greater MVPA within the specified search range, although these model-derived values should not be interpreted as population-wide optimal targets. Conclusion Twenty-four-hour movement composition was associated with cardiometabolic risk, with greater relative sleep and MVPA and lower sedentary behavior associated with favorable profiles. Explainable machine learning identified nonlinear associations and favorable model-derived compositions requiring longitudinal evaluation.