Explainable machine learning predicts adolescent cognitive development from sport engagement in a national longitudinal survey

Globally, over 80% of adolescents fail to meet recommended physical activity levels, raising concerns about cognitive development during a critical neurodevelopmental period. Physical activity is associated with cognitive benefits in young people, yet large-scale predictive evidence on how extracurricular sport engagement and PE-related school resources relate to adolescent cognitive development remains limited. This study applied explainable machine learning to the China Education Panel Survey, a nationally representative longitudinal dataset, to predict follow-up cognitive ability from baseline extracurricular sport engagement, PE-related school resources, and contextual variables among 9366 seventh-grade students. Five models were compared: Random Forest, XGBoost, LightGBM, Support Vector Regression, and a Stacking ensemble. The Stacking model achieved the best performance (R² = 0.593, RMSE = 0.523, Pearson r = 0.770). SHAP analysis revealed that baseline cognitive ability and academic achievement were the strongest predictors, while sport-engagement and PE-related school-resource variables collectively accounted for 5.5% of total SHAP contribution, with individual behavioral engagement (3.7%) exceeding PE-related school resources (1.8%). Among PE predictors, running-track length, weekend sport duration, and frequency of exercising with parents ranked highest. SHAP dependence analysis indicated a negative association between excessive weekday sport time and cognitive outcomes, alongside a positive moderating role of family economic status on weekend sport engagement. Propensity score matching identified a significant association between sport interest and cognitive development (ATT = 0.071, p = 0.004), though sensitivity analysis indicated limited robustness to unobserved confounding. These findings suggest that extracurricular sport engagement, particularly intrinsic sport interest, shows a modest but meaningful association with adolescent cognitive development over one year.