Effect of a sports-medicine–guided football program on physical fitness in adolescents: A controlled school-based trial

Figure 1.. CONSORT participant flow diagram. Participant flow diagram illustrating the number of students screened, classified, allocated to intervention or control groups, completing the intervention, and included in final analyses. Reasons for exclusions and dropouts are indicated at each stage.

Figure 2.. Architecture of the ensemble-based injury prediction system. Schematic overview of the ensemble-based injury prediction system workflow. The pipeline includes input of the football injury dataset, data preprocessing (handling missing values, encoding, normalization), stratified train-test split, model training using a Voting Regressor (CatBoost and Gradient Boosting), evaluation metrics computation, SHAP-based explainability, and an agentic AI feedback module providing individualized injury risk recommendations. Arrows indicate the sequence and data flow between modules.

Figure 3.. Bar plots comparing R2, MAE, and RMSE across regression models. Bar plots presenting the comparative performance of eleven regression models for predicting football player injury duration. Metrics shown include R2 score, mean absolute error (MAE), and root mean squared error (RMSE). The Voting Regressor (CatBoost + Gradient Boosting) demonstrated the highest R2 and lowest error rates, indicating superior predictive performance.

Figure 4.. Scatter plots of actual versus predicted days out. Scatter plots comparing actual versus predicted injury recovery durations (days out) for selected regression models. The red dashed diagonal line indicates the ideal line of perfect prediction. Points clustered closer to the line reflect more accurate model performance.

Figure 5.. Residual distribution plots for all models. Residual distribution plots for all evaluated regression models, depicting the spread and symmetry of prediction errors (predicted minus actual days out). Narrower and more symmetric distributions around zero indicate improved model calibration and predictive consistency.

Figure 6.. Residual distribution plots for all models, illustrating prediction error spread; the Voting Regressor shows the most centralized and symmetric residuals.

Figure 7.. Learning curves depicting training and validation R2 scores across increasing dataset sizes for each model.

Figure 8.. SHAP Beeswarm and Feature Importance Plots. SHAP (SHapley Additive exPlanations) beeswarm plot (left) and feature importance bar plot (right) for the final Voting Regressor model. The beeswarm plot visualizes individual feature effects on model output, while the bar plot summarizes mean absolute SHAP values, highlighting the most influential predictors of injury duration.

Figure 9.. Sample outputs of agentic AI feedback. Examples of individualized feedback generated by the agentic AI module. Each output includes a predicted injury duration, key contributing risk factors, and personalized recommendations for training modification, rehabilitation, or monitoring.