Interpretable multimodal machine learning for injury risk stratification in adolescent athletes: integrating kinematic, physiological, and load data with SHAP/LIME explainability.

Adolescent athletes (aged 13-18 years) face significantly elevated lower-limb injury risk compared to adult populations due to immature neuromuscular control, while existing injury risk assessment methods are broadly limited by single-modal reliance, insufficient model interpretability, and a lack of dedicated research targeting youth populations. This study proposes an interpretable multimodal machine learning framework that integrates three data modalities-kinematics, physiology, and training load-via feature-level concatenation, based on a cross-sectional multimodal dataset of 320 fully anonymised adolescent athletes comprising 12 feature variables acquired under standardized protocols using internationally recognized instruments, including Vicon optical motion capture, AMTI force platforms, Polar H10 heart rate monitors, Supersonic Imagine Aixplorer ultrasound elastography, and Catapult GPS trackers. Through 5-fold stratified cross-validation, the comprehensive performance of five traditional and five modern/deep machine learning models was systematically compared in a binary injury risk stratification task. Experimental results demonstrate that logistic regression achieved the best overall performance [Accuracy = 86.9%, AUC-ROC = 0.965, F1 = 0.866, Matthews Correlation Coefficient (MCC) = 0.742]. Dual interpretability analysis was further conducted on the optimal model using a SHapley Additive exPlanations (SHAP) linear explainer and a Local Interpretable Model-Agnostic Explanations (LIME) local explainer; both methods consistently identified peak hip adduction angle (mean |SHAP| = 2.208), ankle dorsiflexion range of motion (1.426), and dynamic knee valgus angle (1.359) as the most critical injury risk predictors (Spearman ρ = 0.93), in high agreement with established biomechanical findings in sports medicine. An ablation analysis of modality contributions demonstrated that the tri-modal integration achieved a substantive AUC gain over single-modality configurations, supporting the value of multimodal feature integration. We note explicitly that the binary risk label was derived from a clinically-informed weighted composite of the input features rather than from prospective injury follow-up; the framework therefore performs risk stratification with respect to this composite, and validation against true prospective injury outcomes remains an important next step. This study provides an interpretable technical framework for standardized, digitized, and stratified injury risk assessment in adolescent athletes, with important potential for clinical translation and sports injury prevention practice.