Explainable machine learning classification of anterior cruciate ligament related biomechanical risk phenotypes using core stability measures in physical education students

Abstract Anterior cruciate ligament (ACL) injuries remain a major challenge in sports medicine. Machine learning studies of ACL-related risk rarely combine directly measured core function with interpretable modelling in university student-athletes. This cross-sectional study recruited 135 physical education students and constructed a 10-feature assessment combining static stability, dynamic stability, and muscular strength-endurance. Participants were classified into a composite biomechanical high-risk phenotype defined by landing kinematics and strength ratios. No participant was followed for incident ACL injury, so all models classify this surrogate phenotype rather than predict injury. Seven machine learning algorithms were compared using 10 × 10 repeated stratified cross-validation. XGBoost performed best (AUC-ROC = 0.847, sensitivity = 78.6%, specificity = 82.8%), outperforming logistic regression (ΔAUC = 0.085, P < 0.001). SHapley Additive exPlanations (SHAP) analysis identified Y-Balance Test composite score (mean |SHAP| = 0.312), side bridge duration (0.268), and Functional Movement Screen total score (0.241) as the largest contributors to model output, with this ranking stable across folds (Spearman ρ = 0.89); functional dynamic measures contributed more than isolated muscle strength parameters. Given the surrogate outcome, single-center design, and absence of external validation, this framework is exploratory rather than a validated screening tool. Multi-center prospective validation against observed ACL injury is needed.