Position-specific lower extremity performance and machine learning-based injury-risk screening classification in adult male basketball players: a random forest and SHAP approach

Basketball imposes high position-dependent neuromuscular demands that predispose athletes to lower extremity injuries. Machine learning offers the potential to identify multivariable risk-screening patterns that single-threshold approaches miss. This study aimed to (1) compare position-specific lower extremity performance profiles across five playing positions and (2) develop and explain a Random Forest (RF) classifier of a composite, evidence-based injury-risk screening index in adult male basketball players, interpreted via SHapley Additive exPlanations (SHAP). No injuries were prospectively observed; the outcome is therefore a screening classification and not an observed injury endpoint. One hundred adult competitive male basketball players (age 24.5 ± 3.8 years, range 18–30; 20 per position) completed a single-session test battery. Standing height and body mass were measured under standardised conditions immediately before testing, and body mass index (BMI) was computed as body mass (kg) divided by height squared (m²). Performance was assessed with the DeepSport AI-based motion analysis application: countermovement jump height and peak power (CMJ), reactive strength index (RSI), 10-m sprint time, 505 agility time, bilateral single-leg hop distance, and limb asymmetry. Sample size was determined using G*Power 3.1.9.7 (Cohen’s f = 0.45, α = 0.05, 1 − β = 0.95) [19]. A multi-factor threshold model defined the binary screening label. Athletes meeting ≥ 2 of five predefined criteria (limb asymmetry > 15%, RSI < 1.8, 10-m sprint > 1.95 s, 505 agility > 5.0 s, BMI > 27 kg/m²) were classified as high-risk; these five variables also appear among the 13 candidate predictors. An RF classifier was tuned by grid search under stratified 5-fold cross-validation with all preprocessing fitted inside each training fold, and benchmarked against logistic regression, a support vector machine and gradient boosting. SHAP values were computed with TreeSHAP on the complete sample ( n = 100); out-of-fold permutation importance was used as the primary importance measure. A pre-specified sensitivity analysis re-fitted the model after removing all five label-defining variables. Sixty athletes (60%) were classified as high-risk and 40 (40%) as low-risk. BMI differed significantly across positions (F = 3.54, p = 0.010, η² = 0.130) as did RSI (F = 2.50, p = 0.048, η² = 0.095). The RF classifier achieved AUC = 0.947, accuracy = 84.0%, sensitivity = 93.3%, specificity = 70.0%, F1 = 87.5% and κ = 0.655 (mean across 10 cross-validation seeds: AUC = 0.930 ± 0.012). Logistic regression performed comparably (AUC = 0.910, accuracy = 87.0%, κ = 0.726), indicating that the additional complexity of the RF yielded only a marginal discrimination gain. When the five label-defining variables were removed, discrimination fell to AUC = 0.683 (accuracy 62.0%, κ = 0.167), quantifying the extent to which the headline performance reflects recovery of the labelling rule rather than independent signal. Permutation importance ranked limb asymmetry first (ΔAUC = 0.193 ± 0.043), followed by BMI (0.094 ± 0.057) and 10-m sprint time (0.092 ± 0.008). A Random Forest model incorporating field-based DeepSport performance metrics reproduces a composite lower-extremity risk-screening classification with outstanding internal discrimination, but its predictive validity for actual injury remains untested and it should be regarded as a screening-support tool rather than an injury-prediction model. Limb asymmetry, BMI and 10-m sprint time are the dominant screening factors, although this ordering is partly determined by the labelling rule, and BMI should be read as a proxy for body size and joint loading rather than as a modifiable index of adiposity. SHAP explanations provide practitioner-readable rationales for individual risk scores, supporting integration into athlete monitoring workflows. ClinicalTrials.gov, NCT07677371; registered 18 June 2026, i.e. retrospectively, during the May–June 2026 data-collection period.