Machine learning-based diagnostic decision support for suspected anterior cruciate ligament tears in clinically evaluated athletes

Anterior cruciate ligament tears (ACLT) may be difficult to identify at the initial clinical encounter because acute symptoms can limit examination and access to diagnostic tools may be constrained. This study evaluated supervised machine-learning models for ACLT decision support using routinely available initial clinical data. Data from 196 athletes with suspected ACLT (146 confirmed ACLT; 50 symptomatic non-ACLT) were represented by seven manually annotated binary clinical predictors. Participants were stratified into training (n = 137) and independent held-out evaluation (n = 59) sets. Eight classifiers were optimized using nested cross-validation within the training set and evaluated using accuracy, precision, recall, F1-score, and receiver operating characteristic area under the curve (ROC-AUC). Classification used a training-derived F1-optimal threshold of 0.50. Optuna-tuned XGBoost achieved the highest ROC-AUC (0.962) and F1-score (94.25%); XGBoost, support vector machine, and k-nearest neighbors achieved the highest recall (93.18%). SHAP analysis of XGBoost identified joint effusion as the strongest contributor to model output. These findings indicate that supervised machine-learning models based exclusively on initial clinical variables may support ACLT assessment at first presentation. Prospective external validation is required before clinical implementation.