Machine learning application in predicting anterior cruciate ligament injury among basketball players

• Main Author: Longfei, Guo
• Format: Thesis
• Language: English
• Published: 2025
• Subjects: R Medicine; RA440-440.87 Study and teaching. Research
• Online Access: http://eprints.usm.my/63252/

Anterior cruciate ligament (ACL) injury is among the most prevalent injuries in athletes, significantly impacting their competitive performance. Preventing ACL injury is challenging due to their multifactorial nature. Machine learning-based data mining techniques have shown significant potential in identifying risk factors associated with ACL injury. This study aimed to assess the predictive capability of these features using machine learning models. Data on athlete’s profile, physical function, specialized qualities, three-dimensional movement analysis, and simultaneous electromyography were prospectively collected from 104 male basketball players. A one-year follow-up was conducted to monitor ACL injury, identifying n=11 injured players. Four machine learning algorithms—Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), and Logistic Regression (LR)—were developed to predict ACL injury. The optimal model was selected based on the mean area under the receiver operating characteristic curve (AUC-ROC) across 10 cross-validation runs and was used with Shapley Additive exPlanations to analyze the risk factors. The results show that AUC-ROC values varied slightly across repetitions and methods (0.66-0.80), the best classifier was RF. SHAP analysis identified key feature with the highest predictive value for ACL injury during specific sports motions. Emergency Stop phase: Increased knee flexion moment, posterior ground reaction forces, knee flexion angle, and overactivation of the lateral quadriceps and rectus femoris. Initial Acceleration phase: Elevated knee internal rotation torque and lateral stress on the lower limbs. Side-Cutting phase: Reduced tibial inclination and hip flexion angles, increased ankle inversion angle, ankle eversion moment, and excessive lateral thigh muscle activation. Furthermore, poor stability in the non-dominant leg, weak Squat Jump performance, training loads exceeding 15 hours per week, and prior injury history were significant ACL injury predictors. This study emphasizes the Machine Learning model's effectiveness in predicting ACL injury, highlighting biomechanical metrics, functional attributes, and historical feature as critical predictors for targeted prevention strategies.