Predict the Anterior Cruciate Ligament Injuries Among Collegiate Athletes

Abstract

Anterior cruciate ligament (ACL) injuries are prevalent among collegiate athletes and often result in long-term impairment of athletic performance. Machine learning offers technical possibilities for the early prediction of this injury. Addressing the challenges of the Kaggle public dataset “Athlete Injury and Performance” namely its small scale (200 cases total), extreme class imbalance (only 13 injury samples), and the disconnect between high accuracy scores in prior studies and actual patient identification capabilities, this research aims to systematically evaluate the real-world predictive performance of machine learning models. Using “Injury_Indicator” as the target label, irrelevant feature (Athlete_ID) were removed. String features were encoded and standardized. Using leave-one-out cross-validation (LOOCV) combined with oversampling to balance the training set, 12 binary classification machine learning models were constructed. Performance was evaluated using nine metrics. Results indicate that the XGBoost model demonstrated optimal performance, achieving an Accuracy of 0.96, Sensitivity of 0.7857, Kappa coefficient of 0.7118, and AUC of 0.950, with the lowest number of false negatives (fn=3). Although models such as logistic regression and decision trees exhibited high Accuracy, they showed variations in patient identification capabilities. This study demonstrates that machine learning holds potential for predicting ACL injuries in collegiate athletes. However, comprehensive evaluation using multidimensional metrics is essential rather than relying solely on accuracy. The findings provide methodological insights for injury risk prediction in sports medicine.

Ähnliche Arbeiten