Abstract
PURPOSE: To develop a machine learning model capable of predicting the bony risk of non-contact anterior cruciate ligament injury, thereby enabling the identification of factors that contribute to such injuries. METHODS: Data were collected from 400 cases of non-contact ACL-injured and 200 ACL-intact control subjects using Computed Tomography between March 2022 and June 2025. Thirteen features, encompassing demographic, clinical, and radiomic data, as well as six different algorithms, were utilised to develop predictive machine learning models. Shapley Additive Explanations (SHAP) analysis was subsequently performed on the optimal model. RESULTS: The Maximum 2D Diameter Column values for the non-contact ACL injury group and the intact ACL group were 31.12 ± 2.92 mm (95% confidence interval [CI]: 30.83-31.40, p < 0.05) and 32.37 ± 3.07 mm (95% CI: 31.94-32.80, p < 0.05). The Extreme Gradient Boosting classifier was identified as the optimal predictive model, achieving an area under the precision-recall curve of 0.94, the highest among all models evaluated. SHAP analysis revealed that the most predictive feature was the Maximum 2D Diameter Column of the notch, defined as the largest pairwise Euclidean distance between tumour surface mesh vertices in the row-slice plane, followed by the lateral and medial posterior tibial slope. CONCLUSION: The machine learning model developed in this study demonstrated excellent predictive performance for non-contact ACL injuries. The Maximum 2D Diameter Column was the most important predictor, followed by the lateral and medial posterior tibial slope. LEVEL OF EVIDENCE: Level III.