Predicting the Objective and Subjective Clinical Outcomes of Anterior Cruciate Ligament Reconstruction: A Machine Learning Analysis of 432 Patients

Background: Sports levels, baseline patient-reported outcome measures (PROMs), and surgical procedures are correlated with the outcomes of anterior cruciate ligament reconstruction (ACLR). Machine learning may be superior to conventional statistical methods in making repeatable and accurate predictions. Purpose: To identify the best-performing machine learning models for predicting the objective and subjective clinical outcomes of ACLR and to determine the most important predictors. Study Design: Case-control study; Level of evidence, 3. Methods: A total of 432 patients who underwent anatomic double-bundle ACLR with hamstring tendon autograft between January 2010 and February 2019 were included in the machine learning analysis. A total of 15 predictive variables and 6 outcome variables were selected to validate the logistic regression, Gaussian naïve Bayes machine, random forest, Extreme Gradient Boosting (XGBoost), isotonically calibrated XGBoost, and sigmoid calibrated XGBoost models. For each clinical outcome, the best-performing model was determined using the area under the receiver operating characteristic curve (AUC), whereas the importance and direction of each predictive variable were demonstrated in a Shapley Additive Explanations summary plot. Results: The AUC and accuracy of the best-performing model, respectively, were 0.944 (excellent) and 98.6% for graft failure; 0.920 (excellent) and 91.4% for residual laxity; 0.930 (excellent) and 91.0% for failure to achieve the minimal clinically important difference (MCID) of the Lysholm score; 0.942 (excellent) and 95.1% for failure to achieve the MCID of the International Knee Documentation Committee (IKDC) score; 0.773 (fair) and 70.5% for return to preinjury sports; and 0.777 (fair) and 69.2% for return to pivoting sports. Medial meniscal resection, participation in competitive sports, and steep posterior tibial slope were top predictors of graft failure, whereas high-grade preoperative knee laxity, long follow-up period, and participation in competitive sports were top predictors of residual laxity. High preoperative Lysholm and IKDC scores were highly predictive of not achieving the MCIDs of PROMs. Young age, male sex, high preoperative IKDC score, and large graft diameter were important predictors of return to preinjury or pivoting sports. Conclusion: Machine learning analysis can provide reliable predictions for the objective and subjective clinical outcomes (graft failure, residual laxity, PROMs, and return to sports) of ACLR. Patient-specific evaluation and decision making are recommended before and after surgery.