Machine Learning Models Reliably Predict Clinical Outcomes in Medial Patellofemoral Ligament Reconstruction

Abstract

Listen

Translate article

PurposeTo develop the machine learning model to predict clinical outcomes following MPFLR and identify the important predictive indicators. MethodsThis study included patients who underwent MPFLR from January 2018 to December 2022. The exclusion criteria were as follows: 1) concurrent bony procedures, 2) history of other knee surgeries, and 3) follow-up period of less than 12 months. Forty-two predictive models were constructed for seven clinical outcomes (failure to achieve MCID of clinical scores, return to pre-injury sports, pivoting sports, and recurrent instability) using six machine learning algorithms (Random Forest, Logistic Regression, Support Vector Machine, Decision Tree, implemented multilayer perceptron, and K-nearest neighbor). The performance of the model was evaluated using metrics such as the area under the receiver operating characteristic curve (AUC), accuracy, specificity, and sensitivity. Additionally, Shapley Additive Explanation summary plot was employed to identify the important predictive factors of the best-performing model. ResultsA total of 218 patients met criteria. For the best-performing models in predicting failure to achieve the MCID for Lysholm, IKDC, Kujala, and Tegner scores, the AUCs and accuracies were 0.884 (good) and 87.3%, 0.859 (good) and 86.2%, 0.969 (excellent) and 97.0%, and 0.760 (fair) and 76.8%, respectively; 0.952 (excellent) and 95.2% for return to pre-injury sports; 0.756 (fair) and 75.4% for return to pivoting sports; and 0.943 (excellent) and 94.9% for recurrent instability. Low preoperative Tegner score, shorter time to surgery, and absence of severe trochlear dysplasia were significant predictors for return to pre-injury sports, while absence of severe trochlear dysplasia and patellar alta were significant predictors for return to pivoting sports. Older age, female sex, and low preoperative Lysholm score were highly predictive of recurrent instability. ConclusionThe predictive models developed using machine learning algorithms can reliably forecast the clinical outcomes of MPFLR, particularly demonstrating excellent performance in predicting recurrent instability. Level of evidenceLevel III, case-control study.