Developing a Machine-Learning Predictive Model for Retention of Posterior Cruciate Ligament in Patients Undergoing Total Knee Arthroplasty

Abstract

Objective: Predicting whether the posterior cruciate ligament (PCL) should be preserved during total knee arthroplasty (TKA) procedures is a complex task in the preoperative phase. The choice to either retain or excise the PCL has a substantial effect on the surgical outcomes and biomechanical integrity of the knee joint after the operation. To enhance surgeons' ability to predict the removal and retention of the PCL in patients before TKA, we developed machine learning models. We also identified significant feature factors that contribute to accurate predictions during this process.

Methods: Patients' data on TKA continuously performed by a single surgeon who had intended initially to undergo implantation of cruciate-retaining (CR) prostheses was collected. During the sacrifice of PCL, we utilized anterior-stabilized (AS) tibial bearings. The dataset was split into CR and AS categories to form distinct groups. Relevant information regarding age, gender, body mass index (BMI), the affected side, and preoperative diagnosis was extracted by reviewing the medical records of the patients. To ensure the authenticity of the research, an initial step involved capturing X-ray images before the surgery. These images were then analyzed to determine the height of the medial condyle (MMH) and lateral condyle (LMH), as well as the ratios between MLW and MMH and MLW and LMH. Additionally, the insall-salvati index (ISI) was calculated, and the severity of any varus or valgus deformities was assessed. Eight machine-learning methods were developed to predict the retention of PCL in TKA. Risk factor analysis was performed using the SHApley Additive exPlanations method.

Results: A total of 307 knee joints from 266 patients were included, among which there were 254 females and 53 males. A stratified random sampling technique was used to split patients in a 70:30 ratio into a training dataset and a testing dataset. Eight machine-learning models were trained using data feeding. Except for the AUC of the LGBM Classifier, which is 0.70, the AUCs of other machine learning models are all lower than 0.70. In importance-based analysis, ISI, MMH, LMH, deformity, and age were confirmed as important predictive factors for PCL retention in operations.

Conclusion: The LGBM Classifier model achieved the best performance in predicting PCL retention in TKA. Among the potential risk factors, ISI, MMH, LMH, and deformity played essential roles in the prediction of PCL retention.

Keywords: Knee joint; Machine learning; Posterior cruciate ligament; Total knee arthroplasty.

Figure 1

Machine learning development process (A) 7 continuous and 5 categorical predictors of PCL retention were taken into the computational process. (B) A stratified random sampling technique was applied to split patients in a 70:30 ratio to a training dataset and a testing dataset. (C) Training dataset was used to identify the optimal hyperparameters which provided the highest accuracy in a tenfold internal cross‐validation of each model. (D) The performance of all algorithms was evaluated with another, unseen, testing dataset.

Figure 2

(A) A lateral X‐ray of the knee joint illustrated the measurement process of the medial condyle height (MMH), lateral condyle height (LMH), and Insall‐salvati index (ISI). (B) An anteroposterior X‐ray of the knee joint demonstrated the measurement process of the width of the epicondyle (MLW).

Figure 3

Receiver‐operating characteristic curve (ROC) for the models developed with all algorithms. Except for the LGBM Classifier's 0.70, the AUCs of other machine learning models were all lower than 0.70.

Figure 4

Characteristics of the selected model (LGBM Classifier): SHAP Value summary graph of top variables and their impact on the prediction.

Figure 5

Variables importance ratio. Top nine crucial variables on the prediction of posterior cruciate ligament preservation.