Machine Learning Based Prediction of Tibial Insert Thickness in Total Knee Arthroplasty From Intraoperative Knee Joint Laxity Data.

Abstract

Total knee arthroplasty (TKA) represents the gold standard for relieving pain and restoring function in patients with end-stage knee osteoarthritis. Soft-tissue balancing is critical to achieving successful outcomes. One of the factors that contribute to successful soft-tissue management relates to tibial insert thickness, an intraoperative surgical decision based on surgeon experience and preference. This study investigates the relationship between knee joint laxity curves and tibial insert thickness selection in tibia-first TKA and explores predictive modeling to support intraoperative decision making. Data from 1661 tibia-first TKA cases performed by 11 surgeons were used to develop surgeon-dependent and surgeon-independent models. Surgeon-dependent models are personalized to individual surgeons, leveraging data specific to each surgeon. While surgeon-independent models are developed utilizing data from seven expert surgeons (> 70 cases each) to provide generalized recommendations. Three supervised machine learning (ML) algorithms logistic regression (LR), random forest (RF), and XGBoost (XGB) were employed with feature selection methods: correlation-based feature selection (CFS), recursive features elimination (RFE), and Shapley additive explanations (SHAP). The best surgeon-dependent model achieved a mean exact prediction accuracy of 58.3%, mean prediction within 1 mm accuracy of 73%, and mean prediction within 2 mm accuracy of 93.1%. The top-performing surgeon-independent model demonstrated improved accuracy, with a mean exact prediction accuracy of 61.3%, mean prediction within 1 mm accuracy of 73.3%, and mean prediction within 2 mm accuracy of 94.2%. These findings suggest that ML models can assist in standardizing tibial insert thickness selection, potentially reducing variability and improving intraoperative decision-making in TKA.