Sercan ÇAPKIN, Ali İhsan KILIÇ, Mustafa ÇELTİK, Onur KARAGÖZ, Mehmet AKDEMİR
Comprehensive Medicine - 2026;18(3):278-286
Objective: Preoperative component sizing in total knee arthroplasty (TKA) can streamline surgery and support alignment and stability. Linear regression models based on basic demographic and anthropometric measures are practical, but their linear assumptions may miss complex anatomical variation. Although machine learning has been increasingly used for prediction tasks, head-to-head evidence against standard regression models in homogeneous TKA series remains limited. Materials and Methods: We retrospectively reviewed 224 unilateral primary posterior-stabilized TKAs performed by a single surgeon using one implant system. Predictors included age, sex, height, weight, body mass index, and shoe size. Multivariable linear regression was compared with Random Forest and XGBoost. Models were trained on 80% of cases with 10-fold cross-validation and evaluated on a held-out 20% test set using R², RMSE, MAE, and +/-1-size accuracy. Results: Height and shoe size were the strongest predictors. The linear regression fit was good (femoral R²=0.63; tibial R²=0.59). Femoral sizing was comparable across approaches (test R²=0.48; +/-1 accuracy 37.8-44.4%). For tibial sizing, ML outperformed regression: Random Forest achieved a test R²=0.60 with +/-1 accuracy of 48.9%, and XGBoost achieved a test R²=0.56 with +/-1 accuracy of 53.3%, versus linear regression (R²=0.40; +/-1 accuracy 33.3%). Conclusion: In this homogeneous cohort, linear regression predicted femoral size well, whereas tibial sizing showed greater non-linearity and benefited from ensemble ML. Still, accuracy was insufficient to replace intraoperative judgment. These models may assist preoperative inventory planning and trial component preparation, but external validation across implant systems is needed.
