Construction and Internal Validation of a Machine Learning-Based Clinical Prediction Model for Deep Vein Thrombosis after Total Knee Arthroplasty

Objective Deep vein thrombosis (DVT) is among the most serious complications that can occur after total knee arthroplasty (TKA) and poses a significant threat to patients’ postoperative recovery and survival. The aim of this study is to develop and internally validate an individualized machine learning (ML) model to identify important predictors and predict the risk of DVT after TKA. Methods In this study, information on patients who underwent TKA at the First Affiliated Hospital of the University of Science and Technology of China from January 2024 to December 2024 was collected. Least absolute shrinkage and selection operator (LASSO) was used to identify key clinical features, and eight ML models were constructed. Model performance was evaluated using the area under the curve (AUC), sensitivity, specificity, accuracy, Brier score, expected calibration error (ECE), and calibration curves. Shapley additive explanations (SHAP) were used to interpret the final model. Results This study ultimately included 907 patients who underwent TKA. The eXtreme Gradient Boosting (XGB) model performed the best, with an AUC of 0.916 (95% CI: 0.870–0.962), a sensitivity of 0.816, a specificity of 0.880, an accuracy of 0.848, a Brier score of 0.0694 and an ECE of 0.0481. According to the SHAP analysis, the clinical features most strongly associated with the prediction of DVT after TKA were, in descending order, platelet count, uric acid level, age, sex, and augmented renal clearance (ARC). Conclusion This study developed and internally validated an interpretable ML model for predicting the risk of DVT after TKA. Platelet count, uric acid level, age, sex, and elevated preoperative estimated creatinine clearance (CrCl ≥130 mL/min) were identified as important predictors, and the XGB model demonstrated the best overall predictive performance among the eight models.