Development of a machine learning-based predictive model for venous thromboembolism risk assessment in orthopaedic patients with routine prophylaxis.

Abstract

Despite the use of conventional preventive measures, the long-term risk of the development of venous thromboembolism (VTE) in orthopaedic patients remains high in a high-risk patient population. Accurate risk assessment is critical; however, existing assessment tools appear to have certain limitations, and machine learning (ML) models appear to have higher predictive accuracy. Develop an ML model with clinical features to predict VTE in orthopaedic patients on standard prophylaxis. We used 147 clinical variables with XGBoost and CatBoost models for VTE risk prediction, comparing their performance with the Caprini score. Both internal and external validations were conducted to assess the model's efficacy. SHapley Additive exPlanation (SHAP) values were employed to improve interpretability and accurately evaluate predictive efficacy. Using 8182 patients (153 VTE cases), XGBoost and CatBoost achieved internal Area Under the ROC curves (AUCs) of 0.941 and 0.937. In external validation (2121 patients; 31 VTE cases), AUCs were 0.888 and 0.902. They outperformed traditional methods with high accuracy, balanced sensitivity and specificity. SHAP analysis showed feature importance and VTE correlation across algorithms. This study used two models with clinical features to improve VTE risk prediction accuracy in orthopaedic patients under conventional prevention. The models identified VTE risk factors and highlighted key preventive measures.