Clinical study of XGBoost model-based prediction of postoperative deep vein thrombosis risk in patients with lower extremity fractures

LI Tao-tao; ZHANG Yu-lu; MA Li; SONG Tian-tian; CHU Yu-qing; LI Li

doi:10.3969/j.issn.2095-252X.2025.04.008

Abstract：Objective A model for predicting the risk of postoperative deep vein thrombosis(DVT)in patients with lower extremity fractures was constructed based on machine learning(ML)models and Shapley's additive interpretation(SHAP).Methods A total of 202 patients with lower extremity fractures who were surgically treated between February 2022 and December 2023 in our hospital were selected.Important characteristic variables of postoperative DVT were screened by Boruta's algorithm.202 patients were divided into a training set(n=121)and a test set(n=81)in a 3:2 ratio to construct and train 9 ML models.The predictive performance of the 9 ML models was assessed using receiver operating curves(ROC).The ML models were additionally interpreted by SHAP values and constructed to predict the risk of postoperative DVT in patients with lower extremity fractures.Results The prevalence of postoperative DVT in 202 patients with lower limb fractures was 20.3%.Boruta algorithm screened for fasting blood glucose(FBG),red blood cells(RBC),haemoglobin(Hb),D-dimer(D-D),fibrinogen(FIB),and smoking were important characterizing variables for postoperative DVT.Among the nine ML algorithms,the ROC of the training and test sets confirmed that the extreme Gradient Boosting(XGBoost)model had the highest performance in predicting the risk of postoperative DVT.The XGBoost model based on additional interpretation and visualization of SHAP values predicted DVT risk with very high accuracy and generated an online application.Conclusions An XGBoost model based on the interpretation of SHAP values accurately predicts postoperative DVT risk in patients with lower limb fractures.An online application is developed to conveniently calculate patients'postoperative DVT risk.