Machine learning-based identification of key risk factors for deep vein thrombosis within one week after total joint arthroplasty in elderly patients

Background Total joint arthroplasty (TJA) is increasingly performed in aging populations, yet postoperative deep vein thrombosis (DVT) remains a life-threatening complication, with an incidence of 31.3% in elderly patients.There is an urgent need to develop and apply effective preventive measures for DVT in this vulnerable patient population. The prevention of DVT is contingent upon precise risk assessment and stratification. Current risk assessment tools are suboptimal for this cohort. Machine learning (ML) offers a promising approach to improve DVT prediction in elderly TJA patients. This study aimed to develop and compare four ML models for predicting postoperative DVT within one week among TJA patients, and to identify the most effective model with optimal predictive performance. Methods A total of 3,050 elderly patients (aged ≥ 60 years) who underwent TJA between January and December 2024 were retrospectively enrolled. Twenty-two candidate predictors, including demographic, clinical, and surgery-related variables, were collected from electronic medical records. The outcome was DVT detected by venous ultrasound within one week postoperatively. Four ML algorithms—Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), and eXtreme Gradient Boosting (XGBoost)—were developed and validated using a hold-out method (80/20 split) with 5-fold cross-validation. The Synthetic Minority Over-sampling Technique (SMOTE) was applied to address class imbalance. Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) was used to interpret the optimal model. Results Of the 3,050 patients, 302 (9.9%) developed postoperative DVT within one week. The AUC values for the LR, RF, DT, and XGBoost model were 0.7629, 0.9652, 0.8054 and 0.9744, respectively, with F1 scores of 0.71416, 0.905764, 0.816976, and 0.934701, respectively. SHAP analysis identified prothrombin time, C-reactive protein, D-dimer, and sex as the most influential predictors, with preoperative inflammatory and coagulation markers playing dominant roles in early DVT risk stratification for elderly TJA patients. Conclusions The XGBoost model demonstrated superior predictive accuracy for early postoperative DVT among elderly TJA patients compared to LR, RF, and DT. Its high performance, coupled with SHAP-based interpretability, supports its potential integration into clinical decision support systems for individualized DVT risk assessment.