Development and Validation of an Interpretable Machine Learning Model for Predicting Postoperative Venous Thromboembolism in Urological Cancer Patients

Purpose To develop and validate a machine learning-based model for predicting postoperative venous thromboembolism (VTE) in patients with urological cancer, and to compare its performance with a logistic regression baseline model. Methods Clinical data of 474 patients who underwent urological cancer surgery were retrospectively collected and randomly divided into a training set and a test set at a 7:3 ratio. Univariate and multivariate logistic regression analyses were used to identify independent risk factors for VTE. Five machine learning models—support vector machine (SVM), random forest, LightGBM, XGBoost, and LASSO—were developed, with logistic regression as the baseline classifier. Five-fold cross-validation was used for hyperparameter tuning, and class weights were applied to address class imbalance. Model performance was evaluated on the test set using the area under the receiver operating characteristic curve (AUC) as the primary metric, with sensitivity, specificity, positive predictive value, and negative predictive value as secondary metrics. Variable importance was assessed using SVM linear kernel coefficients, and decision curve analysis (DCA) was performed to evaluate clinical net benefit. Results Among the 474 patients, 32 (6.75%) developed postoperative VTE. Multivariate logistic regression identified postoperative bed days (OR = 1.633, P < 0.001) and surgery time (OR = 1.006, P = 0.003) as independent risk factors for VTE. Among the five machine learning models, SVM performed best, with a test set AUC of 0.787, sensitivity of 91.7%, specificity of 60.8%, and negative predictive value of 98.8%. Variable importance analysis showed that postoperative bed days was the most important predictor, followed by surgery time and postoperative APTT. Decision curve analysis indicated that the SVM model provided clinical net benefit across threshold probabilities of 0.05 to 0.30. Conclusion The SVM model developed in this study demonstrated good predictive performance. Its high negative predictive value (98.8%) supports its use as an effective rule-out tool for low-risk VTE patients, providing a reference for clinical nursing decisions. However, the generalizability of the model requires further validation in multicenter prospective studies.