Web-based machine learning prediction tool for venous thromboembolism in patients with rib and spinal fractures: a retrospective cohort study

Venous thromboembolism (VTE) is a major cause of morbidity and mortality in trauma patients, especially in patients with rib and spinal fractures. Early risk stratification remains challenging. This study aimed to develop and validate a machine learning model for predicting VTE risk in this patient population. This multicenter retrospective cohort study included patients with rib fractures combined with spinal fractures from Shanghai Sixth People’s Hospital (development cohort, n = 1007, randomly split 7:3 into training [ n = 704] and internal validation [ n = 303] sets) and Shanghai Fengxian District Central Hospital (external validation cohort, n = 203). The overall VTE incidence was 27.0% (272/1,007) in the development cohort and 21.7% (44/203) in the external validation cohort. All 47 candidate predictors were restricted to variables obtainable within 24 hours of admission. Potential predictors were screened using three methods, including univariate analysis, least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. Seven machine-learning algorithms were used to develop prediction models. The performance of all models was evaluated in terms of discrimination, calibration, and clinical utility using the area under the receiver operating characteristic curve (AUC), Brier score, and decision curve analysis (DCA). Model interpretation was performed using SHapley Additive exPlanations (SHAP). The final model was implemented as a web-based risk estimation tool incorporating a model-derived three-tier risk stratification framework. Seven predictors were retained, including spinal cord injury, spinal fracture stability, D-dimer (within 24 h of admission), thoracic spine–rib–sternum fracture, pelvic or lower limb fracture, age group, and rib fracture count group. The support vector machine (SVM) model achieved an AUC of 0.900 (95% CI, 0.865–0.934) in internal validation, with a sensitivity of 0.939 and a negative predictive value (NPV) of 0.969. In external validation, the AUC was 0.901 (95% CI, 0.846–0.949), with a sensitivity of 0.955 and an NPV of 0.981. Calibration assessment and DCA demonstrated satisfactory model performance and clinical utility. SHAP analysis identified spinal cord injury, spinal fracture stability, and D-dimer as the most influential predictors. A web-based calculator was developed to provide individualized risk estimation using a model-derived three-tier risk stratification framework, with predicted-risk categories of low (<18.0%), intermediate (18.0–41.3%), and high (≥41.3%). The SVM model demonstrated favorable discrimination, calibration, and potential clinical utility in both internal and external validation cohorts. Using routinely available variables within 24 hours of admission, the model enabled early VTE risk estimation and stratification in patients with combined rib and spinal fractures. The web-based tool may facilitate individualized risk assessment. However, further prospective multicenter validation is needed before clinical implementation.