Machine learning prediction of deep vein thrombosis before staged surgery in lower-leg and ankle fractures treated with temporary external fixation

Background Deep vein thrombosis (DVT) detected between temporary external fixation and staged definitive surgery can complicate perioperative management of lower-leg and ankle fractures. We developed machine-learning models to estimate DVT probability during this interval. Methods This retrospective cohort included adults admitted during 2015−2025 who underwent staged fixation and bilateral lower-extremity venous ultrasonography. A stratified 70:30 split defined development and validation sets. Development-set imputation, Boruta feature selection, and five-fold cross-validation were used to develop 12 classifiers. Algorithms were compared in the validation set. The selected Naive Bayes pipeline underwent 500 development-cohort bootstrap replications for optimism correction. Discrimination, precision-recall performance, calibration, and decision curves were evaluated. SHapley additive explanations (SHAP) and a bilingual research prototype web application supported interpretation. Results Among 356 patients, 55 (15.4%) had DVT, including 40 with calf muscular vein thrombosis. The development and validation sets comprised 250 patients (39 events) and 106 patients (16 events). Boruta retained nine predictors. Naive Bayes had the highest validation receiver operating characteristic area under the curve (AUC), 0.807 (95% confidence interval, 0.705−0.909). Precision-recall AUC was 0.343, sensitivity was 0.875, and specificity was 0.589. Beta calibration reduced the Brier score from 0.135 to 0.116 without changing discrimination. The apparent AUC was 0.858, mean bootstrap optimism was 0.084, and optimism-corrected AUC was 0.773. Injury-to-admission time, calcium, age, platelet count, fibrinogen, and hemoglobin had the largest mean absolute SHAP values. Conclusion The model may support risk stratification for an ultrasound-detected composite DVT endpoint in this staged-treatment population. Its proposed use is to inform surveillance timing and intensity while retaining routine bilateral ultrasonography. Prospective validation is required before clinical implementation.