Development and temporal validation of a machine learning model for predicting in-hospital postoperative deep vein thrombosis in adult patients with tibial shaft fractures and concomitant ipsilateral fibular fractures

Background Imaging-detected postoperative deep vein thrombosis (DVT), including asymptomatic events, is an important complication after surgery for tibial shaft fractures with concomitant ipsilateral fibular fractures. We aimed to develop and temporally validate a machine-learning (ML) model for predicting in-hospital postoperative DVT using routinely available admission and preoperative variables. Methods This single-center retrospective study included adults undergoing surgical fixation between January 2015 and December 2025. The outcome was surveillance-detected in-hospital lower-extremity DVT. Patients admitted in 2015–2022 formed the development cohort and those admitted in 2023–2025 the temporal validation cohort. Following feature selection, seven algorithms were compared using development-cohort cross-validation, and the selected model was locked before temporal validation. Additional analyses included repeated nested cross-validation and exploratory cross-validated recalibration. Performance was assessed against a conventional logistic regression benchmark using discrimination, calibration, and decision-curve analysis. SHapley Additive exPlanations (SHAP) characterized predictor contributions. Results Among 643 patients, 104 (16.2%) developed in-hospital DVT. The development and temporal validation cohorts included 473 patients with 73 events and 170 patients with 31 events, respectively. Seven predictors were retained: age, time from injury to surgery, hematocrit, white blood cell count, albumin, fibrinogen, and albumin-to-globulin ratio. In temporal validation, SVM achieved an AUROC of 0.934 (95% CI, 0.892–0.969) and Brier score of 0.084, compared with 0.918 (95% CI, 0.871–0.957) and 0.081 for conventional logistic regression. The AUROC difference was not statistically significant ( P = 0.130). Repeated nested cross-validation of the complete development procedure yielded a mean AUROC of 0.894 (SD 0.004). The original SVM overestimated risk; exploratory cross-validated recalibration reduced the Brier score to 0.077 and moved calibration summaries closer to their ideal values. Decision curves showed broadly similar net-benefit patterns for SVM and logistic regression. Conclusions The SVM model showed favorable discrimination for surveillance-detected in-hospital DVT, but its incremental value over conventional logistic regression was not established. Exploratory recalibration findings require independent confirmation, and clinical utility remains unproven. External validation and prospective impact evaluation are needed before clinical implementation.