Development and Validation of a Nomogram for Predicting Preoperative Venous Thromboembolism in Patients With Lower Extremity Fractures: A Retrospective Cohort Study
Ruoying He, Zuxin Feng, Zhao Fan, Zijian Liu, Li Liang, Yu ZhaoBackground
Patients with lower extremity fractures (LEF) are at high risk of preoperative venous thromboembolism (VTE). However, existing risk assessment models are not specifically developed and validated for this population.
Objective
To develop and validate a prediction model for preoperative VTE in emergency LEF patients, and to compare the performance of traditional logistic regression (LR) with that of machine learning (ML) algorithms.
Design
Retrospective cohort study.
Setting
Peking University Shenzhen Hospital, China.
Participants
486 patients with LEF admitted between January 2022 and July 2025.
Methods
Two modeling strategies were compared: 1) a traditional LR nomogram using backward stepwise selection, and 2) nine ML algorithms with Least Absolute Shrinkage and Selection Operator (LASSO)-based feature selection. Model performance was evaluated using discrimination (area under the curve, AUC), calibration, classification metrics (accuracy, sensitivity, specificity) and decision curve analysis.
Results
Of 486 enrolled patients, 79 (16.3%) developed preoperative VTE. Among the ML models, LASSO-LR and multilayer perceptron (MLP) achieved the highest validation AUCs (0.709 each), the best calibration and highest net benefit. The traditional LR nomogram showed an identical AUC of 0.709 (95% CI: 0.691–0.718), with superior calibration (Hosmer-Lemeshow
Conclusion
The traditional LR nomogram appears promising for preoperative VTE risk prediction in LEF patients. External validation is warranted before clinical implementation.