Explainable machine learning for predicting lower extremity deep vein thrombosis in traumatic brain injury patients: development of a SHAP-guided Random Forest model

Background: Lower extremity deep vein thrombosis (LEDVT) is a common and clinically important complication after traumatic brain injury (TBI). Early identification of patients at increased thrombotic risk remains challenging. We aimed to develop and internally validate an interpretable machine-learning model for predicting LEDVT in patients with TBI. Methods: = 74). A total of 39 demographic, clinical, and laboratory variables were screened. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression, followed by multivariable logistic regression. Eight machine-learning algorithms were developed and compared. Model discrimination, calibration, and decision-curve performance were assessed, and the selected model was interpreted using Shapley additive explanations (SHAP). Results: LEDVT occurred in 88 of 248 patients (35.5%). LASSO retained six variables for model development: age, platelet count, D-dimer, interleukin, lower limb fracture, and prophylactic anticoagulation. In multivariable logistic regression, D-dimer, interleukin, lower limb fracture, and prophylactic anticoagulation were significantly associated with LEDVT, whereas age and platelet count showed borderline associations. In the validation cohort, the Random Forest model showed strong discrimination, with an area under the receiver operating characteristic curve of 0.931 (95% CI 0.878-0.985), sensitivity of 78.8%, specificity of 95.1%, accuracy of 87.8%, precision of 92.9%, and F1 score of 0.852. SHAP analysis identified D-dimer, interleukin, prophylactic anticoagulation, age, platelet count, and lower limb fracture as the main contributors to model output. Conclusion: An interpretable Random Forest model based on routinely available clinical and laboratory variables showed good internal predictive performance for LEDVT after TBI. After external validation, this approach may help support early risk stratification and individualized surveillance in patients at increased thrombotic risk.