Interpretable prediction of in-hospital venous thromboembolism in critically ill patients with heart failure receiving early thromboprophylaxis

Venous thromboembolism (VTE) remains a concern among critically ill patients with heart failure receiving early thromboprophylaxis, yet individualized risk stratification in this population remains limited. We developed nine prediction models using 8,825 patients from MIMIC-IV and externally validated them in 5,425 eICU records. Fifteen predictors were identified through combined least absolute shrinkage and selection operator (LASSO) regression and Boruta feature selection. XGBoost achieved the highest observed discrimination, with ROC AUCs of 0.764 (95% CI, 0.723–0.805) internally and 0.731 (95% CI, 0.676–0.786) externally. SHAP analysis identified atrial tachycardia/fibrillation, respiratory failure, previous thromboembolism, antiplatelet therapy, and mechanical ventilation within 24 h as the leading contributors to model predictions. External validation highlighted the value of recalibration, with exploratory five-fold cross-fitted intercept recalibration improving overall calibration and predictive accuracy. Decision-curve analysis suggested potential net benefit at selected low-risk thresholds following recalibration. These findings demonstrate the feasibility of an interpretable prediction model based on routinely available clinical data, with potential for integration into ICU workflows to support individualized VTE risk stratification and targeted surveillance.