Explainable machine learning for predicting venous thromboembolism in septic shock patients.

Background: Venous thromboembolism (VTE) frequently complicates septic shock, yet precise, individualized risk stratification tools remain scarce. This study aimed to develop and externally validate an explainable machine learning (ML) framework to predict VTE in this critically ill population. Methods: A retrospective cohort study was conducted including adult septic shock patients admitted between January 2020 and December 2025. The study population was partitioned into an internal development cohort (n=733) and an independent external validation cohort from a separate tertiary hospital (n=257, Xi'an No. 3 Hospital). We utilized the Boruta algorithm alongside recursive feature elimination to isolate optimal predictors. Six ML algorithms were trained and evaluated using metrics including the area under the receiver operating characteristic curve (AUC) and F1 score. Shapley Additive Explanations (SHAP) were integrated to establish model transparency. Results: The VTE incidence within the development cohort was 17.74% (130/733). The feature selection pipeline distilled six robust predictors: fibrin degradation products (FDP), prothrombin time (PT), white blood cells (WBC), activated partial thromboplastin time (APTT), D-dimer, and C-reactive protein (CRP). Among the evaluated models, the Random Forest (RF) algorithm exhibited superior discriminative capacity and promising performance in an independent external validation cohort, achieving an AUC of 0.9718 and an F1 score of 0.7917 in the independent external validation cohort, with a sensitivity of 0.7037. SHAP analysis revealed that heightened thrombo-inflammatory markers combined with abbreviated coagulation intervals fundamentally drove VTE risk, offering personalized predictive insights via individual force plots. Conclusions: We successfully established a highly accurate and interpretable RF-based predictive model for VTE in septic shock patients. By leveraging six routine clinical biomarkers and SHAP-derived transparency, this tool bridges complex algorithmic forecasting with clinical intuition, providing a transparent risk assessment framework that may assist in risk stratification for thromboprophylaxis after prospective validation. Future implementation studies are needed to assess its real-world clinical utility and impact on patient outcomes.