A LightGBM model was developed to predict sepsis-induced coagulopathy (SIC) within 72 h of ICU admission using routine clinical data. The model achieved an ROC-AUC of 0.937 (95%CI 0.886–0.963) and a Brier score of 0.106. SHAP analysis identified international normalized ratio, platelet count, and lactate as key predictors. External validation confirmed robustness (ROC-AUC 0.938). Sensitivity analysis excluding SIC diagnostic components yielded an ROC-AUC of 0.754, indicating genuine pathophysiological associations rather than data leakage. To establish a machine learning model for predicting sepsis-induced coagulopathy (SIC) within 72 h of ICU admission. This retrospective cohort study utilized the MIMIC-IV database (2008–2019) to identify first-time ICU admissions with sepsis. Forty variables were extracted. Random Forest importance scores selected the top ten features. Eight machine learning algorithms were compared. Performance was evaluated by ROC-AUC, accuracy, precision, recall, F1 score, and Brier score. SHapley Additive exPlanations (SHAP) analysis was employed for interpretation. Generalizability was assessed using an independent external cohort (2022–2024). A total of 2,237 patients were included (training set: 1,789; external validation: 202). The LightGBM model demonstrated optimal performance (ROC-AUC 0.937, 95%CI 0.886–0.963; accuracy 0.866; Brier score 0.106). SHAP analysis identified international normalized ratio, platelet count, and lactate as the top three predictors. External validation confirmed robust discriminative ability (ROC-AUC 0.938). Sensitivity analysis excluding SIC diagnostic components (INR, PLT, SOFA) yielded ROC-AUC 0.754, indicating genuine pathophysiological associations rather than data leakage. The LightGBM model offers a preliminary tool for early SIC detection. Multi-center prospective validation is warranted to confirm clinical utility.
Li et al. (Sat,) studied this question.