Sepsis-induced coagulopathy (SIC) is a frequent complication of sepsis that significantly impacts clinical outcomes, underscoring the necessity for early identification to facilitate timely intervention. This study aimed to develop an interpretable machine-learning model to predict early onset SIC within 3 days. In this multicenter retrospective cohort study, four distinct cohorts derived from the MIMIC-IV dataset, the eICU Collaborative Research Database, and an ICU cohort from Zhongda Hospital, Southeast University, were used for model derivation and validation. Core predictive variables were identified through multi-step feature engineering using data collected within the first 24 h of ICU admission. Five machine learning models were trained based on these features and subsequently evaluated on three external validation cohorts. The derivation cohort, drawn from MIMIC-IV v2.2 database, included 11,033 patients. External validation encompassed three cohorts totaling 5,572 patients: MIMIC-IV v3.1 dataset (n = 1,371), eICU Database (n = 2,848), and Zhongda Hospital database (n = 1,353). Among five evaluated machine learning models, XGBoost demonstrated the best discriminative performance. The final model incorporated seven predictive variables: International Normalized Ratio, lactate, platelet count, Sequential Organ Failure Assessment (SOFA) circulation, renal, and respiratory subscores, and mechanical ventilation status. The model achieved robust prediction of early-stage SIC across validation sets (internal validation set AUC = 0.83; internal test set AUC = 0.82; external validation set AUC = 0.93,0.86,0.83, respectively). However, the model was ineffective for predicting late-onset SIC, with an internal test set AUC of 0.5 and external validation set AUCs of 0.48, 0.53, and 0.6, respectively. This study establishes the feasibility of an interpretable XGBoost model for accurately predicting early-onset SIC. As a promising clinical decision-support tool, its real-world utility and impact on patient outcomes should be further established through prospective validation prior to deployment.
Peng et al. (Tue,) studied this question.