Predicting Mortality Risk in Sepsis-Induced Early Coagulopathy: A Multicenter Comparison of Machine Learning and Nomogram Approaches

Objective This study aimed to compare the performance of machine learning models with a nomogram model in predicting 28-day mortality risk in patients with sepsis-induced early coagulopathy (SIC), and to quantitatively evaluate the improvement in risk stratification using the net reclassification index (NRI). The goal is to provide evidence for selecting superior early warning tools in clinical practice. Methods A multicenter retrospective cohort study design was employed, utilizing data from the MIMIC-IV database (training and internal validation sets) and the eICU-CRD database (external validation set). Through rigorous variable screening (nomogram used Lasso regression combined with Bayesian Information Criterion (BIC) criteria; eXtreme Gradient Boosting (XGBoost) model used a comprehensive importance assessment based on Gain, Frequency, and Coverage metrics), a nomogram (based on logistic regression) and an XGBoost machine learning model were constructed. Model performance was comprehensively evaluated based on discrimination (Area Under the Curve (AUC)/Concordance index (C-index)), calibration (Brier score, calibration curves), clinical utility (decision curve analysis), and NRI. Results A total of 2275 patients were included (1637 from MIMIC-IV, 638 from eICU-CRD). Both models demonstrated good predictive performance: the nomogram had C-indices of 0.828 and 0.807 in the internal and external validation sets, respectively; the XGBoost model achieved an AUC of 0.968 in the training set, remaining above 0.826 and 0.793 in the internal and external validation sets, respectively. Core predictors included maximum lactate (Lac_max), oxygenation index (OI), age, lactate (Lac), and platelet-related indicators. NRI analysis showed that compared to the nomogram, the XGBoost model achieved significant net reclassification improvement (continuous NRI = 0.458, p  < 0.001), indicating superior discriminatory ability in risk stratification. Conclusion This study successfully constructed and validated a nomogram and an XGBoost model for predicting 28-day mortality risk in patients with sepsis-induced early SIC. Both models exhibited good performance and clinical utility. The nomogram facilitates rapid manual assessment in clinical settings, while the XGBoost model can provide more precise risk stratification when computational resources are available. This research provides new tools for the early identification of high-risk SIC patients, which may aid clinical decision-making and potentially improve patient outcomes.