Cross-algorithm machine learning and consensus features for early disseminated intravascular coagulation risk prediction in sepsis-induced coagulopathy.

backgroundSepsis-induced coagulopathy (SIC) represents an early and potentially reversible stage of coagulation dysfunction in sepsis, whereas progression to disseminated intravascular coagulation (DIC) is associated with multi-organ failure and markedly increased mortality. Although existing diagnostic criteria and machine learning (ML) models have demonstrated utility in broad sepsis populations, they lack specificity for identifying high-risk subgroups within SIC patients who are most likely to deteriorate into DIC. Timely and precise risk stratification at this critical transition point remains an unmet clinical need.

methodsWe conduct a retrospective, multicenter study integrating data from MIMIC-IV v3.1 database and an independent cohort from the First Affiliated Hospital of Wenzhou Medical University. A total of 4,874 SIC patients are included, comprising 1,629 patients from the local cohort and 3,245 patients from MIMIC-IV. Using 61 routinely available clinical variables collected within 48 h of ICU admission, we design a two-stage ML framework. In phase I, 16 ML algorithms are combined with 11 feature importance ranking methods and backward elimination to identify algorithm-specific optimal feature subsets. In phase II, a cross-algorithm consensus strategy is applied to extract globally stable features and construct a minimalist, interpretable predictive model. Model performance is evaluated using accuracy (ACC), area under the receiver operating characteristic curve (AUC), and F1-score. External validation was performed using the MIMIC-IV cohort. Model interpretability was assessed through permutation feature importance, SHapley Additive exPlanations (SHAP), and a clinically deployable nomogram, with calibration curves and decision curve analysis (DCA) used to assess clinical utility.

resultsAcross single-algorithm optimization, the Gradient Boosting Classifier (GB) model achieved the best balance between performance and robustness. Cross-algorithm consensus analysis identified seven highly stable predictors: intensive care unit (ICU) length of stay (LOS), vasopressor, albumin (ALB) level, Modification of Diet in Renal Disease (MDRD) estimate, Logistic Organ Dysfunction System (LODS) score, Sequential Organ Failure Assessment (SOFA) score, Continuous Renal Replacement Therapy (CRRT). The final GB model incorporating these seven features achieved an F1-score of 0.71 and an AUC of 0.75 in the internal validation cohort, while demonstrating excellent generalization in the external validation cohort (F1-score 0.92, AUC 0.74, ACC 0.92). Interpretability analyses confirmed that disease severity, organ dysfunction, and treatment intensity were the primary drivers of DIC progression. The nomogram exhibited good calibration and provided consistent net clinical benefit across a wide range of risk thresholds in DCA.

conclusionsWe propose a robust and interpretable ML framework for early identification of SIC patients at high risk of progression to DIC, based on only seven routinely available clinical indicators. By transitioning from algorithm-specific optimization to cross-algorithm consensus feature selection, this study achieves a favorable balance between predictive performance, generalizability, and clinical usability. This model offers a practical decision-support tool for precision risk stratification and targeted intervention in sepsis-associated coagulopathy. Prospective multicenter studies incorporating dynamic time-series data are warranted to further validate and extend its real-time clinical applicability.