Dynamic Trajectories of Sepsis-Induced Coagulopathy and Machine Learning-Based Early Prediction of High-Risk Phenotypes: A Multicenter Cohort Study

Abstract Background Sepsis-induced coagulopathy (SIC) is usually assessed at a single time point, although coagulation and organ dysfunction evolve rapidly during critical illness. We sought to identify reproducible longitudinal SIC-score trajectories and determine their prognostic relevance across heterogeneous critical-care databases. Methods This multicohort retrospective study included 6,416 patients from MIMIC-IV for development and 1,157 from MIMIC-III CareVue, 2,077 from eICU-CRD, and 641 from NWICU for validation. Group-based trajectory modeling was applied to serial SIC scores during the first 7 days after ICU admission. Associations with 28-day mortality were evaluated using multivariable Cox models, multiple imputation, calibrated inverse probability of treatment weighting (IPTW), proportional-hazards diagnostics, and a day-7 landmark analysis. Incremental discrimination and a nested machine-learning classifier for the persistently high phenotype were also assessed. Results Three trajectories were identified: Low-Rapid Decline (19.25%), Persistently High (32.00%), and Moderate-Decline (48.75%). The three-class model had an entropy of 0.947, class-specific average posterior probabilities above 0.97, and a minimum odds of correct classification of 52.47. Compared with the Low-Rapid Decline class, the Persistently High class had higher adjusted 28-day mortality hazards in MIMIC-IV (hazard ratio [HR] 2.89, 95% confidence interval [CI] 2.40–3.48), MIMIC-III (HR 2.53, 95% CI 1.75–3.67), eICU (HR 2.45, 95% CI 2.01–2.99), and NWICU (HR 3.93, 95% CI 2.58–5.99). Findings were directionally consistent in IPTW and day-7 landmark analyses. Adding trajectory class to SOFA modestly improved discrimination for subsequent mortality. A nested XGBoost classifier identified the Persistently High phenotype with AUCs of 0.830–0.862 across temporal and external validation cohorts. Conclusions Early SIC-score trajectories define reproducible prognostic subphenotypes across four critical-care databases. A persistently high trajectory identifies patients with sustained excess mortality risk, while trajectory information adds a modest temporal dimension beyond static SIC and SOFA assessment. These observational subphenotypes are not causal or treatment-defining and require prospective validation before clinical implementation.