Dynamic cardio-renal-metabolic trajectories identify sepsis subphenotypes for individualized management: a cross-population cohort study

Sepsis heterogeneity remains a major barrier to precision medicine. Although cardio-renal-metabolic interactions critically influence outcomes, their dynamic evolution is rarely integrated into clinical phenotyping. We aimed to identify distinct sepsis subphenotypes based on dynamic cardio-renal-metabolic trajectories to facilitate individualized management. In this retrospective cross-population cohort study, adult patients with sepsis from the MIMIC-IV database comprised the development cohort, with external validation in the MIMIC-III and Fuzhou University Affiliated Provincial Hospital cohorts. Latent class mixed models were applied to jointly characterize 120-hour trajectories of the triglyceride-glucose (TyG) index, SOFA-Cardiovascular subscore, and SOFA-Renal subscore. Multivariable Cox regression evaluated associations between subphenotypes and 30-day mortality. Multiple machine learning algorithms incorporating Shapley Additive exPlanations (SHAP) analysis were used for early prediction of the high-risk subphenotype. Subphenotype-stratified nonlinear risk modeling was used to explore subphenotype-specific associations between early fluid administration and 30-day mortality. In the development cohort ( n = 1,422), four dynamic subphenotypes, labeled Phenotypes A–D, were identified and reproduced in the validation cohorts. Phenotypes C and B exhibited mild-to-moderate organ dysfunction and relatively lower 30-day mortality (22.7% and 27.7%, respectively). Phenotype D was characterized by sustained organ dysfunction and metabolic dysregulation. Notably, Phenotype A presented with moderate initial severity but developed synchronized cardio-renal-metabolic decompensation, with the highest 30-day mortality risk (adjusted HR 1.82 [95% CI 1.20–2.75] vs. Phenotype C; Phenotype D: HR 1.45 [95% CI 1.13–1.85]). LightGBM, XGBoost, and EBM yielded ROC-AUCs of 0.751–0.895 in the development cohort and 0.717–0.762 in the external cohorts. The TyG index, blood urea nitrogen, and Charlson Comorbidity Index were identified as leading predictors of Phenotype A. Associations between early fluid administration and predicted mortality differed across subphenotypes. Integrating dynamic cardio-renal-metabolic trajectories identified four reproducible sepsis subphenotypes with distinct clinical outcomes. Early prediction may support the recognition of patients at risk for delayed deterioration. Associations between early fluid administration and mortality differed across subphenotypes.