Abstract / Summary
Abstract Sepsis-associated acute kidney injury (SAAKI) involves rapidly evolving inflammatory and nutritional disturbances, yet most prognostic studies rely on single baseline biomarkers. We examined whether longitudinal red blood cell distribution width-to-albumin ratio (RAR) trajectories identify reproducible mortality-risk phenotypes. This multicenter retrospective study included MIMIC-IV v3.1 as the development cohort ( n = 4,273), MIMIC-III CareVue as temporal validation ( n = 1,287), and eICU-CRD v2.0 as external validation ( n = 3,195). Group-based trajectory modeling characterized 14-day RAR patterns. Multivariable Cox models evaluated 28-day all-cause mortality. Six machine-learning models predicted the highest-risk trajectory using first-24-h variables, with SHAP used for interpretation. Four trajectories were identified: Persistently Low, Moderate-Stable, Moderate-Elevated, and Persistently High. Mortality increased from 17.54% in Persistently Low to 38.48% in Persistently High. Compared with Persistently Low, Persistently High was associated with higher adjusted mortality in MIMIC-IV (HR 2.22, 95% CI 1.79–2.74), MIMIC-III CareVue (HR 3.32, 95% CI 2.31–4.78), and eICU-CRD (HR 1.71, 95% CI 1.40–2.07). XGBoost achieved AUCs of 0.895, 0.918, and 0.904, respectively. SHAP highlighted baseline RAR, lactate, other infection source, albumin, and creatinine. Longitudinal RAR trajectories identify reproducible mortality-risk phenotypes across the three cohorts and may support early risk stratification, although prospective validation is required.