Development and validation of dynamic clinical subphenotypes for acute pancreatitis using hematocrit and blood urea nitrogen trajectories in intensive care unit: a multicenter retrospective cohort study

• Novel Subphenotype Classification: Identified 5 distinct dynamic clinical subphenotypes of ICU-admitted acute pancreatitis (AP) patients via group-based multi-trajectory modeling (GBMTM) of hematocrit (HCT) and blood urea nitrogen (BUN) trajectories over the first week of admission, namely T1 (Renal Dysfunction), T2 (Fluid-Responsive), T3 (Volume-Deficient), T4 (Stable), and T5 (Hemodilution). • Robust Prognostic Value: Validated across a Chinese development cohort (n = 2027) and U.S. validation cohort (n = 1381) — T1 subtype showed the highest mortality risk (36.6 % in development cohort; adjusted OR = 16.27 in validation cohort), while T4 had the lowest (2.7 %). Trajectory classification outperformed traditional predictors (e.g., APACHE II, creatinine) in mortality prediction via Boruta algorithm. • Subtype-Specific Fluid Therapy Guidance: Revealed divergent fluid tolerance and response across subphenotypes: T1 tolerated 2500–4000 mL on day 1 (excess ↑ mortality); T3 allowed up to 7000 mL on day 1 but required day-2 restriction; T4 had good fluid tolerance, providing a precision framework for individualized resuscitation. • Broad Clinical Applicability: Subgroup and sensitivity analyses confirmed consistent associations between high-risk subtypes (T1, T3, T5) and adverse outcomes (mortality, infected pancreatic necrosis, persistent organ failure) across etiologies, age, gender, and BMI, with robustness to missing data imputation and population restrictions. Acute Pancreatitis (AP) is a common gastrointestinal emergency in which early fluid therapy is critical, yet optimal strategies remain debated. Although Hematocrit (HCT) and Blood Urea Nitrogen (BUN) are recommended to guide fluid management, the prognostic value of their dynamic changes is unclear. This multicenter retrospective cohort study utilized a development cohort from Jiangxi Province (n = 2027) and a validation cohort from the U.S. MIMIC-IV and eICU-CRD databases (n = 1381). Group-based multi-trajectory modeling (GBMTM) was applied to analyze the dynamic changes of HCT and BUN during the first week of ICU admission to identify distinct subphenotypes. Multivariate logistic regression and survival analysis were used to assess the association between each subphenotype and outcomes, including mortality and organ failure. The Boruta algorithm compared predictive importance. Five trajectory subphenotypes were identified: T1 (Renal Dysfunction Subphenotype, n = 287), T2 (Fluid-Responsive Subphenotype, n = 448), T3 (Volume-Deficient Subphenotype, n = 445), T4 (Stable Subphenotype, n = 516), and T5 (Hemodilution Subphenotype, n = 331). The T1 subphenotype had the highest mortality (36.6 % in the development cohort; adjusted OR = 16.27 (95 % CI 6.55–40.45) for in-hospital mortality in the validation cohort), while T4 had the lowest (2.7 %). Subphenotypes exhibited significantly different responses to fluid therapy: for T1, a fluid volume exceeding 4000 mL on the first day increased mortality risk; T3 tolerated a higher initial fluid load (up to 7000 mL on day one) but required fluid restriction on the second day. Consistent results in subgroup and sensitivity analyses demonstrated the robustness of this classification system. This study developed and validated a subphenotype classification system for the acute phase of AP based on the dynamic trajectories of HCT and BUN. It effectively distinguishes patient prognosis and responsiveness to fluid therapy, providing a crucial phenotypic framework for future randomized trials on individualized fluid management.