Early prediction of novel dynamic subphenotypes in acute pancreatitis with hypertriglyceridemia: a multicenter cohort study

Abstract

Concomitant with the escalating global burden of metabolic disorders, acute pancreatitis with hypertriglyceridemia (AP-HTG) has become highly prevalent. However, the initial triglyceride elevation does not strictly correlate with adverse clinical outcomes. Therefore, we aimed to identify novel dynamic subphenotypes reflecting the intrinsic pathophysiological state and develop an early prediction model. Clinicopathological data were collected from MIMIC-IV ( n = 344, development), eICU-CRD ( n = 173, external validation), and an independent local cohort ( n = 319, pragmatic validation). Group-based multi-trajectory modeling (GBMTM) of 7-day white blood cell and calcium trajectories was utilized to identify novel subphenotypes. After evaluating multiple machine-learning algorithms, an optimal early prediction model was established using Boruta, LASSO, and logistic regression based on admission-day indicators. Four novel dynamic subphenotypes were identified, with the Severe-Hyperinflammatory (C3) and Lipotoxic-Hypocalcemic (C4) subphenotypes carrying substantially increased independent risks for persistent organ failure and multiple organ dysfunction syndrome. Furthermore, an early logistic regression prediction model comprising white blood cells, calcium, lymphocytes, albumin, platelets, and hematocrit was developed. This model achieved high macro-average AUCs in the development (0.924) and external validation (0.933) cohorts. Pragmatic validation in the local cohort further confirmed its utility in effectively identifying high-risk subphenotypes. To enhance interpretability and clinical utility, SHAP analysis was applied, and the model was deployed as a free web-based calculator and a nomogram specifically designed for the C4 subphenotype. This study identified four novel dynamic AP-HTG subphenotypes and developed a multicenter-validated early prediction model using admission-day indicators. It facilitates early risk stratification and individualized clinical decision-making.