Predictive value of dynamic creatinine cumulative exposure and trajectory classification onmortality in severe acute pancreatitis: a multicenter retrospective cohort study.

Abstract

This study aims to evaluate the predictive value of cumulative creatinine exposure (CumCr) and dynamic creatinine trajectories for severe acute pancreatitis (SAP) prognosis and construct a machine learning-based risk stratification model. An international multicenter retrospective cohort study included SAP patients from the Nanchang cohort (n = 1,545) and the MIMIC-IV database (n = 530). CumCr during the first 7 days of hospitalization was calculated. Latent class growth modeling (LCGM) identified creatinine trajectory patterns, and restricted cubic spline (RCS) analysis explored non-linear relationships between CumCr and mortality. The Boruta algorithm screened variables, and LASSO regression construct a nomogram model. A non-linear association between CumCr and mortality was observed, with a distinct threshold (K = 964.029 in the Nanchang cohort). Below this threshold, each standard deviation increases in CumCr elevated mortality risk 12-fold (OR = 12.135). LCGM classified four creatinine trajectory patterns. The persistently high-level group (PHL-T4) exhibited the highest mortality (43 % vs. 6.8 % in persistently low-level group (PLL-T1), P < 0.001). The nomogram integrated age, heart rate, calcium, and creatinine trajectories, achieving an AUC of 0.79, outperforming APACHE II (0.68) and SIRS (0.58). External validation yielded an AUC of 0.71. This study combining CumCr and trajectory modeling demonstrates that dynamic creatinine monitoring enhances early identification of high-risk SAP patients.