Published predictor set and recurrent-event burden after acute pancreatitis: a retrospective single-center cohort study

Recurrent acute pancreatitis (RAP) is common after a first acute pancreatitis (AP) episode, but published RAP models are usually derivation studies and rarely address recurrent-event burden. In this retrospective single-center cohort, we enrolled consecutive adults whose index discharge after a first AP admission occurred between January 1, 2018, and October 31, 2021, and administratively truncated follow-up at 36 months after discharge through October 31, 2024. The primary endpoint was first post-discharge RAP within 36 months; secondary endpoints were total RAP episodes, and chronic pancreatitis (CP) progression was summarized descriptively. Mortality was recorded as follow-up disposition/censoring information. The four predictors reported by a published binary nomogram were fitted as a reproducible, source-informed local Cox comparator. No unmodified-source-model absolute risk was evaluated. A separate de novo local Cox model was internally validated with nested multiple-imputation bootstrap, and total event burden was analyzed with a Prentice–Williams–Peterson gap-time model. All 859 patients contributed follow-up through first RAP, death, last contact, or 36 months; 15 event-free patients were lost before 36 months and 14 died. First RAP was observed in 322 patients (crude proportion 37.5%), and 693 recurrent episodes were analyzed. The optimism-corrected Uno’s C was 0.69 (95% CI 0.66–0.72) for the source-informed four-variable local comparator and 0.74 (95% CI 0.72–0.77) for the de novo local Cox model; the latter had a 36-month Brier score of 0.174 (95% CI 0.158–0.190). Apparent risk strata separated cumulative recurrent-event burden. CP progression occurred in 79 patients (9.2%). In this retrospective single-center cohort, a locally fitted model using the published predictor set showed moderate internally validated discrimination, while a separate de novo local model showed higher internally validated performance. Recurrent-event methods added burden information beyond time to first recurrence; independent multicenter validation is required before clinical use.