Multiphasic CT radiomics–clinical model for predicting acute pancreatitis recurrence: a multicenter survival analysis

Abstract Objectives To develop and externally validate a predictive model integrating multiphasic CT radiomics and clinical variables to estimate recurrence risk after first-episode acute pancreatitis (AP). Materials and methods This multicenter retrospective study included 548 patients with first-episode AP (training, n = 272; internal validation, n = 117; external validation, n = 159). Clinical and multiphasic CT data were collected, with recurrence as the primary endpoint. After stability-based feature selection, clinical, radiomics, and combined Cox models were constructed. Performance was evaluated using the C -index and time-dependent AUC, with SHAP for interpretability. Results A total of 548 patients were included (median age 42 years; 363 men). Recurrence rates were similar across cohorts (30.1%, 29.9%, and 28.3%; p = 0.702). The combined model achieved C-indices of 0.783 and 0.775 in the internal and external validation cohorts, respectively, numerically higher than those of the clinical model (0.748 and 0.750) and radiomics model (0.647 and 0.619), although the incremental improvement over the clinical model was not statistically significant. In the external validation cohort, time-dependent AUCs remained acceptable through 36 months but declined at 48 months. High-risk patients had lower 3-year recurrence-free survival than low-risk patients (37.4% vs 86.8%; log-rank p < 0.001). SHAP analysis identified the NC-3mm-Radscore as the most influential imaging-derived variable, whereas etiology-targeted prevention was the dominant protective factor. Conclusions The combined model demonstrated acceptable discrimination for predicting recurrence within the first 24 months after first-episode AP and may facilitate individualized follow-up and secondary prevention strategies. Key Points Question Can a multiphasic CT radiomics–clinical model improve prediction of AP recurrence over clinical factors alone in a multicenter setting? Findings The combined model demonstrated numerically higher predictive performance than the clinical model, though the incremental improvement was not statistically significant. Critical relevance statement This study critically evaluates the added value of multiphasic radiomics for recurrence prediction in AP, demonstrating that while radiomics captures baseline microenvironmental vulnerability, post-discharge clinical interventions remain the dominant dynamic drivers.