Combining MRI dual-sequence radiomics and clinical parameters predicts post-hypertriglyceridemic acute pancreatitis...

Post-acute pancreatitis diabetes mellitus (PPDM-A) represents the most prevalent subtype of diabetes secondary to exocrine pancreatic dysfunction. Patients with acute pancreatitis (AP) caused by hypertriglyceridemia (HTG-AP) face a high risk of PPDM-A. Since clinical predictors alone lack accuracy, integrating magnetic resonance imaging (MRI)-based radiomics may improve prognostic risk stratification.

A retrospective cohort of 210 patients with HTG-AP was included and randomized into training and internal testing cohorts (n = 147 and 63, respectively; ratio 7:3). An independent external validation cohort (n = 119) from a separate hospital campus was also analyzed. Radiomics features from T2-weighted and late arterial phase contrast-enhanced T1-weighted MRI were selected via least absolute shrinkage and selection operator (LASSO) to generate a radiomics score (Rad-score). A random forest model that incorporated the Rad-score and independently significant clinical predictors was established. Model predictive ability was examined using receiver operating characteristic (ROC) curves, decision curve analysis (DCA), and reclassification metrics, including integrated discrimination improvement (IDI) and net reclassification improvement (NRI). SHapley Additive exPlanations (SHAP) analysis was employed to interpret feature contributions.

Seven optimal radiomics features were used to generate the Rad-score. The final combined model incorporated this score with three key clinical variables and achieved areas under the ROC curve (AUCs) of 0.905, 0.904, and 0.900 in the training, testing, and external validation cohorts, respectively, significantly outperforming single-modality models. SHAP analysis identified the Rad-score, length of hospital stay, high-sensitivity C-reactive protein, and recurrence of AP as principal predictive contributors.

Integrating dual-sequence MRI radiomics with clinical features accurately predicts HTG-PPDM-A risk. Enhanced by SHAP interpretability, this non-invasive tool enables transparent long-term risk prediction to guide personalized interventions.