Interpretable machine learning model based on non-contrast computed tomography within 24 hours of admission for early prediction of the risk of severe deterioration of acute pancreatitis.

Background: Early prediction of progression to severe acute pancreatitis (SAP) remains challenging due to the lack of reliable tools within 24 hours of admission. Conventional contrast-enhanced computed tomography (CT) is often unavailable in emergency settings, and non-contrast CT alone provides limited information for severity assessment. Therefore, this study aimed to construct an early prediction model for SAP risk using intrapancreatic and peripancreatic radiomics features from non-contrast CT obtained within 24 hours of admission, combined with clinical characteristics, to provide a theoretical basis for early clinical intervention. Methods: A retrospective analysis was conducted involving 585 patients admitted within 24 hours after acute pancreatitis (AP) onset. Patients were classified according to the Revised Atlanta Classification [2012] at 48 hours after admission, resulting in 148 cases progressing to SAP and 437 cases classified as mild to moderate acute pancreatitis (MAP/MSAP). ITK-SNAP software was used to manually outline the pancreas contour on non-contrast CT images obtained within 24 hours of admission for AP. Regions were expanded outward from the pancreatic contour by 2 mm, 4 mm, and 6 mm, respectively. Radiomics features were extracted and selected via intraclass correlation coefficient (ICC >0.75), Spearman correlation (>0.9), least absolute shrinkage and selection operator (LASSO) regression, and maximum relevance minimum redundancy (mRMR) algorithms. The optimal peripancreatic expansion distance was determined by comparing predictive performance. A combined radiomics-clinical model was built using the eXtreme Gradient Boosting (XGBoost) algorithm and evaluated by area under the receiver operating characteristic (ROC) curve (AUC), DeLong test, calibration curve, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) analysis was used for model interpretability. Results: The 4 mm peripancreatic region showed the best predictive performance. After feature selection, 35-47 robust features were retained across different regions of interest (ROIs). The combined model, incorporating intrapancreatic plus 4 mm peripancreatic radiomics features and clinical parameters [procalcitonin (PCT), white blood cell (WBC) count, calcium (Ca), lymphocyte count (LYM), albumin (ALB), lactate dehydrogenase (LDH)], yielded an AUC of 0.852 [95% confidence interval (CI): 0.782-0.921] in the validation cohort, which was significantly higher than the clinical model alone (AUC =0.807) and the radiomics model alone (AUC =0.829). The calibration curve showed good agreement between predicted and observed probabilities, and DCA demonstrated superior net benefit of the combined model across a threshold range of 0.1-0.6. SHAP analysis revealed that intrapancreatic heterogeneity and peripancreatic density were critical predictive factors. Both SHAP analysis and the developed nomogram demonstrated robust clinical interpretability. Conclusions: Radiomics features extracted from the 4 mm peripancreatic region on non-contrast CT scans obtained within 24 hours of admission, combined with clinical characteristics, effectively predict the risk of progression to SAP. This predictive model can support clinical evaluation and guide treatment decisions.