Integrating 2.5D multi-modal CT imaging with serum triglycerides for early risk stratification in hypertriglyceridemia-induced acute pancreatitis: A multicenter study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Hepatopancreatobiliary, Splenic and Hernia Surgery, The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China, Changsha, Hunan, China.
- Department of Emergency Medicine, The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China, Changsha, Hunan, China.
- Department of Critical Care Medicine, Songgang People's Hospital, Baoan District, Shenzhen, China.
- Department of Emergency Medicine, The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China, Changsha, Hunan, China. Electronic address: [email protected].
- Department of Emergency Medicine, Changsha Traditional Chinese Medicine Hospital (Changsha Eighth Hospital), Changsha, Hunan, China. Electronic address: [email protected].
- Department of Ultrasound, The First Affiliated Hospital of Hunan University of Traditional Chinese Medicine, Changsha, Hunan, China. Electronic address: [email protected].
Abstract
To develop and externally validate a clinical-radiological framework that fuses 2.5D deep learning features from dual-phase computed tomography (CT) with serum triglyceride (TG) levels for early prediction of severe acute pancreatitis (SAP) in patients with hypertriglyceridemia-induced acute pancreatitis (HTGP). In this multicenter retrospective study, 1,518 consecutive HTGP patients from three hospitals were divided into a training cohort (n = 990) and two external validation cohorts (n = 340, n = 188). Patients underwent non-enhanced CT (NECT) and enhanced CT (ECT) within 24 h of admission. A 2.5D SE-ResNet-50 network, pre-trained with self-supervised learning, extracted deep features from NECT and ECT. Features were reduced by principal component analysis and entered into support vector machine classifiers; NECT and ECT probabilities were fused by weighted averaging. Clinical-radiological models further combined imaging probability with TG alone (Fusion-TG) or with TG, total cholesterol, comorbidities, and CT severity index. The fusion imaging model yielded area under the receiver operating characteristic curves (AUCs) of 0.966 in the training cohort and 0.923 and 0.890 externally. Fusion-TG and Fusion-Multivar achieved similar AUCs (0.970-0.971). Incorporating TG increased sensitivity for SAP (e.g., 60.6 % to 94.4 % in external cohort I) while preserving high specificity, and SHapley Additive exPlanations (SHAP) analysis highlighted ECT-derived features and TG as dominant contributors. The parsimonious Fusion-TG model enables robust, interpretable early SAP risk stratification in HTGP using only dual-phase CT and a single biochemical marker, supporting rapid decision-making in emergency care.