CT-based machine learning and multimodal prediction in acute pancreatitis: a narrative review.
Authors
Affiliations (2)
Affiliations (2)
- School of Medical and Life Sciences, Chengdu University of Traditional Chinese Medicine, Chengdu, China.
- Deyang People's Hospital, Deyang, China. [email protected].
Abstract
Acute pancreatitis (AP) has a highly variable clinical course. This narrative review focuses on computed tomography (CT) radiomics, deep learning and clinical-imaging models for severity and adverse-outcome prediction, with clinical-only models as context. A targeted PubMed search checked on 16 October 2026, with a publication cutoff of 14 October 2026, retrieved 80 unique records; ten CT-informed exemplars and eight clinical-context studies were selected for detailed comparison. This is a representative, non-exhaustive synthesis. For severe AP (SAP) prediction, reported same-center internal or temporal areas under the receiver operating characteristic curve (AUCs) ranged from 0.852 to 0.894, and reported external-test AUCs from 0.887 to 0.916; one validation cohort of unconfirmed independence had an AUC of 0.920. These descriptive ranges do not establish model superiority or characterize all retrieved studies. CT protocols and input timing, endpoint definitions, sampling and validation designs varied. Predominantly retrospective evidence, uncertain prediction landmarks, incomplete calibration and threshold reporting, and limited prospective impact assessment constrain translation. Feature explanations describe model associations rather than causal intervention targets. Imaging artificial intelligence (AI) remains an adjunct to clinical assessment; independent validation, calibration, decision-curve appraisal and prospective workflow/outcome evaluation are needed before clinical-pathway integration.