Large language and multimodal foundation models in endoscopic retrograde cholangiopancreatography

Abstract Artificial intelligence in endoscopic retrograde cholangiopancreatography (ERCP) has so far concentrated on image-based tasks. Large language models and multimodal foundation models could extend this work by integrating text, clinical data, endoscopic images, fluoroscopy, and procedural video within a single system. This review aimed to map the plausible roles of these models across the ERCP care pathway, separate direct ERCP evidence from reasoned extrapolation, and define the safeguards required before clinical use. A narrative review was conducted and reported in accordance with the Scale for the Assessment of Narrative Review Articles (SANRA). PubMed, Scopus, and Google Scholar were searched from database inception through April 30, 2026 using predefined strings that combined ERCP-related and foundation-model-related terms. A 2-tier eligibility framework distinguished direct ERCP evidence from transferable evidence in related clinical settings. All empirical studies were peer-reviewed, whereas guidelines, methodological standards, and regulatory documents were included separately for framework and governance context. The source and level of evidence were summarized in tables, and a literature-informed framework was developed for future evaluation. Few studies evaluated a large language model or multimodal foundation model specifically in ERCP or cholangioscopy; the remaining direct ERCP evidence comprised earlier task-specific perceptual models. Structured report drafting, adverse-event extraction, and patient communication emerged as priorities for prospective ERCP evaluation because their outputs are bounded and fully reviewable, but none is ready for clinical use. Indication appraisal and risk or prophylaxis prompting require foundational ERCP studies, whereas autonomous triage and multimodal procedural interpretation remain long-term research objectives. The principal limitations are hallucination, automation bias, privacy risk, domain shift, limited validation, and regulatory uncertainty. All proposed applications remain investigational. They should not inform ERCP clinical decision-making or routine care without prospective ERCP-specific validation, predefined safety thresholds, appropriate governance, and applicable regulatory authorization. The proposed checklist is a literature-informed framework for judging research readiness, not a recommendation for current deployment.