Abstract
INTRODUCTION: Large language models are increasingly used in clinical decision-making, but their role in appropriateness-based indications is uncertain. METHODS: We conducted a vignette-based Italian survey comparing esophagogastroduodenoscopy appropriateness by 5 artificial intelligence (AI) (ChatGPT-4.0, ChatGPT-4.5, Gemini, Claude AI, OpenEvidence) at 2 times (April and September 2025) with gastroenterologists, residents, and general practitioners, using American Society for Gastrointestinal Endoscopy and European Society of Gastrointestinal Endoscopy guidelines. RESULTS: A total of 135 physicians participated. AI performance varied over time: accuracies ranged from 50%–90% in April to 63%–80% in September, with ChatGPT-4.5 and ChatGPT-4.0 outperforming physicians in September. DISCUSSION: Temporal and prompt variability highlight the need for multirun, longitudinal evaluation before clinical adoption.

This publication has 15 references indexed in Scilit: