Guideline Concordance of Large Language Models in the Management of Ureteral Stones: A Clinical Vignette-Based Comparative Study

Vollständiger Abstract

Worum geht es in dieser Arbeit?

ObjectiveLarge language models (LLMs) are increasingly used to answer medical questions, but their reliability in guideline-based urological decisions remains uncertain. This study aimed to evaluate the concordance of three widely used LLMs with the European Association of Urology (EAU) 2026 guideline recommendations for the active management of ureteral stones.Materials and MethodsIn this cross-sectional, vignette-based comparative study, the EAU 2026 ureteral-stone treatment algorithm was converted into 40 standardized clinical vignettes (four groups of ten: proximal 10 mm, distal 10 mm). ChatGPT, Gemini, and Claude were queried with the same standardized prompt, which did not name a specific guideline. Responses were scored against predefined EAU-based reference answers using a binary system. Concordance was compared with Cochran’s Q test.ResultsA total of 120 LLM-generated responses were evaluated. Overall concordance was 96.7% (116/120). ChatGPT achieved complete concordance (40/40, 100%), while Gemini and Claude each achieved 95% (38/40); the difference was not significant (Cochran’s Q=2.67, p=0.264). Concordance was complete in all 10 mm scenarios requiring URS prioritization. All four discordances were “incorrect prioritization” in >10 mm stones, presenting shock-wave lithotripsy as co-equal to ureteroscopy; each involved cross-guideline conflation with American Urological Association (AUA) framing. No unsafe recommendation or guideline hallucination was observed under the predefined scoring categories.ConclusionThe evaluated LLMs showed high concordance with EAU 2026 first-line treatment recommendations for ureteral stones. Concordance was complete in non–priority-sensitive