Evaluation of large language models for guideline-based clinical questions in muscle-invasive and non-muscle-invasive bladder cancer and upper tract urothelial carcinoma: a multidisciplinary Delphi study
Félix Guerrero-Ramos, Jesús María Fernández Gómez, Tomás Fernández Aparicio, Antonio López-Beltrán, Óscar Buisán Rueda, Luis Llanes González, Javier Puente Vázquez, Alberto Carrión, Ana Loizaga Iriarte, José Daniel Subiela Henríquez, Felipe Couñago, Javier Martínez Benavides, Paula Pelechano Gómez, María José Juan Fita, Richard Mast Vilaseca, Antoni Vilaseca, Luis Ladaria Sureda, Alberto Pérez-Lanzac de Lorca, José Luis Domínguez Escrig, Julián Sanz Ortega, Elena Sevillano Fernández, Ana Cristina Plata Bello, Felipe Villacampa Auba, Marta Trassierra Villa, Eloy Vivas Vargas, Isabel Trias Puigsureda, Ángel Borque-FernandoAim: To assess expert panel agreement with responses from three large language models, Claude, Gemini and ChatGPT, to questions based on guidelines for non-muscle-invasive bladder cancer (NMIBC), muscle-invasive bladder cancer (MIBC) and upper tract urothelial carcinoma (UTUC). Materials & methods: The 2025 European Association of Urology (EAU) guidelines were provided to Claude 3.5 Haiku, Gemini 2.5 Flash and GPT-4.1-mini. Figures and tables were converted into plain language text. Each model answered 25 questions per entity. Panels of 7–10 experts rated responses on a 9-point scale using a modified Delphi process with two rounds. Outcomes were consensus in agreement and median expert panel score ≥7. Results: For Claude, Gemini and ChatGPT, respectively, consensus in agreement was achieved for 88%, 100% and 96% of NMIBC responses (p = 0.097); 92%, 100% and 96% of MIBC responses (p = 0.223); and 56%, 100% and 96% of UTUC responses (p < 0.001). For UTUC, significantly fewer Claude responses achieved consensus in agreement than Gemini or ChatGPT responses. Across all questions, consensus in agreement was achieved for 59/75 Claude responses (78.7%), 75/75 Gemini responses (100%), and 72/75 ChatGPT responses (96%) (p < 0.001). Median scores ≥7 were achieved for 88%, 100% and 100% of NMIBC responses; 96%, 100% and 100% of MIBC responses; and 68%, 100% and 96% of UTUC responses. In UTUC, Claude had significantly fewer such responses than Gemini, but the difference from ChatGPT was not statistically significant. Overall, median scores ≥7 were achieved for 63/75 Claude responses (84%), 75/75 Gemini responses (100%) and 74/75 ChatGPT responses (98.7%) (p < 0.001). Conclusion: Under the study conditions, large language model responses frequently achieved expert panel agreement, particularly for Gemini and ChatGPT, supporting the potential of these models for guideline interpretation. Further validation is required to establish clinical utility and safety.