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

This study evaluated expert panel agreement with responses generated by Claude, Gemini and ChatGPT to clinical questions based on the supplied EAU guidelines for NMIBC, MIBC and UTUC. Using a modified Delphi methodology, multidisciplinary expert panels independently assessed the degree of agreement with each response. Gemini and ChatGPT responses frequently achieved consensus in agreement, whereas the corresponding proportion was lower for Claude, particularly in UTUC. The consistently high levels of expert panel agreement observed for Gemini and ChatGPT across the three clinical entities support the potential value of these models for guideline interpretation. These findings should nevertheless be interpreted within the study design, which assessed expert agreement rather than independently establishing factual accuracy, clinical safety or real-world clinical utility.

LLMs are being more frequently used as decision-support tools in oncology [8,18]. In UC management, LLMs and AI-driven tools may support a wide range of healthcare professionals beyond clinicians, including radiologists, pathologists, nurses and healthcare administrators. These technologies may assist in critical tasks such as diagnostic imaging interpretation, detailed pathology report analysis, personalized treatment planning, patient monitoring and enhancing communication across multidisciplinary teams [13,19–21]. Although these technologies demonstrate significant potential to enhance evidence-based decision-making, their clinical accuracy and reliability still vary, particularly in complex or changing situations, emphasizing the need for caution before implementing them in patient care [8,22].

In our study evaluating responses generated by LLMs to guideline-based questions on NMIBC, MIBC and UTUC, expert panel agreement varied across models and clinical entities. For NMIBC and MIBC, overall agreement among experts was high, generally above 88% for all LLMs, and there were no significant pairwise differences among models. In contrast, in the context of UTUC, Claude’s consensus in agreement rate was substantially lower (56%, 14/25 responses), compared with nearly complete consensus in agreement for Gemini and ChatGPT (100%, 25/25 and 96%, 24/25 responses, respectively). Panelist feedback frequently cited incomplete explanations or a lack of sufficient clinical nuance in specific responses, especially for more complex or atypical clinical scenarios. Moreover, the proportion of responses evaluated with a median score of 7 or higher out of 9 was notably lower for Claude (68%, 17/25 responses) than for Gemini (100%, 25/25 responses) and ChatGPT (96%, 24/25 responses). These findings indicate variation in expert panel agreement but do not establish that the responses were factually correct, complete, clinically actionable or safe. Qualitative feedback also included comments that some Claude responses were insufficiently focused, too brief or overly elaborate, suggesting that response style and level of detail, in addition to factual concerns, may have influenced panelist ratings. However, because the comments were not collected using predefined categories, we cannot reliably determine whether these patterns reflected overly cautious or hedged responses rather than differences in the substantive content of the responses.

A notable challenge in evaluating LLMs in medical contexts is the lack of standardized methodology across studies [8]. As recent literature highlights, evaluation approaches vary widely, from multiple-choice assessments and open-ended clinical prompts to expert ratings and consensus methods, making direct comparisons difficult [8]. In our current study, we applied a Delphi methodology involving a multidisciplinary expert panel including urologists, medical oncologists, radiation oncologists, nuclear medicine physicians, pathologists, radiologists and hospital pharmacists, enabling structured, iterative consensus building and rich qualitative feedback. The number and background of experts were similar to our previous study, which used a simpler direct rating system with experts independently scoring responses on a 5-point Likert scale [14]. Interestingly, the previous study found that rephrased prompts did not consistently improve expert ratings and that simple prompts received higher ratings in most comparisons. Claude with simple prompts achieved the highest overall proportion of optimal ratings [14]. In contrast, the current study used only simple prompts and observed that Gemini and ChatGPT, based on later model versions released in 2025, had higher proportions of responses achieving consensus in agreement and median scores ≥7 than Claude, whose version dates from late 2024. Because the studies evaluated different model versions at different time points, this reversal should not be interpreted as a stable ranking of vendors or model families. It may reflect version-specific updates together with differences in evaluation methodology, and confirmation would require contemporaneous testing under identical prompting conditions with repeated sampling. Additionally, the limited and less detailed clinical evidence for rarer tumors such as UTUC [5] may also contribute to variation in expert panel agreement. Together, these findings underscore the importance of rigorous multidisciplinary evaluation and continuous validation when interpreting expert ratings of responses generated by rapidly evolving LLMs.

Several recent studies provide important context for interpreting our findings in both general and urological oncology. Longwell et al. evaluated multiple LLMs with standardized oncology society exam questions, finding the top proprietary LLM answered 85% correctly but with some clinically harmful errors, highlighting the need for expert oversight [11]. Eckrich et al. compared LLM output to urology consultants using case-based questions, noting that while LLMs generated coherent answers, they had lower medical adequacy and a higher rate of potentially hazardous misinformation than experts [10]. Similarly, ‘t Hoen et al. assessed LLMs’ ability to update pediatric urology guidelines and found that, despite scientific accuracy in many cases, consistency and reliability levels were insufficient to replace expert review [12]. These findings collectively emphasize the promise of LLMs in oncology while underscoring the necessity of expert supervision and continuous evaluation before clinical deployment. In the specific context of BC management, recent work by Li et al. comprehensively evaluated six leading LLMs on clinical questions spanning diagnosis, treatment, prognosis and follow-up. Claude-3.5-Sonnet demonstrated the highest baseline accuracy, while iterative fine-tuning enabled GPT-3.5-Turbo to reach perfect accuracy [13]. These results highlight both the potential and current limitations of LLMs in providing reliable, guideline-concordant support, emphasizing the ongoing need for expert validation, sophisticated training methods, and integration of continuous clinical feedback to optimize AI-driven decision support in UC care.

This study has several limitations that should be acknowledged when interpreting its findings. It focused on three urothelial carcinoma subtypes (NMIBC, MIBC and UTUC) using a relatively limited set of 25 predefined clinical questions per clinical entity, which may reduce the generalizability of results to other clinical scenarios or cancers. The evaluation employed a two-round Delphi process with multidisciplinary expert panels including urologists, medical oncologists, radiation oncologists, nuclear medicine physicians, pathologists, radiologists and hospital pharmacists. While this approach facilitates structured consensus building and captures rich qualitative feedback, it inherently involves some subjectivity and potential bias related to panel composition and individual interpretations. Because members of the specialized tumor boards contributed to question development and subsequently evaluated the LLM responses, this overlap of roles may have introduced incorporation bias. However, panelists were blinded to model identity, and the plain language descriptions of figures and tables were reviewed by independent tumor board members. In some questions, especially those with greater specialty specificity, only a subset of the panel voted, and consensus was calculated exclusively among those who responded. While this increases content expertise for each item, it may introduce potential bias due to variable sample sizes across questions, which should be considered when interpreting the results. The limited number of questions and panelists, together with variable voting participation, may also have reduced statistical power, particularly for pairwise model comparisons. Therefore, the absence of statistically significant differences should not be interpreted as evidence of equivalence between models. At the same time, the consistently high expert panel agreement observed for Gemini and ChatGPT provides descriptive evidence of similarly high levels of agreement for these two models under the specific study conditions. Furthermore, rapid and frequent updates of LLM architectures mean that model performance may vary over time. The evaluated models were also released at different times, from October 2024 to May 2025. Consequently, the observed differences should be interpreted as comparisons between the specific versions tested and cannot be attributed solely to vendors or assumed to represent stable differences between model families. Only one response was generated per question and model, so reproducibility across repeated runs could not be assessed, while the use of English guideline materials with Spanish questions may limit generalizability to other language settings. Because the models were provided with the relevant EAU guideline content, including figures and tables converted into plain language textual summaries, the study primarily evaluated their ability to interpret and apply supplied guideline information under controlled conditions. The study did not include a predefined reference standard for independently assessing factual accuracy, completeness, clinical actionability or safety, and the qualitative comments were not collected using standardized categories that would permit reliable retrospective quantification of these outcomes. Therefore, consensus in agreement should not be interpreted as demonstrating that a response was fully correct, complete, clinically actionable or safe, or as evidence of performance in real clinical settings. Nevertheless, the structured preparation of the guideline content, combined with the multidisciplinary Delphi evaluation, strengthens the consistency and clinical relevance of the assessment. Continued evaluation and cautious interpretation of LLM outputs remain essential as these models evolve.

In conclusion, when provided with the 2025 EAU guidelines, including figures and tables converted into plain language textual summaries, the evaluated LLMs frequently generated responses that achieved multidisciplinary expert panel agreement. Agreement was particularly high for Gemini and ChatGPT across the three urothelial carcinoma settings evaluated, supporting the potential value of these models for guideline interpretation. Although the limited number of questions and variable panel participation warrant caution when interpreting model-to-model comparisons, the consistently high expert panel agreement observed for Gemini and ChatGPT provides a strong rationale for further evaluation using broader question sets and in real-world clinical settings. Such studies are needed to establish the clinical utility and safety of these tools as clinical decision-support systems.

A Borque-Fernando led the project and contributed to the study's conceptualization, methodology, project administration and supervision. All authors contributed jointly to the investigation, formal analysis, data curation, validation, resources, visualization, writing of the original draft, and review and editing of the manuscript, and approved the final version for publication.

The authors thank Dr Pablo Rivas for his assistance with medical writing on behalf of Content Ed Net.

This project was funded by the Spanish Association of Urology. The authors received support from Instituto de Investigación Sanitaria Aragón (Spain) for the project “Uro-Onco GuÍAs: Desarrollo de un Chatbot basado en IA Generativa para la Interpretación de Guías Clínicas en Uro-Oncología”.

The authors have no competing interests or relevant affiliations with any organization or entity with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Medical writing of the manuscript through Content Ed Net was funded by the Spanish Association of Urology.

This study involved only expert panelists who participated voluntarily in the Delphi process. No patient data were used. Participation implied informed consent, and formal ethical approval was not required.

Participation by expert panelists was voluntary and implied informed consent.

The results of the Delphi process are presented within the manuscript. The complete set of responses generated by each large language model to all 75 questions is available for consultation at https://www.contentednet.com/proyecto-urooncoguias-aeu/.

This work is licensed under the Attribution-NonCommercial-NoDerivatives 4.0 Unported License. To view a copy of this license, visit https://creativecommons.org/licenses/by-nc-nd/4.0/