Abstract / Summary
Abstract Patients considering robot-assisted radical prostatectomy (RARP) require information that spans diagnosis, treatment selection, functional recovery, and recurrence surveillance. We compared the quality of English-language RARP patient education generated by ChatGPT, DeepSeek, and Doubao. Twenty-five questions covering screening, diagnosis, treatment decision-making, perioperative care, rehabilitation, and follow-up were submitted individually to each model, yielding 75 responses. Three blinded clinicians evaluated the responses using the Global Quality Score, modified DISCERN, Quality Analysis of Medical Artificial Intelligence, Ensuring Quality Information for Patients, and the Patient Education Materials Assessment Tool; six readability indices were also calculated. Overall Global Quality Scores did not differ significantly across models. ChatGPT achieved a higher EQIP score (median, 75.00%) than DeepSeek and Doubao (65.83% each) and greater PEMAT actionability (59.00% vs. 54.33% and 47.67%, respectively). DeepSeek generated comparatively easier-to-read text, although all three models exceeded commonly recommended sixth- to eighth-grade reading levels for public-facing health materials. Under the majority-rater criterion, no factual errors or outdated recommendations were identified, whereas key omissions occurred mainly in DeepSeek and Doubao responses. Inter-rater agreement was low, indicating substantial uncertainty in multidimensional human assessment. LLMs may support the preparation of clinician-reviewed RARP educational materials but should not replace individualized counseling by urologists.