Large Language Models as Emerging Tools to Support Patient Education and Informed Consent for Uterine Artery Embolization

Purpose: To explore the role of artificial intelligence (AI) in supporting and improving patient education and informed consent in radiology; To evaluate the accuracy and quality of information provided by large language models (LLMs), taking uterine artery embolization (UAE) as a representative procedure.

Methods/Materials: 3 accessible LLMs, Chat GPT-5 (OpenAI), Gemini (Google), and Claude Opus 4.1 (Anthropic), were given 10 questions about UAE, focusing on the procedure, benefits, risks, alternatives, fertility, recovery, and long-term outcomes. Responses were anonymized and randomized. 5 interventional radiologists and 2 minimally invasive gynecologic surgeons independently evaluated responses using a Likert scale (1-5) across domains of accuracy, comprehensiveness, readability, compassion, and overall quality. Performance within domains was characterized with one-way ANOVA.

Results: A total of 1,050 ratings were collected. Significant differences between LLMs were observed for accuracy(p=0.0012) and comprehensiveness(p=0.0003), with GPT-5(3.86, 3.91) and Claude(3.80, 3.91) outperforming Gemini(3.42, 3.44). Overall quality differed significantly(p=0.0188), with GPT-5(3.80) outperforming Claude(3.64) and Gemini(3.44). GPT-5 demonstrated the strongest performance, Claude performed comparably in comprehensiveness but lower in quality, and Gemini scored lowest within most domains.

Conclusions: LLMs demonstrate potential to support patient education and informed consent for UAE, producing accurate, readable, and comprehensive information outside clinical encounters. However, specific limitations were noted, such as limited discussion of radial access, limited GnRH antagonists information, insufficient guidance on fertility considerations, and post-procedural red flags warranting emergent care. When used in an AI-assisted, physician-supervised framework, LLMs may help deliver consistent information, empower patients to engage more in shared decision-making, and reduce strain on healthcare systems by addressing patient questions that arise outside clinical encounters. Future prospective studies incorporating physician-supervised patient utilization of LLMs are needed to optimize patient-centered implementation of AI.