Incorporating information retrieval into AI chatbots for patient education on thyroid eye disease.

PURPOSE: To evaluate the performance of general-purpose, retrieval-augmented, and medicine-specific AI chatbots in answering common thyroid eye disease (TED) patient questions.

DESIGN

Cross-sectional comparative evaluation. Online TED forum discussion posts were collected and synthesized into 15 representative patient questions across five groups spanning clinical (treatment, diagnosis, management, epidemiology) and non-clinical topics, grouped into three difficulty levels. Three differing large language models (LLMs) generated responses were randomized and anonymized for blinded assessment.

SUBJECTS, PARTICIPANTS, AND/OR CONTROLS: Three oculoplastic surgeons evaluated clinical metrics; 3 medical students assessed non-clinical metrics.

METHODS

Three AI models generated responses: GPT-4o-mini (ChatGPT), a retrieval-augmented generation model grounded in TED literature (ChatGPT-RAG), and a specially trained LLM for healthcare professionals (OpenEvidence). Blinded raters assessed randomized responses. Statistical analysis used paired Wilcoxon signed-rank tests with Hedges' g for effect sizes.

MAIN OUTCOMES MEASURED: Clinical evaluation of responses was conducted using a 7-point Likert scale for relevance, accuracy, balance, and scope. Non-clinical metrics of empathy, understandability, and readability were also assessed using validated tools.

RESULTS

OpenEvidence significantly outperformed both ChatGPT (mean clinical score 5.96 vs 4.94; Hedges' g = 1.21, P < 0.001) and ChatGPT-RAG (5.96 vs 5.55; g = 0.53, P < 0.001) in clinical rankings and across most clinical metrics, including accuracy and relevance. However, performance patterns reversed for non-clinical metrics, with ChatGPT consistently outperforming specialized models in empathy, understandability, and actionability (18.4 vs 14.96 for OpenEvidence; g = 1.25, P < 0.001). Across both domains, ChatGPT-RAG achieved intermediate performance, more closely trailing OpenEvidence clinically (g = 0.53) and ChatGPT with respect to non-clinical metrics (g = 0.44). Limitations include a modest sample of raters and synthesized questions from online forums, which may affect generalizability.

CONCLUSIONS

Specialized medical AI models may have better clinical accuracy, while general-purpose models may outperform in patient communication and accessibility. The development of retrieval augmented generation-based approaches combining clinical precision with effective communication represents a promising direction for AI-powered patient education in TED and, potentially, other complex conditions.