Evaluation of artificial intelligence-based patient education models for irritable bowel syndrome.

Abstract

Background: Irritable bowel syndrome (IBS) is a common functional gastrointestinal disorder with a significant psycho-social burden. Despite medical advancements, patient education on IBS remains inadequate. This study compared two large language models (LLMs)-ChatGPT-4 and Gemini-1-for their performance in addressing IBS-related patient queries.

Methods: Thirty-nine IBS-related frequently asked questions (FAQs) from IBS organizations and hospital websites were categorized into six domains: general understanding, symptoms and diagnosis, causes, dietary considerations, treatment and lifestyle factors. Responses from ChatGPT-4 and Gemini-1 were evaluated by two independent gastroenterologists for comprehensiveness and accuracy, with a third reviewer resolving disagreements. Readability was measured using five standardized indices (Flesch Reading Ease [FRE], Simple Measure of Gobbledygook [SMOG], Gunning Fog Index [GFI], Automated Readability Index [ARI], Reading Level Consensus [ARC]) and empathy was rated on a 4-point Likert scale by three reviewers.

Results: Gemini produced comprehensive and accurate answers for 94.9% (37/39) of questions, with two rated as mixed (vague/outdated). ChatGPT achieved 89.7% (35/39) comprehensive responses, with four rated mixed. Domain-wise, both models performed best in "symptoms and diagnosis" and "treatment", while mixed responses were most frequent in "general understanding" and "lifestyle". There was no significant difference in comprehensiveness (p = 0.67). Readability analysis showed both LLMs generated difficult-to-read content: Gemini's FRE score was 35.83 ± 3.31 vs. ChatGPT's 32.33 ± 5.57 (p = 0.21), corresponding to college-level proficiency. ChatGPT's responses were more empathetic, with all responses rated moderately empathetic; Gemini was mostly rated minimally empathetic (66.7%).

Conclusion: While ChatGPT and Gemini provided extensive information, their limitations-such as complex language and occasional inaccuracies-must be addressed. Future improvements should focus on enhancing readability, contextual relevance and accuracy to better meet the diverse needs of patients and clinicians.