Evaluating Large Language Models for Accuracy and Completeness of Vitiligo Patient Education: A Comparative Analysis.

BACKGROUND

Vitiligo causes significant psychological stress, creating a strong demand for accessible educational resources beyond clinical settings. This demand remains largely unmet. Large language models (LLMs) have the potential to bridge this gap by enhancing patient education. However, uncertainties exist regarding their ability to accurately address individualized patient inquiries and whether comprehension capabilities vary between LLMs.

PURPOSE

This study aims to evaluate the applicability, accuracy, and potential limitations of OpenAI o1, DeepSeek-R1, and Grok 3 for vitiligo patient education.

METHODS

Three dermatology experts first developed sixteen vitiligo-related questions based on common patient concerns, which were categorized as descriptive or recommendatory with basic and advanced levels. The responses from the three LLMs were then evaluated by three vitiligo-specialized dermatologists for accuracy, comprehensibility, and relevance using a Likert scale. Additionally, three patients rated the comprehensibility of the responses, and a readability analysis was performed.

RESULTS

All three LLMs demonstrated satisfactory accuracy, comprehensibility, and completeness, although their performance varied. They achieved 100% accuracy in responding to basic descriptive questions but exhibited inconsistency when addressing complex recommendatory queries, particularly regarding treatment recommendations for specific populations. Pairwise comparisons indicated that DeepSeek-R1 outperformed OpenAI o1 in accuracy scores (p = 0.042), while no significant difference was observed compared to Grok 3 (p = 0.157). Readability assessments revealed elevated reading difficulty across all models, with DeepSeek-R1 exhibiting the lowest readability (mean Flesch Reading Ease score of 19.7; pairwise comparisons showed DeepSeek-R1 scores were significantly lower than those of OpenAI o1 and Grok 3, both p < 0.01), potentially reducing accessibility for diverse patient populations.

CONCLUSION

Reasoning-LLMs demonstrate high accuracy in responding to simple vitiligo-related questions, but the quality of treatment recommendations declines as question complexity increases. Current models exhibit errors in providing vitiligo treatment advice, necessitating enhanced filtering mechanisms by developers and mandatory human oversight for medical decision-making.