Benchmark analysis of myopia-related issues using large language models: a comparison of ChatGPT-4o and deepseek

This study evaluated the accuracy and comprehensiveness of responses generated by ChatGPT-4o and DeepSeek regarding commonly asked questions about myopia.

Thirty myopia-related questions spanning six clinical domains were submitted to both chatbots. Three medical professionals independently rated each response for accuracy and comprehensiveness. Inter-rater reliability was assessed using Fleiss' Kappa, and Shapiro-Wilk tests were conducted to examine normality in rating distributions. Statistical comparisons were performed using the Chi-square test, with significance set at p < 0.05.

DeepSeek outperformed ChatGPT-4o in overall accuracy, with significantly more responses rated as "Good" (p < 0.0001). Both models demonstrated high comprehensiveness scores when accuracy was rated "Good," though performance declined in treatment-related queries, particularly regarding commercial products like DIMS lenses. Fleiss' Kappa values indicated poor inter-rater agreement (DeepSeek: [Formula: see text] = 0.106; ChatGPT-4o: [Formula: see text] = - 0.0221), and normality tests showed non-normal score distributions (p < 0.0001 across domains).

Both ChatGPT-4o and DeepSeek can deliver useful responses to myopia-related questions, though limitations remain in areas requiring up-to-date, region-specific treatment information. DeepSeek's stronger performance suggests that localized LLMs may offer competitive advantages. Ongoing refinement, regular data updates, and domain-specific fine-tuning are essential for improving the reliability of AI chatbots in clinical communication.