Assessing the proficiency of large language models on funduscopic disease knowledge

Abstract

AIM: To assess the performance of five distinct large language models (LLMs; ChatGPT-3.5, ChatGPT-4, PaLM2, Claude 2, and SenseNova) in comparison to two human cohorts (a group of funduscopic disease experts and a group of ophthalmologists) on the specialized subject of funduscopic disease. METHODS: Five distinct LLMs and two distinct human groups independently completed a 100-item funduscopic disease test. The performance of these entities was assessed by comparing their average scores, response stability, and answer confidence, thereby establishing a basis for evaluation. RESULTS: Among all the LLMs, ChatGPT-4 and PaLM2 exhibited the most substantial average correlation. Additionally, ChatGPT-4 achieved the highest average score and demonstrated the utmost confidence during the exam. In comparison to human cohorts, ChatGPT-4 exhibited comparable performance to ophthalmologists, albeit falling short of the expertise demonstrated by funduscopic disease specialists. CONCLUSION: The study provides evidence of the exceptional performance of ChatGPT-4 in the domain of funduscopic disease. With continued enhancements, validated LLMs have the potential to yield unforeseen advantages in enhancing healthcare for both patients and physicians.