Visual-textual integration in LLMs for medical diagnosis: A preliminary quantitative analysis

Background and aim: Visual data from images is essential for many medical diagnoses. This study evaluates the performance of multimodal Large Language Models (LLMs) in integrating textual and visual information for diagnostic purposes. Methods: We tested GPT-4o and Claude Sonnet 3.5 on 120 clinical vignettes with and without accompanying images. Each vignette included patient demographics, a chief concern, and relevant medical history. Vignettes were paired with either clinical or radiological images from two sources: 100 images from the OPENi database and 20 images from recent NEJM challenges, ensuring they were not in the LLMs' training sets. Three primary care physicians served as a human benchmark. We analyzed diagnostic accuracy and the models' explanations for a subset of cases. Results: LLMs outperformed physicians in text-only scenarios (GPT-4o: 70.8 %, Claude Sonnet 3.5: 59.5 %, Physicians: 39.5 %, p < 0.001, Bonferroni-adjusted). With image integration, all improved, but physicians showed the largest gain (GPT-4o: 84.5 %, p < 0.001; Claude Sonnet 3.5: 67.3 %, p = 0.060; Physicians: 78.8 %, p < 0.001, all Bonferroni-adjusted). LLMs altered their explanatory reasoning in 45-60 % of cases when images were provided. Conclusion: Multimodal LLMs showed higher diagnostic accuracy than physicians in text-only scenarios, even in cases designed to require visual interpretation, suggesting that while images can enhance diagnostic accuracy, they may not be essential in every instance. Although adding images further improved LLM performance, the magnitude of this improvement was smaller than that observed in physicians. These findings suggest that enhanced visual data processing may be needed for LLMs to achieve the degree of image-related performance gains seen in human examiners.