Assessment of Differential Diagnoses for Oculoplastics Cases Produced by Large Language Models.
Academic Article
Overview
abstract
PURPOSE: This study aimed to evaluate the accuracy of different large language models (LLMs) in generating differential diagnoses for oculoplastic cases. METHODS: Differential diagnoses were generated for 20 oculoplastic cases sourced from University of Iowa EyeRounds.org using 6 LLMs: Chat Generative Pre-Trained Transformer (ChatGPT) 3.5, ChatGPT 4.0, OcuSmart/EyeGPT, Google Gemini 1.5, Claude 3.5, and Microsoft CoPilot. Outputs were compared against the EyeRounds expert-curated differentials examining (1) top diagnosis match rate (2) inclusion of the correct diagnosis within the generated list, as well as (3) recall and (4) precision, calculated to assess the overlap and conciseness of LLM outputs. RESULTS: OcuSmart/EyeGPT achieved the highest top diagnosis match rate (85 ± 36%), while Claude 3.5 demonstrated the highest rate of inclusion of correct diagnosis in differential, as well as recall rate (100 ± 0% and 55 ± 27%, respectively). Google Gemini produced the most precise differentials at 43 ± 24%. Claude 3.5 generated the largest but least concise lists. LLM performance varied by case; for example, idiopathic orbital inflammation cases yielded highest recall and top diagnosis match across all models, while floppy eyelid syndrome cases demonstrated lowest performance. CONCLUSIONS: LLMs show promising potential in diagnosing oculoplastic cases, with OcuSmart/EyeGPT and Claude 3.5 performing best for determining the case diagnosis and recall, and ChatGPT 3.5, OcuSmart/EyeGPT, and Gemini generating concise and relevant differentials. However, further research and development are necessary to validate LLMs' capabilities and integrate them into the clinical workflow.
