Electronic Ophthalmic Records Diagnosis using Role-specific Guided Large Language Model Based on Stylistic Differentiation

Laiyi Fu^1*Binbin Fan¹Hongkai Du¹Yanxiang Feng¹Chunhua Li²Huping Song^2*

• ¹Xi'an Jiaotong University, Xi'an, China
• ²Xi’an Fourth Hospital, xi'an, China

Ophthalmology consultations play a vital role in the diagnosis, treatment, and prevention of eye diseases. Nevertheless, the surging demand for consultations has outstripped the supply of ophthalmologists. In the era of digital health, electronic diagnosis and online treatment have emerged as potential solutions. By harnessing large pre-trained language models, we can design effective dialogues for specific ophthalmology scenarios, thereby facilitating consultations. However, conventional fine-tuning strategies for question-answering tasks have proven impractical due to the escalating model size, the frequent oversight of the patient-doctor role function during consultations, and the lack of exploration into the underlying mechanisms of ophthalmic diseases. In this paper, we propose EyeDoctor, an ophthalmic medical questioning large language model which augments its accuracy via doctor-patient role perception guidance and an enhanced knowledge base incorporating external disease information and electronic medical records using Retrieval-Augmented Generation (RAG). Specifically, the knowledge base incorporates curated genetic and disease-related information from PubMed to provide a deeper understanding of ophthalmic disease mechanisms and improve responses to medical queries. Experimental results show EyeDoctor achieves higher questionanswering precision in ophthalmology consultations. Notably, EyeDoctor demonstrated a 7.25% improvement in Rouge-1 scores and a 10.16% improvement in F1 scores on multi-round datasets compared to second best model ChatGPT, highlighting the importance of doctor-patient role differentiation, external knowledge integration, and a deeper focus on disease mechanism exploration. Moreover, EyeDoctor demonstrates promising clinical utility by enhancing the relevance and accuracy of ophthalmic consultation responses. It assists physicians in preliminary triage and symptom interpretation, helps reduce consultation time, and improves consistency in diagnostic support. These capabilities contribute to better alignment with real-world clinical practices and improved patient experience, particularly in resource-constrained settings. EyeDoc also serves as a free available web based service and source code is available at https://github.com/sperfu/EyeDoc.