Yingbin Zheng ^1* Yiwei Yan ¹ Sai Chen ² Yunping Cai ² Kun Ren ² Yishan Liu ³ Jiaying Zhuang ¹ Min ZHAO ^1*

• ¹ Biomedical Big Data Center, First Affiliated Hospital of Xiamen University, Xiamen, Fujian Province, China
• ² Meteorological Disaster Prevention Technology Center, Xiamen Meteorological Bureau, Xiamen, China
• ³ Software Engineering, School of Software, Taiyuan University of Technology, Taiyuan, Shanxi Province, China

Background: Web-based medical services have significantly improved access to healthcare by enabling remote consultations, streamlining scheduling, and improving access to medical information. However, providing personalized physician recommendations remains a challenge, often relying on manual triage by schedulers, which can be limited by scalability and availability. Objective: This study aimed to develop and validate a Retrieval-Augmented Generation-Based Physician Recommendation (RAGPR) model for better triage performance. Methods: This study utilizes a comprehensive dataset consisting of 646,383 consultation records from the Internet Hospital of the First Affiliated Hospital of Xiamen University. The research primarily evaluates the performance of various embedding models, including FastText, SBERT, and OpenAI, for the purposes of clustering and classifying medical condition labels. Additionally, the study assesses the effectiveness of large language models (LLMs) by comparing Mistral, GPT-4o-mini, and GPT-4o. Furthermore, the study includes the participation of three triage staff members who contributed to the evaluation of the efficiency of the RAGPR model through questionnaires. Results: The results of the study highlight the different performance levels of different models in text embedding tasks. FastText has an F1-score of 46%, while the SBERT and OpenAI significantly outperform it, achieving F1-scores of 95% and 96%, respectively. The analysis highlights the effectiveness of LLMs, with GPT-4o achieving the highest F1-score of 95%, followed by Mistral and GPT-4o-mini with F1-scores of 94% and 92%, respectively. In addition, the performance ratings for the models are as follows: Mistral with 4.56, GPT-4o-mini with 4.45 and GPT-4o with 4.67. Among these, SBERT and Mistral are identified as the optimal choices due to their balanced performance, cost effectiveness, and ease of implementation. Conclusions: The RAGPR model can significantly improve the accuracy and personalization of web-based medical services, providing a scalable solution for improving patient-physician matching.