Evaluating Large Language Models in Pediatric Fever Management: A two-layer study

Guijun Yang¹Hejun Jiang¹Shuhua Yuan¹Mingyu Tang¹Jing Zhang¹Jilei Lin¹Jiande Chen¹Jiajun Yuan²Liebin Zhao³Yin Yong^1*

• ¹Department of Respiratory Medicine, Shanghai Children's Medical Center, Shanghai, China
• ²Medical Department, Shanghai Children's Medical Center, Shanghai, China
• ³Pediatric AI Clinical Application and Research Center, Shanghai Children's Medical Center, Shanghai, China

Background: Pediatric fever is a prevalent concern, often causing parental anxiety and frequent medical consultations. While large language models (LLMs) such as ChatGPT, Perplexity, and YouChat show promise in enhancing medical communication and education, their efficacy in addressing complex pediatric fever-related questions remains underexplored, particularly from the perspectives of medical professionals and patients' relatives. Objective: This study aimed to explore the differences and similarities among four common large language models (ChatGPT3.5, ChatGPT4.0, YouChat, and Perplexity) in answering thirty pediatric fever-related questions and to examine how doctors and pediatric patients' relatives evaluate the LLM-generated answers based on predefined criteria.to examine the differing perspectives of doctors and pediatric patients' relatives on these responses. Methods: The study selected thirty fever-related pediatric questions answered by the four models. Twenty doctors rated these responses across four dimensions. To conduct the survey among pediatric patients' relatives, we eliminated certain responses that we deemed to pose safety risks or be misleading. Based on the doctors' questionnaire, the thirty questions were divided into six groups, each evaluated by twenty pediatric relatives. The Tukey post-hoc test was used to check for significant differences. Some of pediatric relatives was revisited for deeper insights into the results. Results: In the doctors' questionnaire, ChatGPT3.5 and ChatGPT4.0 outperformed YouChat and Perplexity in all dimensions, with no significant difference between ChatGPT3.5 and ChatGPT4.0 or between YouChat and Perplexity. All models scored significantly better in accuracy than other dimensions. In the pediatric relatives' questionnaire, no significant differences were found among the models, with revisits revealing some reasons for these results. Conclusions: Internet searches (YouChat and Perplexity) did not improve the ability of large language models to answer medical questions as expected. Patients lacked the ability to understand and analyze model responses due to a lack of professional knowledge and a lack of central points in model answers. When developing large language models for patient use, it's important to highlight the central points of the answers and ensure they are easily understandable.