Clinical Management of Wasp Stings Using Large Language Models: Cross-Sectional Evaluation Study.

BACKGROUND

Wasp stings are a significant public health concern in many parts of the world, particularly in tropical and subtropical regions. The venom of wasps contains a variety of bioactive compounds that can lead to a wide range of clinical effects, from mild localized pain and swelling to severe, life-threatening allergic reactions, such as anaphylaxis. With the rapid development of artificial intelligence (AI) technologies, large language models (LLMs) are increasingly being used in health care, including emergency medicine and toxicology. These models have the potential to assist health care professionals in making fast and informed clinical decisions. This study aimed to assess the performance of 4 leading LLMs-ERNIE Bot 3.5 (Baidu), ERNIE Bot 4.0 (Baidu), Claude Pro (Anthropic), and ChatGPT 4.0-in managing wasp sting cases, with a focus on their accuracy, comprehensiveness, and decision-making abilities.

OBJECTIVE

The objective of this research was to systematically evaluate and compare the capabilities of the 4 LLMs in the context of wasp sting management. This involved analyzing their responses to a series of standardized questions and real-world clinical scenarios. The study aimed to determine which LLMs provided the most accurate, complete, and clinically relevant information for the management of wasp stings.

METHODS

This study used a cross-sectional design, creating 50 standardized questions that covered 10 key domains in the management of wasp stings, along with 20 real-world clinical case scenarios. Responses from the 4 LLMs were independently evaluated by 8 domain experts, who rated them on a 5-point Likert scale based on accuracy, completeness, and usefulness in clinical decision-making. Statistical comparisons between the models were made using the Wilcoxon signed-rank test, and the consistency of expert ratings was assessed using the Kendall coefficient of concordance.

RESULTS

Claude Pro achieved the highest average score of 4.7 (SD 0.603) out of 5, followed closely by ChatGPT 4.0 with a score of 4.5. ERNIE Bot 4.0 and ERNIE Bot 3.5 received average scores of 4 (SD 0.600) and 3.8, respectively. In analyzing the 20 complex clinical cases, Claude Pro significantly outperformed ERNIE Bot 3.5, particularly in areas such as managing complications and assessing the severity of reactions (P<.001). The expert ratings showed moderate agreement (Kendall W=0.67), indicating that the assessments were consistently reliable.

CONCLUSIONS

The results of this study suggest that Claude Pro and ChatGPT 4.0 are highly capable of providing accurate and comprehensive support for the clinical management of wasp stings, particularly in complex decision-making scenarios. These findings support the increasing role of AI in emergency and toxicological medicine and suggest that the choice of AI tool should be based on the specific needs of the clinical situation, ensuring that the most appropriate model is selected for different health care applications.