Evaluation of Six Large Language Models for Clinical Decision Support: Application in Transfusion Decision-making for RhD Blood-type Patients.

Abstract

Background Large language models (LLMs) have the potential for clinical decision support; however, their use in specific tasks, such as determining the RhD blood type for transfusion, remains underexplored. Therefore, we evaluated the accuracy of six LLMs in addressing RhD blood type-related issues in Korean healthcare. Methods Fifteen multiple-choice and true/false questions, based on real-world transfusion scenarios and reviewed by specialists, were developed. The questions were administered twice to six LLMs (Clova X, Gemini 1.0, Gemini 1.5, ChatGPT-3.5, GPT-4.0, and GPT-4o) in both Korean and English. Results were compared against the performance of 22 transfusion medicine experts. For particularly challenging questions, prompt engineering was applied, and the questions were reevaluated. Results GPT-4o demonstrated the highest accuracy rate in Korean (0.6), with significant differences compared with those of Clova X and Gemini (P <0.05). In English, the results were similar across all models. The transfusion experts achieved a higher accuracy rate (0.8). Among the five questions subjected to prompt engineering, only GPT-4o correctly responded to one, whereas the other models failed. All LLM models changed their responses or did not respond when the same question was repeated. Conclusions GPT-4o showed the best overall performance among the models tested and may be beneficial in RhD blood product transfusion decision-making. However, its performance suggests that it may serve best in a supportive role rather than as a primary decision-making tool.