Related Experiment Video
Updated: Jan 9, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Comparative study of the performance of ChatGPT-4, Claude, Gemini, Mistral, and perplexity on multiple-choice
Martin Wendlassida Nacanabo1, Yannick Laurent Tchenadoyo Bayala2, André Arthur Taryètba Seghda3
1Department of Cardiology, Bogodogo University Hospital, Ouagadougou, Burkina Faso. nacmartinwend@gmail.com.
Background:
Artificial intelligence, particularly Large Language Models (LLMs), has revolutionized the field of medicine. Their ability to understand and answer medical questions is generating growing interest, especially in cardiology, where diagnostic and therapeutic accuracy is essential.
Objective:
The objective of our study was to assess and compare the performance of five LLMs on multiple-choice questions (MCQs) in cardiology.
Materials And Methods:
This was a comparative study conducted in the cardiology department of the Bogodogo University Hospital, Ouagadougou, involving 83 MCQs derived from the 2020 French national cardiology curriculum. The questions were submitted to ChatGPT-4, Claude, Gemini, Mistral, and Perplexity. Performance was evaluated based on overall and thematic accuracy, as well as the number of discordances. Agreement between the LLMs was assessed using the Kruskal-Wallis test.
Results:
Claude achieved the highest overall accuracy (78.31%), followed by ChatGPT-4 and Gemini (75.90%), then Mistral (72.29%) and Perplexity (68.67%). Each LLM demonstrated a distinct performance profile by topic, with Claude excelling in heart failure (100%) and arrhythmias (90.9%), and ChatGPT-4 in diagnostic investigations (87.5%). The analysis of discordances showed a slightly higher precision for ChatGPT-4. The Kruskal-Wallis test with effect size revealed statistically significant differences in performance between the LLMs, whether globally, by topic (p < 0.05) and with generally large effect sizes.
Conclusion:
Despite variations in their performance profiles, these five LLMs studied have relatively similar capabilities for answering well-structured cardiology multiple-choice questions. They could therefore be valuable tools in medical education in our resource-limited context.
