AI-Powered Test Question Generation in Medical Education: The DailyMed Approach

IntroductionLarge language models (LLMs) presents opportunities to improve the efficiency and quality of tools in medical education, such as the generation of multiple-choice questions (MCQs). However, ensuring that these questions are clinically relevant, accurate, and easily accesible and reusable remains challenging. Here, we developed DailyMed, an online automated pipeline using LLMs to generate high-quality medical MCQs. MethodsOur DailyMed pipeline involves several key steps: 1) topic generation, 2) question creation, 3) validation using Semantic Scholar, 4) difficulty grading, 5) iterative improvement of simpler questions, and 6) final human review. The Chain-of-Thought (CoT) prompting technique was applied to enhance LLM reasoning. Three state-of the art LLMs--OpenBioLLM-70B, GPT-4o, and Claude 3.5 Sonnet--were evaluated within the area of clinical genetics, and the generated questions were rated by clinical experts for validity, clarity, originality, relevance, and difficulty. ResultsGPT-4o produced the highest-rated questions, excelling in validity, originality, clarity, and relevance. Although OpenBioLLM was more cost-efficient, it consistently scored lower in all categories. GPT-4o also achieved the greatest topic diversity (89.8%), followed by Claude Sonnet (86.9%) and OpenBioLLM (80.0%). In terms of cost and performance, GPT-4o was the most efficient model, with an average cost of $0.51 per quiz and a runtime of 16 seconds per question. ConclusionsOur pipeline provides a scalable, effective and online-accessible solution for generating diverse, clinically relevant MCQs. GPT-4o demonstrated the highest overall performance, making it the preferred model for this task, while OpenBioLLM offers a cost-effective alternative.