Retrieval-augmented generation for medication safety: A case study using drug package inserts.

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Education Technology
• Natural Language Processing

Background:

• Retrieval-augmented generation (RAG) shows potential for reducing large language model (LLM) hallucinations.
• The application of RAG in medication safety education is not well-studied.
• This study evaluates RAG-based LLMs for generating patient medication safety education materials.

Purpose of the Study:

• To assess the performance of RAG-based LLMs in creating patient medication safety education.
• To compare LLM-generated content with materials created by clinical experts.
• To evaluate the agreement between LLM-as-a-judge assessments and clinical expert judgments.

Main Methods:

• Clinical pharmacists developed an ontology for medication safety information.
• GPT-4o, OAGLLM, and human experts generated educational materials.
• Seven pharmacists conducted blinded evaluations using a 5-point Likert scale across six dimensions.

Main Results:

• Human experts achieved the highest overall score (0.89 ± 0.16), followed by GPT-4o (0.86 ± 0.20) and OAGLLM (0.84 ± 0.17).
• Human-generated materials excelled in Coverage, Relevance, and Accuracy.
• LLM evaluations showed fair agreement (κ = 0.21-0.24) and low-to-moderate alignment (ρ = 0.40-0.53) with human experts.

Conclusions:

• Clinical experts remain superior in generating medication safety education materials.
• GPT-4o demonstrates significant potential for medication safety education.
• Expert oversight is crucial due to limited agreement between LLM evaluations and human judgments.