Building Trustworthy Generative Artificial Intelligence for Diabetes Care and Limb Preservation: A Medical Knowledge...

Area of Science:

• Artificial Intelligence in Medicine
• Health Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) show promise for medical information extraction but risk inaccuracies.
• Diabetes and diabetic foot care knowledge gaps impact patient outcomes, particularly limb loss.
• Improving patient health literacy is crucial for diabetes self-management.

Purpose of the Study:

• To develop and validate a retriever-augmented generation (RAG) model for accurate diabetes and diabetic foot care information delivery.
• To create a user-friendly AI tool for laypersons with an eighth-grade literacy level.
• To enhance patient self-education and self-management capabilities.

Main Methods:

• Utilized a RAG architecture with GPT-4 and Pinecone vector database.
• Built a question-and-answer AI model based on NIH National Standards for Diabetes Self-Management Education.
• Validated model outputs via expert review against guidelines and literature, testing with 175 questions.

Main Results:

• The RAG model achieved 98% accuracy with optimized content volume and few-shot learning prompts.
• Demonstrated capability to deliver user-friendly and comprehensible medical information.
• Successfully extracted knowledge on diabetes and diabetic foot care.

Conclusions:

• The RAG model is a promising tool for disseminating reliable medical knowledge to the public.
• Effective for diabetes self-education and self-management.
• Highlights the importance of content validation and prompt engineering in AI applications.