RAG-Enhanced Open SLMs for Hypertension Management Chatbots.

Area of Science:

• Digital Health
• Artificial Intelligence in Healthcare
• Computational Linguistics

Background:

• Chronic disease management necessitates continuous patient support and monitoring.
• Chatbots offer a scalable solution for patient engagement and assistance in home-based healthcare.
• Large language models (LLMs) present privacy and computational challenges for widespread deployment.

Purpose of the Study:

• To develop a privacy-preserving chatbot for hypertensive patients using open-source small language models (SLMs).
• To enable chatbot deployment on resource-constrained personal devices for continuous patient support.
• To evaluate the effectiveness of retrieval-augmented generation (RAG) with SLMs for question-answering (QA) tasks.

Main Methods:

• Developed a chatbot utilizing open-source SLMs for deployment on personal devices.
• Implemented retrieval-augmented generation (RAG) with a local knowledge base to enhance conversational performance and ensure data privacy.
• Experimented with eight SLMs, two prompt configurations, and various RAG strategies (embedding and retrieval).
• Evaluated chatbot performance using reference metrics and expert assessment.

Main Results:

• Retrieval-augmented generation (RAG) enhanced SLMs demonstrated improved response clarity and content accuracy in QA tasks.
• Newer SLMs, such as Qwen3, exhibited strong performance even without RAG, indicating potential for simplified future implementations.
• The developed chatbot achieved competitive accuracy with low computational costs suitable for end-user devices.

Conclusions:

• SLM-based chatbots, particularly when enhanced with RAG, offer a viable and privacy-conscious solution for chronic disease management.
• The evolution of SLMs may reduce the reliance on complex retrieval mechanisms for achieving high performance.
• This approach facilitates accessible, continuous patient support for hypertension management on personal devices.