Chatbot To Help Patients Understand Their Health

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Patient engagement is crucial for effective healthcare.
• Current AI tools often lack the nuanced communication needed for patient education.
• A gap exists in AI systems trained for patient understanding without extensive human-labeled data.

Purpose of the Study:

• To introduce NoteAid-Chatbot, a conversational AI designed to improve patient comprehension.
• To demonstrate a novel 'learning as conversation' framework using large language models (LLMs) and reinforcement learning (RL).
• To validate the chatbot's effectiveness in promoting patient understanding in simulated healthcare scenarios.

Main Methods:

• Developed NoteAid-Chatbot using a lightweight 3B-parameter LLaMA 3.2 model.
• Employed a two-stage training process: supervised fine-tuning on synthetic medical conversations and RL using patient understanding assessments.
• Utilized Proximal Policy Optimization (PPO)-based reward modeling for training without human-labeled data.

Main Results:

• NoteAid-Chatbot demonstrated emergent behaviors like clarity, relevance, and structured dialogue.
• The AI successfully handled multi-turn interactions and incorporated diverse educational strategies.
• A Turing test indicated NoteAid-Chatbot outperformed non-expert humans in simulated patient education.

Conclusions:

• Low-cost, PPO-based RL can effectively train domain-specific chatbots for complex conversational tasks.
• The 'learning as conversation' framework shows promise for realistic, open-ended dialogue applications.
• NoteAid-Chatbot's success broadens the applicability of RL-based alignment methods beyond healthcare.