A knowledge-enhanced domain-aware large language model agent for atrial fibrillation management

Area of Science:

• Artificial Intelligence in Medicine
• Digital Health
• Patient Self-Management

Background:

• Large language models (LLMs) offer potential for atrial fibrillation (AF) management but face accuracy challenges.
• Current clinical LLM applications often yield suboptimal results for patient self-care.

Purpose of the Study:

• To develop PULSE, a novel knowledge-enhanced, domain-aware LLM agent for improved AF patient self-management.
• To integrate multimodal inputs, curated knowledge bases, prompt engineering, and retrieval-augmented generation.
• To enhance the accuracy, integrity, utility, safety, and readability of LLM outputs for AF patients.

Main Methods:

• Developed PULSE, an agent-based LLM framework with knowledge enhancement and domain awareness.
• Integrated multimodal data, clinical knowledge bases, optimized prompts, and retrieval-augmented generation.
• Evaluated PULSE against four base LLMs on response quality and readability, followed by expert assessment of 75 real-world queries.

Main Results:

• PULSE significantly improved clinical accuracy, content integrity, utility, and patient safety (P < 0.05) compared to base LLMs.
• Enhanced empathy and clarity of responses while maintaining conciseness.
• Demonstrated superior performance across all evaluated metrics.

Conclusions:

• PULSE enhances factual accuracy and readability of LLM-generated medical information for AF patients.
• Agent-driven LLM systems show significant clinical potential for chronic disease self-management.
• PULSE can improve long-term patient outcomes in AF management.