MediCARE: Medical Collaborative Agents REasoning over Interpretable Heterogeneous Graphs.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Informatics
• Machine Learning for Healthcare

Background:

• Large Language Models (LLMs) show promise as medical reasoning agents.
• LLM hallucinations present a significant challenge in the medical domain, risking inaccurate patient information.
• Existing AI approaches require robust methods to ensure reliability and interpretability in clinical decision support.

Purpose of the Study:

• To propose and evaluate the Predict→Interpret→Explain (PIE) paradigm for reliable medical AI.
• To develop a framework (MediCARE) that integrates Graph Neural Networks (GNNs) with LLMs for enhanced medical reporting.
• To mitigate LLM hallucinations and improve the precision of AI-driven medical recommendations.

Main Methods:

• Utilized graph data structures to represent patient clinical data for Graph Neural Network (GNN) modeling.
• Employed eXplainable AI (XAI) techniques, including GNNExplainer and Integrated Gradients, for model interpretation.
• Leveraged a pool of LLM-based collaborative agents to process predictions and generate comprehensive medical reports.

Main Results:

• The MediCARE framework demonstrated the capability of LLMs in generating accurate medical reports.
• The integration of LLMs and GNNs led to improved prediction precision in medication recommendation tasks.
• The PIE paradigm effectively addressed the challenge of LLM hallucinations in a medical context.

Conclusions:

• The MediCARE framework, based on the PIE paradigm, offers a promising approach to reliable AI in medicine.
• LLMs can be effectively utilized to enhance the accuracy and comprehensibility of AI-generated medical assessments.
• This study highlights the potential of combining predictive modeling, XAI, and LLMs for safer clinical decision support systems.