KERAP: A Knowledge-Enhanced Reasoning Approach for Accurate Zero-shot Diagnosis Prediction Using Multi-agent LLMs

Area of Science:

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

Background:

• Machine learning (ML) models for medical diagnosis prediction struggle with generalization due to labeled data costs.
• Large language models (LLMs) show potential but suffer from hallucinations and lack structured reasoning.
• Current methods face limitations in reliable and scalable zero-shot medical diagnosis prediction.

Purpose of the Study:

• To develop a knowledge graph (KG)-enhanced reasoning approach (KERAP) to improve LLM-based medical diagnosis prediction.
• To address challenges of hallucinations and lack of structured reasoning in LLMs for healthcare.
• To provide a scalable and interpretable solution for zero-shot diagnosis prediction.

Main Methods:

• Proposed KERAP, a multi-agent architecture integrating knowledge graphs with LLMs.
• Implemented a linkage agent for attribute mapping and a retrieval agent for structured knowledge extraction.
• Utilized a prediction agent for iterative refinement of diagnosis predictions.

Main Results:

• KERAP demonstrated enhanced diagnostic reliability in zero-shot medical diagnosis prediction.
• The approach efficiently improves the performance of LLM-based diagnostic tools.
• Experimental results validate the scalability and interpretability of the proposed framework.

Conclusions:

• KERAP offers a robust solution for improving LLM-based medical diagnosis prediction.
• The knowledge graph integration mitigates LLM limitations like hallucinations and unstructured reasoning.
• This framework advances personalized healthcare through more reliable and interpretable AI-driven diagnostics.