CGX: OCR-enhanced knowledge graph retrieval for explainable heart failure analysis

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Medicine
• Cardiovascular Medicine

Background:

• Existing graph-based retrieval systems face challenges in cost, scalability, and interpretability, especially in complex fields like cardiovascular medicine.
• Heterogeneous, noisy data and intricate relationships in cardiovascular medicine amplify these retrieval system limitations.

Purpose of the Study:

• To introduce CGX, a domain-oriented GraphRAG framework designed for explainable heart failure analysis that mirrors clinical reasoning.
• To address the limitations of existing systems by improving knowledge organization, retrieval accuracy, and interpretability in clinical practice.

Main Methods:

• CGX employs a three-layer knowledge hierarchy (patient observations, guideline evidence, ontologies) for cardiovascular data.
• An OCR-enhanced pipeline and zero-shot biomedical transformer convert diverse data into semantic triples, reducing errors compared to standard RAG.
• A Hybrid U-Retrieval mechanism utilizes graph topology for efficient evidence retrieval and explicit evidence chain generation.

Main Results:

• CGX demonstrated improved evidence retrieval quality and answer reliability in heart failure clinical question answering.
• The framework reduced total graph construction time by 69.7% compared to baseline methods.
• Blinded expert evaluation showed CGX reduced clinically risky answers from 12.4%-14.0% to 8.3% and improved expert-rated scores.

Conclusions:

• CGX provides a scalable and reusable GraphRAG architecture for integrating medical knowledge with LLMs.
• The framework supports trustworthy clinical decision-making by offering explainable analysis and reliable evidence retrieval.
• CGX represents a significant advancement in applying AI to complex medical domains like cardiology.