Refine Medical Diagnosis Using Generation Augmented Retrieval and Clinical Practice Guidelines.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Natural Language Processing

Background:

• Current medical language models often rely on ICD codes, which lack clinical nuance.
• Existing models struggle to replicate clinicians' evidence-based reasoning using diverse patient data and guidelines.
• This limits the practical application of AI in clinical diagnosis.

Purpose of the Study:

• To introduce GARMLE-G, a novel framework grounding medical language models in clinical practice guidelines (CPGs).
• To enable hallucination-free, clinically aligned diagnostic recommendations by directly retrieving CPG content.
• To improve the utility and reliability of AI in medical diagnosis.

Main Methods:

• GARMLE-G integrates LLM predictions with EHR data for semantically rich queries.
• It retrieves relevant CPG knowledge snippets using embedding similarity.
• Guideline content is fused with model output for clinically aligned recommendations.

Main Results:

• GARMLE-G demonstrated superior retrieval precision and semantic relevance compared to RAG baselines.
• The framework showed enhanced clinical guideline adherence in hypertension and coronary heart disease diagnosis.
• It maintains a lightweight architecture suitable for localized healthcare deployment.

Conclusions:

• GARMLE-G offers a scalable, low-cost, and hallucination-free method for grounding medical language models in evidence-based practice.
• This approach significantly improves the clinical utility of AI-driven diagnostic tools.
• The framework has strong potential for widespread clinical deployment.