MedCoRAG: Interpretable Hepatology Diagnosis via Hybrid Evidence Retrieval and Multispecialty Consensus

MedCoRAG is an interpretable, end-to-end framework for hepatic disease diagnosis that combines hybrid evidence retrieval from knowledge graphs and clinical guidelines with a multi-agent collaborative reasoning process to emulate multidisciplinary consultation and outperform existing methods in accuracy and transparency.

The Gist

Imagine you are a patient walking into a hospital with a mysterious liver problem. In the real world, doctors don't just guess; they gather your history, check the latest medical textbooks, look at your lab results, and often hold a "team meeting" where a liver specialist, a radiologist, and a drug expert all weigh in to figure out what's wrong.

MedCoRAG is a computer program designed to act like that perfect, super-smart medical team. It's an AI system built specifically to diagnose liver diseases accurately and, most importantly, to explain how it reached its conclusion so doctors can trust it.

Here is how it works, broken down into simple analogies:

1. The Problem: The "Know-It-All" Who Makes Things Up

Standard AI models (like the chatbots you might use) are like students who have read a lot of books but haven't been updated in years. If you ask them a tricky medical question, they might confidently give you an answer that sounds good but is actually wrong or outdated. They also can't show their work, which is dangerous in medicine.

2. The Solution: MedCoRAG's "Three-Step Detective Process"

MedCoRAG doesn't just guess. It follows a strict, three-step investigation process:

Step A: The "Evidence Collector" (Hybrid RAG)

Imagine the AI has two different libraries:

1. The Rulebook Library: This contains official medical guidelines (the "laws" doctors must follow).
2. The Map Library: This is a giant, connected map of medical terms (called a Knowledge Graph) showing how symptoms link to diseases.

When a patient comes in, MedCoRAG doesn't just search for keywords. It acts like a detective who:

• Pulls the relevant pages from the Rulebook.
• Draws a specific path on the Map connecting the patient's symptoms to possible diseases.
• Crucially: It then acts like a strict editor, throwing away any pages or map paths that don't make sense for this specific patient. This ensures the AI only looks at evidence that actually fits the case.

Step B: The "Smart Dispatcher" (The Router Agent)

Once the evidence is gathered, a "Router Agent" (think of it as a triage nurse or a manager) looks at the case.

• Is it a simple case? (e.g., a common liver cyst). The manager says, "No need to call the whole team." A Generalist Agent (a general doctor AI) looks at the evidence and makes the diagnosis quickly.
• Is it a complex case? (e.g., symptoms that could be cancer, autoimmune disease, or drug reaction). The manager says, "This is tricky!" and calls in the Specialist Agents.

Step C: The "Roundtable Discussion" (Multi-Agent Collaboration)

If the case is complex, the system simulates a real hospital meeting.

• A Liver Specialist looks at the evidence.
• A Radiology Agent looks at the imaging clues.
• A Drug Expert checks if a medication caused the issue.

These "agents" don't just agree blindly. They debate!

• Agent A: "I think it's Disease X because of this symptom."
• Agent B: "Wait, the guidelines say if it were Disease X, this other lab result would be different. Let's go look for more info."

If they realize they are missing a piece of the puzzle, they trigger a re-search to find more specific evidence. They keep talking and checking until they reach a Consensus.

3. The Result: A "Traceable" Diagnosis

Finally, the Generalist Agent writes the final report. But unlike other AIs that just say "It's Disease X," MedCoRAG provides a traceable report. It says:

"We diagnosed Disease X because:

1. The patient had Symptom A.
2. The Medical Rulebook says A + B = Disease X.
3. The Liver Specialist and Drug Expert both agreed after checking the guidelines.
4. We ruled out Disease Y because..."

Why This Matters

• No Hallucinations: Because it constantly checks against real medical rules and maps, it's much less likely to make things up.
• Trust: Doctors can see the "chain of thought," making it safe to use in real hospitals.
• Efficiency: It only calls in the "experts" when the case is actually hard, saving time and computing power.

In short: MedCoRAG is like a digital medical team that reads the latest rules, draws the right maps, holds a debate when things get complicated, and hands you a diagnosis with a full "receipt" showing exactly how they figured it out.

Technical Summary

Here is a detailed technical summary of the paper "MedCoRAG: Interpretable Hepatology Diagnosis via Hybrid Evidence Retrieval and Multispecialty Consensus."

1. Problem Statement

Accurate and interpretable diagnosis of hepatic (liver) diseases from real-world Electronic Health Records (EHRs) is critical but challenging. Existing AI approaches face three primary limitations:

• Lack of Transparency: Large Language Models (LLMs) often provide "black box" diagnoses without traceable, step-by-step justification, which is unacceptable for high-stakes medical decisions.
• Static and Outdated Knowledge: Standard LLMs rely on pre-trained knowledge that may be outdated, leading to confident but incorrect diagnoses.
• Inadequate Reasoning Structures:
  • Traditional Retrieval-Augmented Generation (RAG) relies on unstructured text, failing to support multi-hop clinical inference.
  • Knowledge Graph (KG) approaches often retrieve irrelevant paths and lack integration with clinical practice guidelines.
  • Existing Multi-Agent systems typically use static teams of specialists regardless of case complexity and rely on loosely retrieved or internal knowledge, lacking deep integration of authoritative guidelines and structured KGs.

2. Methodology: MedCoRAG Framework

MedCoRAG (Medical Collaborative RAG) is an end-to-end framework designed to emulate multidisciplinary team (MDT) consultations. It integrates Hybrid RAG (combining structured KGs and unstructured guidelines) with Dynamic Multi-Agent Reasoning.

The workflow consists of three core modules:

A. Abnormal Findings and Preliminary Diagnosis

1. Entity Extraction & Standardization: The system extracts raw abnormal entities from the patient narrative using an LLM and maps them to standardized medical terminology (UMLS) via a knowledge graph matching process.
2. Hypothesis Generation: Using the standardized entities and the full case description, the LLM generates a focused list of initial differential diagnoses (up to 4 candidates).

B. Hybrid RAG (Evidence Construction)

For each diagnostic hypothesis, the system constructs a patient-specific evidence package by retrieving two complementary sources:

1. Clinical Guideline Retrieval: A two-stage retrieval process (Bi-encoder for dense retrieval, Cross-encoder for re-ranking) fetches relevant excerpts from authoritative clinical guidelines (e.g., AASLD, EASL).
2. Knowledge Graph Path Retrieval & Pruning: The system retrieves semantic paths (up to 3 hops) connecting abnormal entities to the diagnosis from the UMLS knowledge graph.
   • Pruning Mechanism: An LLM evaluates these raw paths against the full clinical narrative and the retrieved guidelines. It discards paths that are clinically incoherent or unsupported, retaining only a "pruned" set of valid reasoning paths.
   • Output: A structured evidence package containing guideline excerpts and validated KG paths.

C. Multi-Agent Collaborative Reasoning

A Router Agent assesses case complexity based on the clinical narrative and abnormal findings:

• Simple Cases: Routed directly to a Generalist Agent, which synthesizes a diagnosis from the initial evidence package.
• Complex Cases: The Router dynamically dispatches relevant Specialist Agents (e.g., Hepatology, Oncology, Radiology) based on the specific clinical context (e.g., a case with jaundice and kidney issues triggers Hepatology and Nephrology agents).
  • Iterative Deliberation: Specialists iteratively reason over the shared evidence, providing stances (Support/Neutral/Oppose), confidence scores, and sufficiency judgments.
  • Targeted Re-retrieval: If evidence is deemed insufficient (by >50% of agents), the system triggers a targeted re-retrieval round to fetch specific missing data (e.g., specific biomarker thresholds).
  • Consensus Formation: The Generalist Agent performs holistic adjudication, synthesizing all deliberations, disagreements, and uncertainties into a single, traceable consensus diagnosis.

3. Key Contributions

1. Dynamic Multi-Specialty RAG: Unlike static agent frameworks, MedCoRAG dynamically routes cases to on-demand specialists based on abnormal findings, emulating real-world MDT workflows.
2. MDT-Aligned Hybrid Reasoning: It unifies pruned Knowledge Graph paths and clinical guideline excerpts into a single evidential space. This "hybrid" approach ensures reasoning is grounded in both structured medical relationships and authoritative practice standards.
3. Iterative Evidence-Constrained Loop: The framework introduces a closed-loop mechanism where agents can trigger re-retrieval if evidence is insufficient, ensuring the reasoning process evolves until diagnostic confidence stabilizes.
4. Interpretability: The system produces a traceable consensus report that explicitly links the final diagnosis to supporting/contradicting evidence, guideline excerpts, and the deliberation history of specialist agents.

4. Experimental Results

The framework was evaluated on 3,470 real-world hepatic disease cases from the MIMIC-IV dataset, covering 13 distinct liver disease categories.

• Performance: MedCoRAG achieved state-of-the-art results, outperforming:
  • Specialized Medical Models (up to 8B parameters).
  • Large Proprietary Models (e.g., GPT-4o, Gemini-2.5-Pro, GLM-4.6).
  • Medium/Small General Models (14B–32B).
  • Existing Agent Frameworks (e.g., ColaCare, MedAgent-Pro).
• Key Metrics:
  • Weighted F1-Score: 79.12% (vs. ~77% for the next best baseline).
  • Precision: 81.32%.
  • Recall: 79.18%.
• Ablation Studies: Removing any single component (Clinical Guidelines, Multi-Agent deliberation, Knowledge Graph, or Teacher Distillation) resulted in significant performance drops. The most severe drop occurred when both KG and Guidelines were removed (F1 dropped to 55.32%), proving the necessity of hybrid evidence.
• Efficiency: Simple cases were resolved in ~10 seconds, while complex multi-agent cases averaged ~33 seconds.

5. Significance and Conclusion

MedCoRAG represents a significant step forward in Trustworthy Medical AI.

• Clinical Relevance: It addresses the "black box" problem by providing a transparent, evidence-grounded reasoning path that aligns with how human specialists collaborate.
• Adaptability: By dynamically routing agents based on case complexity, it optimizes computational resources while ensuring complex cases receive deep, multi-disciplinary scrutiny.
• Generalizability: While tested on hepatology, the architecture (Hybrid RAG + Dynamic Routing + Consensus) offers a blueprint for AI-driven diagnosis in other complex medical domains requiring multidisciplinary input.

The authors conclude that MedCoRAG successfully bridges the gap between high-performance LLMs and the rigorous, explainable standards required for clinical decision support in high-stakes environments.