HEAL-KGGen: A Hierarchical Multi-Agent LLM Framework with Knowledge Graph Enhancement for Genetic Biomarker-Based Medical Diagnosis
Kaiwen Zuo, Zixuan Zhong, Peizhou Huang, Shiyan Tang, Yuyan Chen, Yirui Jiang · 2025 · 13 citationsRead the paper
ABSTRACT The discovery and validation of genetic biomarkers across diverse diseases demand intelligent systems capable of integrating complex multi-omics data with clinical relevance. We introduce HEAL-KGGen, an end-to-end framework that enhances Large Language Models (LLMs) through a hierarchical multi-agent architecture and an automatically constructed medical knowledge graph. The system includes a General Practitioner (GP) agent for initial biomarker triage and specialist agents for genomics, transcriptomics, proteomics, and clinical interpretation. The core innovation of HEAL-KGGen lies in its dynamic knowledge graph pipeline, which combines entity extraction based on patterns and semantics, ontology-aligned normalization (using UMLS, MeSH, SNOMED CT) and the construction of multi-source relationships from biomedical databases and literature. Retrieved subgraphs are transformed into contextual prompts that guide LLM reasoning via structured, explainable pathways. Our experiments show that HEAL-KGGen significantly improves question-answering accuracy across multiple mainstream large language models, with the highest improvement observed on Claude 3.5 Sonnetachieving a 43.75% increase in accuracy., confirming the value of domain-specific graph knowledge in advancing LLM performance for genetic and molecular diagnostics.
What the full text shows
Working prototype- HEAL-KGGen increased question-answering accuracy across all tested LLMs, with the highest gain seen in Claude 3.5 Sonnet. — 43.75% increase in accuracy (Compared to baseline Claude 3.5 Sonnet on 80 biomedical MCQs)
- Average accuracy improvement across all tested LLMs. — 27.8125% average improvement (Across Claude 3.5 Sonnet, GPT-4, Gemini 1.5 Pro, and LLaMA 3.1 405B)
- Improvement in accuracy for GPT-4. — 10.00% increase (70.00% to 80.00%) (On 80 biomedical MCQs)
- Improvement in accuracy for Gemini 1.5 Pro. — 40.00% increase (15.00% to 55.00%) (On 80 biomedical MCQs)
- Improvement in accuracy for LLaMA 3.1 405B. — 17.50% increase (30.00% to 47.50%) (On 80 biomedical MCQs)
- Efficiency of medical entity extraction. — 3.76 entities per question on average (Using HEAL-KGGen framework)
- Setup
- Benchmark dataset of 80 single-optioned multiple-choice questions (MCQs) derived from genomic databases; tested with GPT-4, Claude 3.5 Sonnet, Gemini 1.5 Pro, and LLaMA 3.1 405B.
- How
- The framework uses a hierarchical multi-agent architecture (GP agent and specialists in genomics, transcriptomics, proteomics, and clinical interpretation) integrated with a Neo4j-based medical knowledge graph. It employs a pipeline of pattern-based entity extraction, ontology-aligned normalization (UMLS, MeSH, SNOMED CT), and subgraph-to-prompt transformation to guide LLM reasoning.
- Limitations
- Underperforms in specific domains such as chromosomal abnormalities and rare diseases. · Knowledge graph coverage is not exhaustive and requires labor-intensive curation. · Chain-of-thought traces are not yet embedded in agent outputs for full auditability. · Benchmark questions are simpler than real-world clinical scenarios with long causal chains.
Summarized by Gemma 4 from the open-access full text · 8 Oct 2026. Read it