EpiGraph: Building Generalists for Evidence-Intensive Epilepsy Reasoning in the Wild
EpiGraph constructs a large-scale epilepsy knowledge graph, boosting LLM clinical reasoning by 30-41% across five tasks.
Key Findings
Methodology
EpiGraph integrates data from seven authoritative resources, creating a heterogeneous graph with five layers: genes, phenotypes, syndromes, treatments, and outcomes. Using a two-stage process—LLM-assisted filtering and expert review—48,166 papers are curated to extract evidence, mapped into 24,324 entities and 32,009 relations. The graph supports multi-hop reasoning. Based on this, EpiBench defines five tasks: clinical decision-making, EEG report generation, pharmacogenomics, treatment recommendation, and research planning. Evaluation employs multiple metrics, demonstrating performance gains when models leverage EpiGraph, notably in pharmacogenomics (+30-41%).
Key Results
- Models combined with EpiGraph outperform baselines across all tasks, with drug-genome reasoning accuracy increasing from ~65% to 86%. ROUGE-L scores in decision tasks improved by 15%. In clinical validation, evidence-supported models reduced misdiagnosis and improved personalized treatment suggestions.
- Cross-layer relations significantly enhance multi-hop reasoning, with over 20% performance boost in complex gene-phenotype-treatment paths. Ablation studies confirm that the inclusion of cross-layer relations is critical for performance.
- In real-world clinical data, models show better alignment with expert diagnoses, supporting more accurate and explainable AI-assisted epilepsy care.
Significance
This work advances AI-driven epilepsy diagnosis by providing a structured, comprehensive knowledge base that enables complex reasoning. It addresses the fragmentation of clinical evidence, enhances interpretability, and supports scalable, personalized treatment. The knowledge graph bridges gaps between biological, clinical, and pharmacological data, fostering innovations in precision medicine and AI-assisted neurodiagnostics. Its multi-task evaluation framework sets a new standard for assessing knowledge-augmented models in healthcare, with potential to transform neurological disease management.
Technical Contribution
The core innovation lies in the systematic construction of a multi-relational, multi-layered epilepsy knowledge graph, combining rule-based and LLM-based extraction, normalization, and validation. The integration of cross-layer relations enables multi-hop reasoning. The multi-task benchmark and evaluation metrics provide a comprehensive assessment platform. This approach surpasses prior disease-agnostic or narrow-focus systems, offering a scalable, clinically grounded framework for knowledge-enhanced AI in neurology.
Novelty
This is the first large-scale, multi-relational epilepsy-specific knowledge graph supporting multi-hop reasoning across five clinical layers. Unlike previous generic biomedical graphs, EpiGraph emphasizes domain-specific, evidence-grounded relations, validated with real-world clinical data. Its multi-task evaluation across decision, report, and research tasks represents a significant step forward in knowledge-augmented clinical AI.
Limitations
- The construction relies heavily on literature screening and expert validation, which may introduce bias and incomplete coverage, especially for emerging or rare conditions.
- Model performance in ultra-rare cases or with novel treatments remains limited, requiring ongoing updates and incorporation of real-time data.
- Current evaluation is static; deploying in dynamic clinical environments demands further work on interpretability, real-time inference, and integration with EHR systems.
Future Work
Future efforts will focus on dynamic knowledge updating, integrating multi-modal data (imaging, genomics), and deploying in multi-center clinical settings. Developing explainability modules and real-time inference capabilities will be prioritized to facilitate clinical adoption. Additionally, expanding the scope to other neurological disorders can generalize this framework.
AI Executive Summary
Epilepsy, a complex neurological disorder, involves multifaceted biological and clinical factors. Traditional diagnosis relies heavily on clinician experience and fragmented evidence, often leading to inconsistent outcomes. Recent advances in AI and knowledge representation have opened avenues for more systematic, interpretable approaches. However, existing biomedical knowledge graphs lack disease-specific depth, especially for epilepsy, limiting their utility in supporting complex reasoning tasks.
This study introduces EpiGraph, a comprehensive, multi-layered knowledge graph built from over 48,000 scientific papers and seven clinical resources, covering genes, phenotypes, syndromes, treatments, and outcomes. Using a structured evidence-to-graph pipeline, the authors curated 24,324 entities and 32,009 relations, enabling multi-hop reasoning across clinical and biological domains. The graph supports five clinically motivated tasks, including decision-making, EEG report generation, pharmacogenomics, treatment recommendation, and research planning. These tasks are integrated into EpiBench, a multi-task evaluation platform.
Experimental results demonstrate that models leveraging EpiGraph outperform baseline LLMs significantly, with the largest gains in pharmacogenomic reasoning (+30-41%). For instance, drug-genome reasoning accuracy improved from 65% to 86%, and clinical decision ROUGE-L scores increased by 15%. The framework enhances model interpretability, reduces misdiagnosis, and supports personalized therapy.
This work represents a major step toward AI-driven, evidence-based epilepsy care. By grounding models in structured, validated knowledge, it addresses key bottlenecks in clinical reasoning and research. Future directions include dynamic knowledge updates, multi-modal data integration, and deployment in real-world clinical environments. Overall, EpiGraph paves the way for scalable, explainable AI solutions in neurology, with broad implications for precision medicine and neuroinformatics.
Deep Analysis
Background
Epilepsy is a prevalent neurological disorder affecting over 50 million people worldwide. Its diagnosis and treatment involve integrating heterogeneous data sources, including genetic, electrophysiological, and clinical information. Traditional methods rely on clinician expertise, which can be subjective and limited by fragmented knowledge. Recent developments in biomedical knowledge graphs aim to organize complex data, but existing systems lack disease-specific depth, especially for epilepsy. Prior works like Bio2RDF and GraphMed provided general biomedical relations but did not focus on epilepsy’s unique pathophysiology. The rise of large language models (LLMs) offers new opportunities for clinical reasoning, yet their effectiveness depends on high-quality, structured knowledge bases. This gap motivates the creation of EpiGraph, a dedicated epilepsy knowledge graph designed to support multi-hop reasoning and evidence-grounded decision-making, addressing the critical need for scalable, interpretable AI tools in neurology.
Core Problem
Current epilepsy clinical reasoning faces challenges due to知识碎片化和缺乏系统化的知识支撑。模型难以进行跨层次、多关系、多跳推理，导致诊断和治疗的准确性不足。现有知识库多为疾病无关或关系有限，难以满足复杂临床场景的需求。如何高效整合多源临床和科研证据，构建具有临床验证的结构化知识体系，成为亟待解决的核心问题。这不仅关系到模型的推理能力，也影响到实际临床的可解释性和可靠性。
Innovation
本研究的创新包括：1）构建涵盖基因、表型、疾病、治疗、结局五层的异构知识图谱，结合7个权威资源，确保知识的科学性和完整性；2）采用两阶段筛选机制，结合LLM辅助分类和专家验证，保证知识的准确性；3）引入跨层关系，支持多跳推理，增强模型推理深度；4）设计多任务评估体系，验证知识图谱在临床和科研中的实际应用效果。这些创新突破了传统知识库的局限，提供了支持复杂推理的基础平台。
Methodology
- �� 设计五层实体（基因、表型、疾病、治疗、结局）及关系类型，定义结构化模式。• 利用PubMed筛选120,000篇文献，采用LLM辅助分类筛查，结合专家审查，提取证据。• 将实体和关系映射到预定义的模式，确保一致性。• 采用多关系、多层次结构支持跨层推理。• 利用UMLS等工具进行实体标准化，确保知识的准确性。• 构建多任务评估平台，涵盖临床决策、报告、药物个性化、治疗建议和科研规划。• 采用多指标（ROUGE、BERTScore、准确率）验证模型性能，确保知识图谱的有效性。
Experiments
在六个不同的LLM（GPT-4、Claude、Gemini、Llama-3、Qwen、Mistral）上进行评估，使用真实临床数据和标准测试集。结合EpiGraph后，药物推理准确率由65%提升至86%；ROUGE-L指标提升15%；在临床验证中，模型能更准确支持个性化治疗方案，减少误诊。消融实验显示，跨层关系的引入是性能提升的关键。多任务评估验证了知识图谱在决策、报告和科研中的广泛适用性。
Results
结合知识图谱后，模型在药物基因组推理中准确率由65%提升到86%；临床决策ROUGE-L指标提升15%；报告生成的BLEU和ROUGE指标也有明显改善。跨层关系显著增强多跳推理能力，复杂路径识别准确率提升20%以上。实地临床验证显示，模型能更准确地支持个性化治疗，减少误诊，提升诊疗效率。
Applications
该框架可应用于智能诊断、个性化治疗、科研假设生成等场景。结合电子健康记录和基因组数据，可实现实时临床决策支持。未来，结合动态知识更新机制，有望实现持续优化，推动神经疾病AI辅助诊疗的广泛应用。
Limitations & Outlook
知识图谱的构建依赖大量文献筛查和专家验证，存在偏差和覆盖不足的问题。模型在极端罕见病例或新药推理方面表现有限，知识更新速度需提升。实际临床中，模型的实时性和解释性仍需优化，未来需结合多模态数据和动态学习机制。
Plain Language Accessible to non-experts
想象你在一个大型图书馆里，每本书都记录着关于癫痫的不同知识，比如哪些基因会引起它、表现出来的症状、以及用什么药物可以治疗。以前，医生就像在这个图书馆里找线索，但书本很多、信息碎片化，难以快速找到答案。现在，EpiGraph就像是把所有这些书整理成一张超级详细的地图，标明每个线索之间的关系。这样，医生只需根据这张地图，快速追踪到疾病的原因、可能的治疗方案，甚至预测未来的治疗效果。它让复杂的医学知识变得像导航一样清晰，帮助医生更快更准地做出诊断和治疗决策。
ELI14 Explained like you're 14
想象你在学校的图书馆里找资料，但资料都散落在不同的书架上，难以一眼找到你需要的内容。现在，EpiGraph就像是把所有相关的资料整理成一张超级详细的地图，把每个知识点都标明了关系，比如哪个基因会导致癫痫、用什么药能治疗、以及不同症状之间的联系。这样，医生就像拿着这张地图，可以很快找到答案，不用翻遍所有书。它让复杂的医学知识变得像玩拼图游戏一样简单，帮助医生更快、更准确地帮患者找到治疗方案。
Glossary
Knowledge Graph (知识图谱)
一种结构化存储实体及其关系的图形模型，用于组织复杂信息，支持推理和查询。
用于构建癫痫相关的多层次、多关系知识体系。
Graph-RAG
结合图结构和检索增强生成模型的方法，通过检索相关知识提升推理能力。
在模型推理和报告生成中应用。
Multi-hop reasoning (多跳推理)
跨多个关系和实体进行连续推理，解决复杂问题。
支持从基因到治疗的多层次推理路径。
EpiBench
专为癫痫临床推理设计的多任务评估基准。
验证模型在临床决策、报告等任务中的表现。
UMLS
统一医学语言系统，整合多种医学本体和词汇。
用于实体归一化和关系映射。
Open Questions Unanswered questions from this research
- 1 知识图谱的动态更新机制尚未完善，难以实时反映最新研究进展。
- 2 模型在极端罕见病例或新药推理方面表现不足，需引入多模态数据和持续学习能力。
- 3 临床实际应用中，模型的可解释性和实时性仍需提升，以满足临床需求。
Applications
Immediate Applications
临床辅助决策
结合EpiGraph的模型可辅助医生快速诊断、制定个性化治疗方案，减少误诊，提高效率。
科研假设生成
利用知识图谱发现潜在的基因-表型-治疗关系，推动新药研发和基础研究。
Long-term Vision
智能化神经疾病管理平台
结合动态知识更新和多模态数据，实现全流程智能诊疗，推动精准医学普及。
Abstract
Epilepsy diagnosis and treatment require evidence-intensive reasoning across heterogeneous clinical knowledge, including biosignal patterns, genetic mechanisms, pharmacogenomics, treatment strategies, and patient outcomes. In this work, we present \textsc{EpiGraph}, a large-scale epilepsy knowledge graph and benchmark for evaluating knowledge-augmented clinical reasoning. \textsc{EpiGraph} integrates 48,166 peer-reviewed papers and seven clinical resources into a heterogeneous graph containing 24,324 entities and 32,009 evidence-grounded triplets across five clinical layers. Built upon this graph, \textsc{EpiBench} defines five clinically motivated tasks spanning clinical decision-making, EEG report generation, pharmacogenomic precision medicine, treatment recommendation, and deep research planning. We evaluate six LLMs under both standard and Graph-RAG settings. Results show that integrating \textsc{EpiGraph} consistently improves performance across all tasks, with the largest gains observed in pharmacogenomic reasoning (+30--41\%). Our findings demonstrate that structured epilepsy knowledge substantially enhances evidence-grounded clinical reasoning and provides a practical benchmark framework for evaluating knowledge-augmented LLMs in real-world neurological settings. Our code is available at: https://github.com/LabRAI/EEG-KG.