基于多源异构信息融合的肺癌专病知识图谱构建研究

Research on the Construction of a Lung Cancer-Specific Knowledge Graph Based on Multi-Source Heterogeneous Information Fusion

摘要: 肺癌是全球重大公共卫生挑战。医学知识图谱(MKG)可为智能诊疗提供关键支持，但现有图谱常面临信息源单一、覆盖不全、缺乏真实案例等问题。为此，本研究融合MIMIC-IV电子病历、DrugBank、PubMed、ICD-10等多源异构数据，构建肺癌专病知识图谱。创新性地采用模块化子图融合方法：先构建患者、疾病、药物三个子图，再通过实体对齐融合为总图谱。实验验证：1) 基于微调BioBERT的医疗实体识别模型性能优于基线；2) 利用TransE/TransH生成的图谱嵌入在药物/手术预测任务中，Top-3和Top-5命中率均≥92%。该图谱为肺癌临床决策提供了可靠知识支撑，其构建框架为多源医学数据融合与知识图谱构建提供了可复用的参考方案。

Abstract: Lung cancer is a major global public health challenge. Medical Knowledge Graphs (MKG) can provide crucial support for intelligent diagnosis and treatment, but existing graphs often face issues such as single information sources, incomplete coverage, and a lack of real cases. To address these problems, this study constructs a lung cancer-specific knowledge graph by integrating multi-source heterogeneous data, including MIMIC-IV electronic medical records, DrugBank, PubMed, and ICD-10. It innovatively adopts a modular subgraph fusion approach: first constructing three subgraphs for patients, diseases, and drugs, then fusing them into an overall graph through entity alignment. Experimental verification shows that: 1) The medical entity recognition model based on fine-tuned BioBERT outperforms the baseline; 2) The graph embeddings generated using TransE/TransH achieve a hit rate of ≥92% for both Top-3 and Top-5 in drug/surgery prediction tasks. This graph provides reliable knowledge support for clinical decision-making in lung cancer, and its construction framework offers a reusable reference scheme for multi-source medical data fusion and knowledge graph construction.