面向医学知识图谱演进的图神经网络知识融合及补全方法研究

The construction and evolution of medical knowledge graph can rich computer intelligence system in large-scale and high-quality medical knowledge, which is significantly support the evolution for the research of medical artificial intelligence from perceptual intelligence to cognitive intelligence. However, the current medical knowledge graph is still in its infancy, and has many problems such as sparse and low-quality data, poor dynamic scalability, which restricts the development and industrialization of AI applications such as medical intelligence QA system and intelligent diagnosis. There are many researches on knowledge graph fusion and complement, but there are still many shortcomings in accuracy and efficiency of existing methods because of the inability to utilize structural information and timing information. The research of this project is mainly carried out from three aspects. Firstly, we focus on the knowledge fusion model which is based on efficient graph neural networks. Secondly, we research on the knowledge complement model with graph neural network which is combined with adversarial learning and attention mechanism. Thirdly, we research on the representation of newly dynamic knowledge graph with multi information. Through the method, we will utilize structural information and represent timing information which can learn more accurately representation of entities and relationships to improve the accuracy of knowledge graph fusion and completion. By the way, our method can solve the problem of dynamic expansion of knowledge graph. Meanwhile, we do some experiments about our project and the results to verify the validity of our methods.

医学知识图谱构建和持续性演进增强，可以让计算机智能系统拥有大规模、高质量的形式化医学知识，可以显著支撑医学人工智能领域研究从感知智能向认知智能升级。但当前医学知识图谱发展还处于数据稀疏、质量低、动态扩展性差等初期阶段，制约着我国医学智能问答、智能诊断等AI应用的研发和产业化落地。已有针对知识图谱融合、补全的方法，由于未能利用图谱结构、时序信息，在精度和效率上还存在诸多不足。因此本课题从最新的图神经网络结构出发，围绕基于高效的图神经网络的知识融合模型、结合对抗学习和注意力机制的图神经网络知识补全模型、基于多信息融合的新型动态知识图谱表示模型三个方面展开研究，通过一整套方法实现知识图谱结构信息的高效利用和时序信息的有效表示，准确地学习实体和关系的表示，提高知识图谱融合和补全的精度，并解决知识更新所带来的知识图谱动态扩展的问题。最后，通过医学资源平台，对课题取得的成果与关键技术进行验证。

课题组围绕基于图神经网络的知识融合与实体对齐技术研究，完成了面向多模态医学知识图谱融合的知识图谱实体对齐、面向医学决策场景的超关系医学知识图谱表示、复杂知识图谱的补全及多跳推理、动态知识图谱补全与推理问答技术研究，开展成果落地应用。.具体包括：1）提出了基于回响网络和自举策略的实体对齐框架EchoEA，大幅减少对医学数据标注的依赖，同时在3个跨语言对齐国际标准数据集上，将top1准确率提升到96%，成为当前SOTA模型。进一步提出了医学影像与疾病实体标签的分类对齐模型，建成了眼科湿性AMD疾病的首个多模态医学知识图谱。2）提出了超关系结构表示复杂医学知识的新方法，建成了高血压疾病的首个超关系医学知识图谱，有效支撑了高血压用药辅助决策的智能推理需求，开发的具备高可解释性的高血压临床决策支持系统，已在安贞医院高血压科开展临床应用。 3）提出了面向层次多元的复杂知识图谱的补全及多跳推理模型，构建了超关系知识图谱上的双视图知识图谱数据集及复杂查询数据集，为同行在医学知识图谱的知识检索、辅助诊断等领域研究，提供了标准评测指标和数据集。4）提出了基于时序知识图谱的递归时态事实嵌入框架，为在医疗领域的动态知识图谱研究提供了关键技术支持。.发表及录用学术论文13篇，其中CCF A类国际会议录取2篇，CCF B类国际会议1篇，SCI中科院二区4篇，SCI中科院三区1篇，CCF中文T1和T2期刊各1篇。申请国家发明专利11项，获授权1项；国际发明专利1项。软件著作权申请6项。构建了2个专科疾病精细化知识图谱高质量数据集。.课题组进一步研发了基于精细化高血压知识图谱的智能慢病管理服务平台，为高血压慢病患者提供了高血压智能预问诊和高血压慢病智能管理小程序、为医生提供了慢病患者智能管理和高血压用药智能辅助决策等智能化AI医疗服务。平台已为18所医院的25位医生、300余名患者提供了智能高血压服务。