Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (5): 815-824.DOI: 10.3778/j.issn.1673-9418.1908018

Previous Articles     Next Articles

Research on Disease Diagnosis Method Combining Knowledge Graph and Deep Learning

DONG Lili, CHENG Jiong, ZHANG Xiang, YE Na   

  1. School of Information and Control Engineering, Xi??an University of Architecture and Technology, Xi??an 710055, China
  • Online:2020-05-01 Published:2020-05-08

融合知识图谱与深度学习的疾病诊断方法研究

董丽丽程炯张翔叶娜   

  1. 西安建筑科技大学 信息与控制工程学院,西安 710055

Abstract:

Focusing on the problem that existing disease diagnosis methods using deep learning rely heavily on labeled data in the auxiliary diagnosis process, and lack the experience and knowledge of doctors or experts, a disease diagnosis method combining medical knowledge graph and deep learning is proposed. The core of this method is a knowledge driven convolutional neural network (CNN) model. By using the entity linking and disam-biguation and knowledge graph embedding technologies to get the structured disease knowledge of medical knowledge graph, the word vector of disease features in the disease description text and the entity vector of corresponding knowledge are taken as multi-channel input of CNN, so as to represent different types of diseases from semantic and knowledge levels in the convolution process. Through training and testing on multiple disease description datasets, the experimental results show that the diagnostic performance of the proposed method is better than that of the single CNN model and other disease diagnosis methods, and verify that this method combining knowledge and data training is more suitable for the preliminary diagnosis of disease types of disease description.

Key words: knowledge graph embedding, expert experience, convolutional neural network (CNN), multi-channel, disease diagnosis

摘要:

针对现有深度学习疾病诊断方法在辅助诊断过程中大规模依赖标注数据,且缺乏医生或专家经验知识的问题,提出一种融合医学知识图谱与深度学习的疾病诊断方法。该方法的核心是一个知识驱动的卷积神经网络(CNN)模型,通过实体链接消歧与知识图谱嵌入抽取得到医学知识图谱中的结构化疾病知识,并将病情描述文本中的疾病特征词向量与相应知识实体向量作为CNN的多通道输入。在卷积过程中从语义和知识两个层面表示不同类型疾病。通过在多类病情描述文本数据集上进行训练和测试,实验结果表明该方法的诊断性能要优于单一CNN模型与其他疾病诊断方法,并验证了这种知识与数据联合训练的方法更适用于初步诊断病情描述的疾病类型。

关键词: 知识图谱嵌入, 专家经验知识, 卷积神经网络(CNN), 多通道, 疾病诊断