【摘要】 针对辅助诊断过程中病人所患疾病不单一，多种疾病之间存在内在关联，及长病历文本特征提取较为困难等问题，该文提出一种基于知识增强的多视野表征学习方法。该方法首先使用Bi-LSTM和注意力网络、医疗知识图融合、预训练模型分别从字符视野、实体视野、文档视野提取疾病表征，并通过融合多视野信息从长病历文本中准确抽取疾病诊断相关特征。而后建模疾病间内在关联关系，基于图神经网络方法进行知识融合以增强疾病表征，并实现疾病预测。该模型利用多视野表征学习与知识增强方法，提升了疾病预测的性能，通过结果可视化为模型提供了可解释性。在华为云杯评测数据上的实验表明，该方法优于其他基线方法，消融实验验证了该方法各模块的有效性。更多还原

【Abstract】 To model internal correlations between diseases and extract features from long medical records, we propose a multi-view representation learning network based on knowledge augmentation for auxiliary diagnosis. Firstly, the method combines the Bi-LSTM, the attention network, the medical knowledge graph, and the pre-trained models to extract disease representations from character view, entity view, and document view, respectively. Then, the features related to disease diagnosis are accurately extracted from the long medical record text by the fusion of multi-view information. Secondly, the internal correlation between diseases is modeled by knowledge fusion based on the graph neural network to enhance disease representation. Finally, the model uses multi-view representation learning and knowledge enhancement methods to predict disease. Experiments on Huawei Cloud evaluation dataset show that the model is superior to baseline methods, and ablation studies prove the effectiveness of each module in this method.更多还原