基于并行神经网络的疾病特征实体识别方法

Disease Feature Entity Recognition Method Based on Parallel Neural Network

摘要: 针对静脉血栓栓塞症电子病历文本语义复杂，疾病信息多维性导致的疾病特征学习不彻底、实体识别不准确的问题，本文提出了一种并行神经网络的疾病特征实体识别方法。首先，通过RoBERTa模型，更好地学习到病历实体中的特征信息。然后，通过双向长短期记忆网络，提取病历中的全局特征，再经过并行的迭代膨胀卷积神经网络提取病历中的局部特征。最后，利用CRF推理层修正神经网络输出的疾病特征标签。在医院提供的2000份静脉血栓电子病历上，本方法的平均准确率为85.26%，相对于单纯的卷积神经网络，该方案的识别准确率提高了13.52%。

Abstract: In response to the intricate semantics of electronic medical records for venous thromboembolism and the multi-dimensionality of disease information that gives rise to incomplete acquisition of disease features and inaccurate entity recognition, this paper presents a parallel neural network-based disease feature and entity recognition approach. Firstly, the language representation RoBERTa model is employed to more effectively acquire the feature information of medical record entities. Subsequently, a bidirectional long short-term memory network is utilized to extract global features from the medical record, followed by a parallel iterative dilated convolutional neural network for extracting local features from the medical record. Eventually, the CRF inference layer is adopted to rectify the disease feature labels output by the neural network. On the 2000 venous thromboembolism electronic medical records provided by the hospital, the average accuracy of the proposed method is 85.26%, which is 13.52% higher than that of the pure convolutional neural network.