生物学杂志

生物学杂志 ›› 2020, Vol. 37 ›› Issue (1): 11-.doi: 10.3969/j.issn.2095-1736.2020.01.011

• 研究报告 • 上一篇    下一篇

基于图注意力深度神经网络的癫痫脑电尖波识别

  

  1. 1．清华大学 脑与智能实验室及生物医学工程系， 北京 100084； 2．北京未来芯片高精尖中心及清华大学类脑计算研究中心， 北京 100084； 3．清华-IDG/麦戈文脑科学联合研究院， 北京 100084
  • 出版日期:2020-02-18 发布日期:2020-03-09
  • 通讯作者: 宋森，博士，研究员，主要研究方向为类脑计算与神经科学，E-mail:songsen@tsinghua.edu.cn
  • 作者简介:崔昊天，硕士，主要从事深度学习和生物医学工程研究，E-mail:cht15@mails.tsinghua.edu.cn
  • 基金资助:
    国家自然科学基金项目（61836004）；国家自然科学基金项目（31871071）；北京脑科学专项（Z181100001518006）

Spike detection in epileptic EEG recording based on graph attention deep neural networks

  1. 1. Laboratory for Brain and Intelligence and Department of Biomedical Engineering, Tsinghua University,  Beijing 100084; 2. Beijing Innovation Center for Future Chip and Center for Brain-inspired Computing  Research, Tsinghua University, Beijing 100084; 3. Tsinghua-IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China
  • Online:2020-02-18 Published:2020-03-09

PDF (PC)

98

摘要： 头皮脑电记录中的尖波识别能够为癫痫诊断提供重要的帮助和参考，但已报道方法中存在特异性低，忽略多通道间信息等问题。通过建立一种深度神经网络算法，实现癫痫脑电中尖波自动识别。研究中以临床头皮脑电记录为样本，首次建立层叠的卷积与图注意力深度神经网络模型进行识别，分别对脑电电极通道内和通道间的特征作有效提取。实验不仅发现算法的识别准确率达到93.55%，特异性96.83%，相较于对比方法有明显提升；同时通过电极通道间的注意力指数得到和尖波放电明显相关的电极关系图，有尖波发生的电极区域在图中明显聚集并分离于背景电极，可以为癫痫病灶定位提供辅助信息。上述算法，创新地利用多通道间信息，取得了领先的识别准确率，具有良好的临床应用前景。

关键词: 头皮脑电, 尖波识别, 深度学习, 注意力模型, 卷积神经网络

Abstract: Spike recognition in scalp EEG recording can contribute greatly to diagnosis process of epilepsy, however, problems exist in reported methods including low specificity and lack of inter-channel information. This study implements a deep neural network based algorithm to achieve automatic detection of spikes in epileptic EEG. The study used clinical scalp EEG records and established a stratified convolutional and graph attention deep neural network model for the first time, which achieved effective feature extraction from inner-channels and inter-channels of the EEG electrodes. The experiment not only found that the classification result reached the state of art 93.55% accuracy and 96.83% specificity. At the same time, it was found that the attention index between different electrode channels had a positive correlation with the discharge frequency of the region - the electrodes with spikes were aggregating and detached from the background electrodes in the graph, which could be used as auxiliary information for focus localization. The above algorithm innovatively utilizes cross channel relations in a data-driven manner, achieves superior accuracy and has promising clinical application prospects.

Key words: scalp EEG, spike detection, deep learning, attention model, convolutional neural network

中图分类号: