| 基于双向LSTM卷积网络与注意力机制的自动睡眠分期模型 |
| |
| 引用本文： | 李倩玉,王蓓,金晶,张涛,王行愚.基于双向LSTM卷积网络与注意力机制的自动睡眠分期模型[J].智能系统学报,2022,17(3):523-530. |
| |
| 作者姓名： | 李倩玉 王蓓 金晶 张涛 王行愚 |
| |
| 作者单位： | 1. 华东理工大学 信息科学与工程学院, 上海 200237;2. 清华大学 自动化系, 北京 100086 |
| |
| 摘 要： | 针对现阶段深度睡眠分期模型存在的梯度消失、对时序信息学习能力较弱等问题，提出一种基于双向长短时记忆卷积网络与注意力机制的自动睡眠分期模型。将少样本类别的睡眠脑电数据通过过采样方式进行数据增强后，利用带残差块的卷积神经网络学习数据特征表示，再通过带注意力层的双向长短时记忆网络挖掘深层时序信息，使用Softmax层实现睡眠分期的自动判别。实验使用Sleep-EDF数据集中19晚单通道脑电信号对模型进行交叉验证，取得了较高的分类准确率和宏平均F1值，优于对比方法。该方法能够有效缓解睡眠分期判别中少数类分类性能较低的问题，并提高了深度睡眠分期模型的整体分类性能。
|
| 关 键 词： | 睡眠分期 脑电图 卷积神经网络 残差网络 双向长短时记忆网络 注意力机制 类不平衡 过采样 |
Automatic sleep staging model based on the bi-directional LSTM convolutional network and attention mechanism |
| |
| LI Qianyu,WANG Bei,JIN Jing,ZHANG Tao,WANG Xingyu.Automatic sleep staging model based on the bi-directional LSTM convolutional network and attention mechanism[J].CAAL Transactions on Intelligent Systems,2022,17(3):523-530. |
| |
| Authors: | LI Qianyu WANG Bei JIN Jing ZHANG Tao WANG Xingyu |
| |
| Affiliation: | 1. School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China;2. Department of Automation, Tsinghua University, Beijing 100086, China |
| |
| Abstract: | Owing to the problems of gradient disappearance and the weak learning ability of time series information in the current deep sleep staging model, we propose an automatic sleep staging model based on the bi-directional long short-tern memory convolutional network and attention mechanism. After the sleep EEG data of minority classes are enhanced by oversampling, a convolutional neural network with a residual block is designed to learn the data feature representation, and then, an attention layer is combined with the BiLSTM network to extract the deep time sequence information, a softmax layer is adopted to realize the automatic discrimination of sleep stages. A total of 19-night single-channel EEG signals from the Sleep-EDF dataset are analyzed to cross-verify the proposed model. The obtained classification accuracy and macro-F1-score (MF1) are more satisfied than the comparison methods. The effect of low classification performance of minority classes in sleep staging is reduced effectively. The overall classification performance by the proposed deep sleep staging model is sufficiently improved. |
| |
| Keywords: | sleep staging electroencephalogram convolutional neural network residual network Bi-directional long short-term memory network attention mechanism class imbalance oversampling |
|
| 点击此处可从《智能系统学报》浏览原始摘要信息 |
|
点击此处可从《智能系统学报》下载全文 |
