采用小波变换和双向长短期记忆网络的脑电睡眠分期模型  

EEG Sleep Staging Model Using Wavelet Transform and Bidirectional Long Short-Term Memory Network

作　　者：王天宇 陈晗 王刚[2,3] 吴宁[1] WANG Tianyu;CHEN Han;WANG Gang;WU Ning(Faculty of Electronic and Information Engineering,Xi’an Jiaotong University,Xi’an 710049,China;The Key Laboratory of Biomedical Information Engineering of Ministry of Education,Xi’an Jiaotong University,Xi’an 710049,China;School of Life Science and Technology,Xi’an Jiaotong University,Xi’an 710049,China)

机构地区：[1]西安交通大学电子与信息学部,西安710049 [2]西安交通大学生物医学信息工程教育部重点实验室,西安710049 [3]西安交通大学生命科学与技术学院,西安710049

出　　处：《西安交通大学学报》2022年第9期104-111,共8页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(32071372);陕西省重点研发计划资助项目(2021GXLH-Z-066)。

摘　　要：针对睡眠生理信号采集难度大、睡眠分期精度低的问题,提出一种采用小波变换和双向长短期记忆网络的脑电睡眠分期模型。首先使用连续小波变换提取睡眠脑电的时频图;然后使用卷积神经网络从脑电信号的时频图中提取睡眠相关的脑电特征,作为单个睡眠片段的分期依据,再使用双向长短期记忆网络进一步提取睡眠片段之间的状态转换规则;最后利用深度学习方法建立特征、规则与睡眠阶段的映射,使用数据扩充和两步训练法训练模型,削弱数据不均衡的影响,完成连续片段的睡眠分期。采用SHHS公开数据库的5793名被试者的睡眠脑电数据对该模型进行验证,实验结果表明,睡眠分期准确率达到85.82%,整体F1达到78.39,Kappa系数达到0.799,和现有方法相比性能明显提升。For the difficulty in acquisition of physiological signals and low accuracy of sleep staging,an EEG sleep staging model using wavelet transform and bidirectional long short-term memory network is proposed.First,the time-frequency map of raw sleep EEG is extracted by using continuous wavelet transform.Then,the sleep-related EEG features are extracted from the time-frequency map through the VGG convolutional network as the staging basis for a single sleep segment,and the transition rules of sleep state are further extracted by the bidirectional long short-term memory network.Finally,the deep learning method is used to establish the mapping of features,rules and sleep stages,and the data augmentation and two-step training methods are used to train the model to weaken the influence of data imbalance and complete the continuous sleep staging work.The model is verified by the sleep EEG data of 5793 subjects from the public database SHHS.The experimental results show that the sleep staging accuracy of the model reaches 85.82%,the F1 reaches 78.39,and the Kappa index reaches 0.799.Compared with existing methods,the performance of the proposed model is significantly improved.

关 键 词：睡眠分期 脑电信号 连续小波变换 卷积神经网络 双向长短期记忆网络 

分 类 号：TN911.6[电子电信—通信与信息系统]