摘要
传统上,自动睡眠分期是一项非常具有挑战性且费时费力的任务。大多数现有的自动睡眠分期方法都基于单通道的脑电(electroencephalography, EEG)数据,然而,这些方法忽略了医师从整体上观测多个通道EEG信号进行睡眠阶段评分的过程。为了解决这一问题,我们优化了数据结构,对医师的评分过程进行了详细的学习与建模,提出了一种基于多通道脑电图的自动睡眠评分方法。我们介绍了在原始EEG与EOG样本上使用深度卷积神经网络(convolutional neural network, CNN)进行睡眠阶段评分的监督学习。该网络具有11层,每30 s的睡眠数据作为一个分期,并且不需要任何信号预处理或特征提取。本文使用来自福建省某医院的EEG与EOG及专家评估的多导睡眠图(polysomnography, PSG)数据对系统进行训练和评估。实验结果表明,在自动睡眠分期的研究中不应该忽略EOG数据。我们的系统性能与中级睡眠分期专家的结果相当。
In the field of medical informatics,the automatic sleep staging is achallenging and time-consuming task,and most existing automatic sleep stagingmethods are based on single channel electroencephalography(EEG)data.However,these methods ignore the physician’s overall observation of multiple channelEEG and EOG signals for the sleep stage scoring.To resolve this problem,wepropose an automatic sleep scoring method based on multi-channel EEG,includingthree-channel EEG and two-channel Electrooculogram(EOG)data.We introduce theuse of a deep convolutional neural network(CNN)on raw EEG samples forsupervised learning of sleep stage prediction,which does not require anysignal preprocessing or feature extraction.We use the EEG and EOG ofpolysomnography(PSG)data which have been assessed by medical expert from aHospital of Fujian Province to train and evaluate our system.Comparing withthe staging result with single-channel EEG data,we indicate that the EOG datashould not be ignored for a better sleep staging.It shows that the performanceof our system is comparable to that of mid-level experts.
出处
《生物物理学》
2019年第2期11-25,共15页
Biophysics
基金
国家自然科学基金资助项目(批准号:11874310,11675134)
国家111项目(批准号:b16029)。
