基于多尺度深度网络的自动睡眠分期

Multi-Scale Deep Network for Automatic Sleep Staging

传统睡眠质量评估与诊断高度依赖医生的经验以及对长时间睡眠监测数据的分析统计,耗时耗力,且传统机器学习技术所实现的自动睡眠分期依赖人工构造的特征,在发掘深层次分期特征上效果有限,对部分分期的辨识效果欠佳。提出一种基于多尺度深度网络(MSDNet)的自动睡眠分期算法,能够自动分析提取睡眠信号特征,并基于不同睡眠阶段的分期难度进行相应的多尺度分析与判别。该算法可提前输出特征简单明确的W、N2、N3、REM等阶段的分类结果,然后让特征相对不够清晰的过渡阶段N1进入更深尺度的网络,利用更精细的特征数据以及更深的网络进行判别,能提高整体分类效率以及N1阶段的分类精度。在Sleep-EDFx数据集中,提取197组样本数据用于训练和测试,并在仅采用单通道EEG脑电信号的前提下,平均分类精度达到0.83,Kappa值为0.749,N1阶段的F1-score达到0.51,较传统机器学习算法以及多种深度网络,提高了整体分类精度,特别是N1阶段精度,且计算量没有显著增加,适于自动化实时分析。

Sleep quality assessment and diagnosis highly depends on the doctors’effort and experience.It is quite labor intensive and time consuming for the doctor to inspect the long-term sleep monitoring records.The current automatic sleep staging mainly uses traditional machine learning and it highly relies on the features designed by experts.However,these features are usually incapable to capture the deep level features hidden in the measured data,and the behave not well for some staging such as N1.This paper proposed an automatic sleep staging algorithm based on multi-scale deep network.It used the deep network to automatically extract sleep signal features and used multi-scale analysis and discrimination criteria relating to the difficulty measure in the classification of different sleep stages.The classification results of stage W,N2,N3 and REM were selected and output in advance based on shallow layer features,and the fallible transition stage N1 entered a deeper network for further analysis.This policy improved the overall classification efficiency and especially the classification accuracy of the N1 stage.When extracting 197 sets of sample data for training and testing in the Sleep-EDFx data set and using only single-channel EEG signals,the average classification accuracy achieved 83%,and Kappa value was 0.749,which indicated that the constructed models were highly consistent.F1-score at stage N1 achieved 0.51.Compared with traditional machine learning algorithms and a variety of deep networks,the overall classification accuracy and accuracy of the N1 stage were improved.And at the same time,there was no apparent calculation increase.It is suitable for automatic real-time analysis.