Abstract: In order to achieve efficient automatic sleep staging, a time-domain sleep signal processing method based on the principle of period segmentation is proposed in this paper. Merger of Increasing and Decreasing Sequences（MIDS） is employed to segment the three-channel polysomnography records（2-channel electroencephalogram and 1-channel electrooculogram） periods, then the characteristic waveforms of each stage of sleep are marked according to the signal period, the proportion of the duration and the average amplitude of the feature waves are extracted as features. Bi-directional Long Short-Term Memory（Bi-LSTM） is used as a classifier to solve the shortcomings of traditional machine learning methods that can not utilize the temporal context information of sleep data. The average accuracy for 42?699 samples is 84.8% by cross-validation, and the merging sequence method can reduce the complexity of EEG signal analysis which is an effective time-domain signal processing method. Bidirectional LSTM neural network can effectively improve the accuracy of sleep staging, and has good application prospects.

Key words: sleep stage, period segmentation, merger increase and decrease sequence, deep learning, Bi-directional Long Short-Term Memory（Bi-LSTM）

摘要： 为实现高效的自动睡眠分期，提出一种基于周期分割的时域信号处理方法，采用合并增减序列方法对三个通道多导睡眠图记录（2路脑电，1路眼电）进行周期分割，根据信号波形的周期标记睡眠各期的特征波形，提取特征波形在每一帧数据的时长占比与平均幅值作为特征。双向长短时记忆网络（Bi-directional Long Short-Term Memory，Bi-LSTM）作为分类器，解决传统机器学习方法无法利用睡眠数据时间上下文信息的缺点。对42?699个样本使用交叉验证方法得到了84.8%的平均准确率，实验结果表明合并增减序列方法可以降低脑电信号分析的复杂度，是一种有效的时域信号处理方法，双向长短时记忆网络可以有效提高睡眠分期准确率，具有良好的应用前景。

关键词: 睡眠分期, 周期分割, 合并增减序列, 深度学习, 双向长短时记忆网络