基于多层级卷积融合网络的自动睡眠分期方法

Automatic Sleep Staging Method Based on Multi-Layer Convolution Fusion Network

针对原始数据不均衡、多层级特征利用不充分等原因造成的基于脑电(EEG)信号分析的睡眠分期方法精度长期驻足不前的问题,提出了一种多层级卷积融合网络的睡眠分期方法,在以SMOTE算法对原始EEG数据进行均衡化处理,并构建时-空信息特征矩阵对网络的输入量进行二维化处理的基础上,通过搭建和优化不同深度的卷积网络,自动提取并融合多层级、多尺度的EEG信号睡眠特征,实现睡眠分期。实验结果表明:所提方法在Sleep-EDF数据集上的分期精度可达到92.35%,宏F1-score达到84.4%,分期精度最低的N1期F1-score可达到53.3%,睡眠分期性能优于其他深度学习模型。

A multi-layer convolutional fusion network is proposed,targeting at the long-term lack of accuracy of sleep staging methods based on EEG signal analysis due to unbalanced raw data and insufficient utilization of multi-layer features.The sleep staging method is based on equalizing the original EEG data with the SMOTE algorithm,and constructing a spatio-temporal information feature matrix for two-dimensional processing of the network input.By building and optimizing convolutional networks of different depths,automatic extraction and fusion of the sleep features of multi-layer and multi-scale EEG signal is realized to achieve sleep staging.The experimental results show that the staging accuracy of the proposed method on the Sleep-EDF dataset can reach 92.35%,the macro F1-score can reach 84.4%,the N1 stage F1-score with the lowest staging accuracy can reach 53.3%,and the sleep staging performance is better than other deep learning models.