基于深度学习的多通道脑电信号睡眠分期方法

Multi-Channel EEG Sleep Staging Method Based on Deep Learning Sleep Staging Method Using Multi-Channel Electroencephalogram Based on Deep Learning

目的利用脑电信号进行睡眠分期是分析人睡眠状态的重要方法,通过引入人工智能深度学习来研究基于多通道脑电信号的睡眠分期方法。方法提出基于注意力的多通道脑电信号睡眠分期方法(Attention Based Multi-Channel EEG Sleep Net,AMCSleepNet),该方法利用多个分支卷积网络分别提取脑电信号不同通道的时频域特征,设计新型压缩聚合层和残差层自适应融合多通道脑电信号的特征,并通过Transformer编码器挖掘特征的时间相关性。结果在2021全国智能睡眠科学大赛提供的多通道脑电数据上,AMCSleepNet的平均准确率达到了81.86%,相较于基于注意力的单通道模型和多通道深度卷积模型分别提升了5.69%和11.06%。结论AMCSleepNet方法能够利用多通道脑电信号提升睡眠分期的准确率,可更充分利用睡眠数据,潜在应用价值较高。

Objective Sleep staging based on electroencephalogram(EEG)signal is an important method to analyze human sleep state.This paper introduced artificial intelligence deep learning approach to process the method of sleep staging based on multi-channel EEG signals.Methods An attention-based multi-channel EEG sleep net(AMCSleepNet)was proposed.In this method,multiple branched convolutional networks were used to extract the time-frequency domain features of different channels of EEG signals.The AMCSleepNet designed squeeze-excitation layer and residual layer to adaptively fuse the features of multi-channel EEG signals,and applied a transformer encoder to mine the temporal correlation of the features.Results Testing on the multi-channel EEG data provided by the 2021 National Intelligent Sleep Science Competition,the average accuracy rate of the AMCSleepNet reached 81.86%,which was 5.69%and 11.06%higher than that of attention-based single-channel model and multi-channel deep convolution model respectively.Conclusion The AMCSleepNet method can improve the accuracy of sleep staging by using multi-channel EEG signal,make full use of sleep data,and has high potential application value.