融合多种网络的半监督分层睡眠分期算法

Semi-supervised EEG Sleep Staging Algorithm Based on Hybrid Neural Network

当前的睡眠阶段时空特征提取依赖于给定数据集标签的监督学习,在一定程度上受到限制。提出一种基于脑电信号的半监督睡眠分期算法,利用由改进卷积编码-解码器和生成对抗网络构建的浅层特征提取网络提取浅层时空特征,采用Hard swish激活函数来加速模型收敛。为充分提取脑电信号高质量的深层时序依赖特征,模型的深层特征提取网络将传统的长短时记忆网络改进为参数较少的双向门控循环单元。在特征融合后使用加权交叉熵损失函数训练以提高模型的分类准确性。实验使用Sleep-EDF数据集在Fpz-Cz通道上对模型进行20折交叉验证,得到模型总体准确率和MF1值分别为86.3%和81.2%,相比于卷积循环网络分别提高了3.1%和3.3%。

Current spatial-temporal feature extraction of sleep phases relies on supervised learning of labels from given datasets,which is limited to a certain extent.A semi-supervised sleep staging algorithm based on EEG signals is proposed.The shallow feature extraction network constructed by improved convolutional codec and generative adversarial network is used to extract the shallow spatial-temporal features,and the Hard Swish activation function is used to accelerate the convergence of the model.In order to fully extract the high-quality deep time-dependent features of EEG signals,the deep feature extraction network of the model improves the traditional long short term memory network into a bidirectional gated recurrent unit with fewer parameters.After feature fusion,the weighted cross-entropy loss function is used to train the model to improve the classification accuracy.The experiment uses Sleep-EDF dataset to perform 20-fold cross-validation on Fpz-Cz channel,and the overall accuracy and MF1 value of the model are 86.3%and 81.2%,respectively,which are improved by 3.1%and 3.3%compared with the convolutional recurrent network.