Abstract: The classification accuracy of current automatic sleep staging is determined by the small data set of imba-lanced classes and hand-engineered features.Aiming at this problem,this paper proposed an automatic sleep staging model based on deep hybrid neural network.For the construction of model’s main structure,the multi-scale Convolutional Neural Networks are used to automatically learn the high-level time-invariant features,the Recurrent Neural Networks constructed by bidirectional Gated Recurrent Unit are used to decode the temporal information from the time invariant features,and the residual connection is used to fully combine the time invariant features with the time information features.For model optimization,in order to reduce the impact of the dataset of imbalanced class on the classification effect of minority class,the experimental data set reconstructed by MSMOTE (Modified Synthetic Minority Oversampling Technique) is used for pre-training.The Swish activation function is used to accelerate the training convergence rate.The experiment was set up on the initial single-channel EEG signal of Fpz-Cz in Sleep-EDF Database.The 15-fold cross-validation experiments show that the overall classification accuracy is 86.85% and the Macro-averaged F1-score is 81.63%.This model can effectively avoid the subjectivity of feature selection and the limitation of class imba-lanced small dataset of imbalanced class in deep learning.

Key words: Deep learning, End to end, Gated recurrent unit, Single-channel, Sleep staging, Swish