| Sleep staging is an essential step for evaluating sleep quality and diagnosing sleep related diseases.It is a time-consuming,labor-intensive and a bit subjective task for human experts.At the same time,due to the limited knowledge of physiological mechanism,the results of sleep staging are often mutable.Therefore,it has significant clinical value to use artificial intelligence technology to realize automaticity and objectivity of sleep staging.This work focuses on the deep learning method of sleep staging from two perspectives.Main contributions and results are following:(1)In order to explore the influence of different channel data on sleep stages,from the perspective of visualization,a new hybrid neural network model based on hierarchical attention is proposed.In this model,a convolutional neural network is used for single sleep epoch’s representation learning,and a bi-direction recurrent neural network is used for successive sleep epochs’sequence learning.By taking advantages of hierarchical attention,it is convenient to view the importance changes of different channels within a sleep epoch.Visualization of hierarchical attentions from the model,comparing with power spectrum of the original data,then the results show that the model’s channel usage is basically consistent with human cognition.Besides,of the model’s 10-fold cross validation on the public database sleep-EDF,the results also turn out that its performances have ability to compare with the state-of-the-art models in recent years;(2)In order to learn the characteristics of transition between sleep stages and improve accuracies on transition set and non-transition set,a sleep staging method suitable for sequence classification is proposed,which is a sleep staging method with multi-task learning based on sequence learning.By reconstructing the basic loss function,it is a creative behavior for the method to view whether the sleep stage shift or not as a binary classification task,which is the first time to be used in the field of sleep staging to the best knowledge.The baseline models with the method take 10-fold cross validations on the public database sleep-EDF.It is turned out that this method can improve the accuracy and macro FI score by 1.5%and 1.46%respectively.The experiment results prove that the model and the method proposed in this work both have good utility for automatic sleep staging. |
