End-to-end Automatic Sleep Staging Research Based On Deep Learning

Current available sleep electroencephalogram data sets for sleep staging are all class imbalanced small data sets,which is hard to achieve the ideal staging result by direct migration application of deep learning models.A deep automatic sleep staging model for class imbalanced small data sets was proposed,from the aspect of data oversampling and model training optimization.Its main contents and results are as follows :(1)A deep end-to-end automatic sleep staging model was Proposed.Using the original sleep EEG dataset of the Sleep-EDF database Fpz-Cz channel to perform a 15-fold cross validation experiment on the model,the overall accuracy and Macro-averaged F1 score are 86.73% and 81.70%,respectively.Compared with the classification results obtained from different references using the same dataset set in recent years,the advancement of the proposed automatic sleep staging model was verified.(2)Based on the Modified Synthetic Minority Oversa mpling Technique(MSMOTE),from the perspective of reducing decision domains,only the safety class data in minority classes were augmented,and the method for data generation was improved.A Dimension Modified Synthetic Minority Oversampling Technique(DMSMOTE)was proposed.Experiments show that the application effect of DMSMOTE in this model is better than that of MSMOTE..(3)In order to reduce the negative impact of class imbalance on the classification of minority classes in the original data set.Considering that the reconstructed class balanced dataset will contain data that does not belong to any classification,using the class balanced dataset which was reconstructed by DMSMO TE to pre-activate the model,the original EEG datase t was used for fine-tuning the model.Experiments show that the pre-activation of the model using the class-balanced dataset generated by DMSMOTE can increase the minimum class classification F1 from 45.16% to 53.64%?The research results show that the model can realize end-to-end learning of a small amount of raw sleep EEG data set,and the overall classification effect is better than the high-level model in recent years;given its end-to-end nature,compared with the traditional feature-based machine learning model,The model is more suitable for the portable sleep monitors that work in conjunction with background servers.