Automatic Sleep Staging Fusion Network Research Based On Transfer Learning

With the growth of sleep disorders,there has been a rapid development of sleep detection devices and automated sleep staging techniques.The EEG signal contains a large amount of physiological information and has been used as an important basis for automatic sleep staging.At the meantime,the development of deep learning technology has also provided important technical support for sleep wave feature learning and automatic sleep staging applications.However,EEG signals are affected by the surrounding environment containing a lot of noise,and the variability of EEG signal acquisition instruments leads to no uniform standard of data.This causes great interference to the intelligent analysis of EEG data.In addition,due to the weak interpretability of current deep learning,single feature learning cannot describe the complex and variable EEG signal features,which ultimately leads to unsatisfactory classification results.In order to solve the above problems,this paper conducts research from both knowledge transfer and multi-feature fusion,and proposes a transfer learning-based automatic sleep staging fusion architecture method,whose main research content is as follows.1.A transfer learning-based deep neural network sleep staging method is proposed to address the problem of insufficient EEG data for sleep monitoring and channel mismatch between different EEG datasets.The study was conducted using two neural networks commonly used for sleep-edf classification,Seqsleepnet and Deepsleepnet.The model parameters were first trained on the source domain large dataset Montreal Archive of Sleep Studies(MASS)to initialize the upper layer network parameters.Then,a sub-dataset sleepcassette of the target domain sleep-edf expanded was applied to design four different strategies of fine-tuning the target network for the network structure.Finally,the effectiveness of the transfer strategies was verified through five sets of controlled experiments.The experimental results show that the method in this paper can be well adapted to accomplish knowledge transfer in different EEG signal channel environments and can effectively improve the accuracy of sleep-cassette on small samples.2.In response to the problems of weak interpretability of deep learning networks and inadequate representation of EEG signals by a single neural network,a multi-network fusion method for sleep fractionation is proposed.This method combines Seqsleepnet and Deepsleepnet networks to extract EEG time-frequency features using Fast Fourier Transform,and combines attention mechanism and long-short term model to learn the features of each stage of sleep.The CNN network was also used to extract the sleep time-frequency features at different scales from the raw EEG data,and a LightGBM classifier was designed to classify the mixed features to obtain the final classification results.In this study,six experimental groups were designed and two sets of data samples were used for validation.The experimental results show that this method improves the classification accuracy by2.4%-4.94% for the sleep data of healthy individuals with SC and 70.04% for the data set of patients with ST sleep disorder compared to other comparative algorithms.