Research On Neural Network For Insomnia Detection Based On Time-Space Characteristics Of EEG Signals

Insomnia disorder is a kind of sleep disorder.Patients have trouble in getting to sleep or keeping sleep.Clinical diagnosis is based on patients’main complaint,scale evaluation and polysomnography(PSG).The first two items are easy to be misdiagnosed due to subjective factors.PSG collects physiological signals such as sleep EEG,which is more objective but requires doctors to be professional.Using machine learning method based on EEG signals collected by PSG to design an objective,efficient and convenient automatic detection algorithm of insomnia disorder can improve the diagnosis efficiency.The existing insomnia disorder detection algorithm includes two steps:sleep stage and insomnia disorder recognition,which has the error propagation problem.From the application scenario,there are individual differences in EEG signals of different person.The existing algorithm lacks the research of universality.Based on these disadvantages,this paper works as follows:(1)In order to improve the shortcomings of two-stage algorithm,two kinds of neural networks based on the time-space characteristics of EEG signals are designed to realize the direct detection of insomnia.CNNBiLSTM constructs the feature matrix according to the electrode position,extracts the high-level features of spatial dimension through CNN,and then extracts the timing information by BiLSTM.GCN-BiLSTM constructs brain function network in light of adaptive threshold and the Euclidean distance between eigenvectors to reflect the spatial connection of brain.GCN is used to extract spatial dimension features,and BiLSTM is used to extract timing information.In the single-subject experiment,the performances of the two models are slightly higher than two-stage method,which shows that two models are feasible to detect insomnia directly.In the cross-subject experiment,the performance of both models decreased,indicating that the universality of the model needs to be improved.However,the performance of GCN-BiLSTM is better than CNN-BiLSTM in single-subject experiment and cross-subject experiment,which shows that the connection between vertices and edges in brain function network can reflect the influence of brain interval.(2)Because of the individual differences of EEG signals,the model is not universal in new subject data in practical application.Transfer learning is applied aim to enhance the cross-subject performance of CNN-BiLSTM and GCN-BiLSTM.As a part of the loss function,the multi kernel maximum mean discrepancy aligns characteristic distribution of the two domains.Then add center loss to improve the classification ability of the mode.The performance of the improved two algorithms is greatly improved in the cross-subject experiment.(3)The existing research is based on the sleep EEG collected by PSG.The collection work is time-consuming overnight,and the subjects’unconscious body movements at night will loosen the electrode and reduce the data quality.According to the medical literature,in the awake and sleep stages,patients with insomnia show the increase of EEG high-frequency activity and the decrease of EEG low-frequency activity.Based on the awake resting state EEG collected in cooperation with the Hospital of Beijing University of Posts and Telecommunications,we retrain the models in work(1)and work(2),and then test the models’ performance of single-subject experiment and cross-subject experiment to explore the feasibility of detecting insomnia disorder in the awake state of subjects,which can make the diagnosis faster.