Data Augmentation Of Raw EEG Signals And Its Application In Depression Detection

At present,the incidence of psychosocial diseases represented by depression is increasing.Due to social neglect,pressure and shortage of medical resources,the current mainstream depression detection methods cannot diagnose and treat depression in a timely and effective manner.Therefore,a more objective and effective method is urgently needed for the detection of depression disorder.Electroencephalograph is an indicator that uses electrical signals to record various activities of the cerebral cortex.It can intuitively reflect various activities of the human central nervous system,and is intuitive and difficult to disguise.In recent years,the research on the detection of depression based on EEG signals has been increasing.However,in the combination of EEG signal and machine learning,due to the complexity of the collection method and the few diseased samples,the EEG signal data set is smaller than the image and speech data set,which will have a negative impact on the performance of the machine learning model.In addition,the characteristics of EEG signals and their unique individual differences hinder the application performance of traditional data augmentation methods in EEG signals.Therefore,how to find a more effective and more suitable augmentation method for the characteristics of the EEG signal from the raw EEG signal is an important prerequisite for constructing a depressive disorder diagnosis and recognition model with stronger generalization performance and superior performance.Taking into account the characteristics of the raw EEG signal's time series signal and its own individual differences,it is very necessary to find a data augmentation method suitable for the raw EEG signal.This article starts with the characteristics of the raw EEG signal,explores a more diversified and more efficient augmentation method based on the raw data set,and extends it to Generative Adversarial Networks,Applying this algorithm to the objective detection of depressive disorder.The experiment obtained a relatively superior diagnostic effect.The main work and achievements of this paper are summarized as follows:1)This paper proposes a data augmentation method dp EMD combining Determinantal Point Processes(DPP)and Empirical Mode Decomposition(EMD).Use the DPP algorithm to mine the individual difference information of the raw EEG signal,and integrate the difference information into the participant selection of the EMD algorithm.Through the DPP algorithm,the augmented subjects can more effectively combine the characteristics of multiple subjects in the raw EEG signal data set.This method increases the diversity of the augmented EEG signal by fusing the individual difference information of the original EEG signal,thereby effectively improving the performance of the EEG signal in the depression detection model.2)This paper proposes a Determinantal Point Processes Generative Adversarial Networks(dp GAN)algorithm.By introducing the kernel matrix of determinant point process algorithm,the loss function of generative adversarial network is improved.The individual difference information of the raw EEG signal is introduced into the network,and the diversity of the generated augmented EEG signal data is controlled by simulating the distribution of the raw EEG signal.This method can obtain better results in the augmentation of the raw EEG signal data set with a small sample size.Compared with other original EEG signal augmentation methods,dp GAN can generate higher-quality augmented EEG signals,thereby effectively strengthening the detection ability of the depression detection model.3)This article uses the dp EMD and dp GAN augmentation method on the data of170 subjects(81 depressed patients and 89 normal controls)to generate relevant augmented data.Use the augmented EEG data and the raw EEG data to build a model,and then compare the performance of the model to verify the hypothesis.The results show that the experimental results after using the dp EMD augmentation method and dp GAN augmentation method to augment the raw EEG signal are better than other commonly used raw EEG signal augmentation methods.Among them,compared with other commonly used raw EEG signal augmentation methods,the diagnosis accuracy of the depression diagnosis model constructed by dp EMD is 5-9% higher.It proves that the dp EMD augmentation method has achieved a certain effect on the original EEG signal augmentation method.However,since the EMD algorithm essentially recombines the characteristics of the raw EEG signal,its contribution to the detection model of depression in terms of the amount of information is not as great as the contribution of the newly augmentation EEG signal data generated by the GAN.The performance improvement of dp EMD augmentation method to machine learning method is not as good as that of dp GAN augmentation method.In the small sample size augmentation experiment,dp GAN can achieve better results.Under the same number of raw EEG signals,the detection accuracy of dp GAN method is 2?7% higher than that of dp EMD,which proves that this method is effective in the augmentation of raw EEG signals.Therefore,dp GAN has a better effect when processing small sample data set augmentation,and can solve the problem of too little data in the raw EEG data set.In summary,the raw EEG augmentation method that incorporates the DPP algorithm proposed in this article can deeply explore the individual difference information among EEG subjects with depression,and generate effective augmented data for the identification of depression.This means that it is necessary to study the diverse sampling methods of raw EEG signals to further improve the robustness and generalization of the classification model,which provides a new idea for data augmentation methods for raw EEG signals.