A Study On Mild Depression Recognition Based On EEG Signals Using Deep Neural Networks

Studies based on Electroencephalography(EEG)focus on the use of data mining methods for depression recognition,while those on mild depression are yet in infancy,especially in effective monitoring and quantitative measurement.For the recognition of mild depression,this study proposed two computer-aided detection(CAD)models using Convolutional Neural Network(CNN).This study used EEG data from 24 mildly depressed and 24 healthy control college students during the emotional face browsing task.In the first model,a deep neural network mild depression recognition model based on EEG power features,inspired by the great success of CNN in the field of image recognition,the model converted EEG time-series signals into RGB three-channel image form as the input of deep neural networks.EEG signals contain spectral,spatial,and temporal information.The spectral features are usually studied using spectrogram of the signal,so the power features of the EEG are extracted in this study.The spatial information of the EEG is reflected by the electrode positions,which are distributed in three-dimensional space.For the convenience of processing,the Azimuthal equidistant projection(AEP)is used to project the three-dimensional electrode position onto a two-dimensional plane to form an image to retain the spatial information of the EEG.The EEG time-series signal reflects the temporal evolution of brain activity.To this end,this study divided a trial into multiple frames with a length of one second,and explained the time evolution of the EEG signal through the sequence of frames.In order to explore the role of different aspects of EEG information in the recognition of mild depression,different input forms were designed for the recognition model: the EEG power is organized into a feature vector form to explore the role of EEG spectral information.The power of the corresponding electrode is filled into the electrode projection image,so that it maintains the spectral and spatial information,and compared it with the feature vector form to observe the role of the EEG spatial information.Organize the above two input forms respectively in trial form and frame form to compare and observe the effect of EEG temporal information.In addition to building CAD systems,brain functional imaging has also been very effective in the studies of depression.The calculation result of the functional connectivity metric used in brain functional imaging research is a two-dimensional grid data structure,and CNN has unique advantages in processing grid data.Therefore,this paper built a second model,a convolutional neural network mild depression recognition model based on functional connectivity matrix.Before the model was built,this paper investigated whether the brain functional connectivity of mild depression showed abnormal patterns like those in patients with depression by using the graph theory method widely used in brain functional imaging research.This paper calculated four types of functional connectivity metrics: coherence,correlation,phase locking value,and phase lag index,to explore the mode of brain functional connectivity of mild depression from the aspects of characteristic path length,clustering coefficient,and small-world network properties.Furthermore,CNN is applied to four functional connectivity matrices from five EEG bands(delta,theta,alpha,beta and gamma)to build a novel mild depression recognition model.Through the first model,this paper found that the spectral information of the EEG signals plays a major role in the recognition of mild depression,the spatial information cannot improve the recognition accuracy,while the temporal information of EEG provides a statistically significantly improvement to the recognition accuracy.The first model provided an accuracy of 85.62%,which can effectively recognize mild depression and normal controls.It is found that the brain functional network of the mild depression group has a longer characteristic path length and a lower clustering coefficient,showing a deviation from the small world network,through graph theory.The second model achieved an accuracy of 80.74%.All accuracy was evaluated using the 24-fold cross-validation to ensure that the training and test sets are divided by the participants in this study.This study showed that deep neural network may be the key to solving the early diagnosis of diseases such as depression,and it can help doctors to quickly and accurately diagnose mild depression.