Depression Recognition Based On Deep Learning And EEG Signals

Depression is a mental illness that negatively affects people’s behavior,emotions,and thinking abilities.With the accelerated pace of social life and the increase in daily stress,the incidence of depression shows an upward trend.It is one of the main causes of disease burden in today’s society,and it is expected to rise to the first place in the world by 2030.Early and accurate diagnosis and timely treatment of depression are particularly important.At present,the diagnosis of depression lacks objective criteria,mainly based on doctors’ clinical interviews and psychiatric questionnaires.This diagnostic method is affected by the clinical experience of doctors and the authenticity of patients’ self-description,and has strong subjectivity.Finding an objective and effective quantitative diagnosis method for depression has become a scientific problem to be solved urgently in the medical field.Changes in mental state such as depression can cause changes in brain nerve activity.Electroencephalogram(EEG)is the most direct manifestation of brain nerve activity,closely related to human emotional state,and it is also the physiological signal with the greatest potential to be applied to the evaluation of emotion and mental state.Functional magnetic resonance imaging studies have shown that depression can lead to lesions in the brain structure and function of patients.It has become a hot issue in the study of psychological diseases such as depression to explore the changes in the mechanism of brain nerve through relatively convenient EEG.At present,researchers have studied the computer-aided depression detection method based on EEG signals,and their thinking is usually to extract the relevant features of EEG signals and send them to the traditional machine learning classifier.This kind of method requires to extract features manually,and it may be difficult to find effective features.In recent years,deep learning has become more and more popular in the medical field due to its automatic decision-making ability.It can overcome the shortcomings of the previous methods,and automatically and quickly learn from the samples to multi-level abstraction and representation.Based on the resting frontal lobe three-electrode EEG data collected from 40 patients with depression and the same number of healthy subjects,this study explored an accurate and efficient method for identifying depression through deep learning technology for different application scenarios,and provided theoretical and technical support for computer-aided diagnosis of depression.The main work of this thesis is as follows:(1)Complete the experimental data acquisition and the data preprocessing.In this study,the data were collected from 80 subjects,and the noise components in the signals were removed using wavelet decomposition and reconstruction.Then the signal is segmented into EEG samples by data enhancement to increase the sample size.(2)A one-dimensional hybrid model combining convolutional neural network(CNN)and long short-term memory(LSTM)is designed for depression recognition of single-channel EEG signals.The CNN and LSTM parts of the model are used to learn the local features and sequence features of the signal,respectively.Based on this model,the recognition effects of different electrodes and different frequency bands of the same electrode were explored in this study.In order to further verify the effectiveness of the model,with the help of the idea of ablation research,two sub-parts of the model were taken as comparative models and their performances were observed.The highest classification accuracy obtained by this model is 90.13%,and the model input is the full-band signal of Fz channel.(3)A two-dimensional spatio-temporal convolutional model based on EEGNet is established for depression recognition of multi-channel EEG signals.The convolution operation of the model includes temporal convolution,spatial convolution and separable convolution.In which the temporal and spatial convolutions extract the time sequence features of the EEG signal and the spatial information on different electrodes,respectively.Spatial convolution uses deep convolution,which utilizes different convolution kernels for different feature maps,thus the model can learn specific spatial information from specific feature maps.In this thesis,the comparison experiment of different models and the optimization experiment of model parameters are designed.The results show that the depression recognition model based on EEGNet has the highest average classification accuracy of 93.74%,and the highest classification accuracy achieved in the parameter optimization experiment is 94.27%.Compared with the one-dimensional hybrid model,the two-dimensional spatio-temporal convolution model based on EEGNet has better performance.(4)The validity of the deep learning models is analyzed with the reduced-dimension visualization method.In this thesis,the dimension reduction visualization is performed on the untrained samples and the feature maps learned from the trained model.The results show that the features learned from the one-dimensional hybrid model and the two-dimensional spatiotemporal convolutional model can be well aggregated towards two clusters.Therefore,the methods in this thesis can extract the differential characteristics from the EEG to distinguish depression patients and healthy subjects,and has great application potential in computer-aided diagnosis of depression.