| Depression is a psychiatric disorder that extremely affects physical and mental health,and in serious cases,self-harm and suicidal behavior may occur,so it is of great practical significance to identify and screen for mild depression.At present,the diagnosis of depression is mainly made by patients filling out self-assessment scales and clinical interviews with doctors,which has problems such as large artificial interference factors,lack of scientific quantification,poor specificity,and inaccurate results of single tests.EEG signal is a direct reflection of the neural activity of the brain and is closely related to the emotional state of a person,which becomes an objective and reliable way to assess depression,and EEG has the advantages of high temporal resolution,relatively low cost and no harm,etc.More and more researchers use EEG data to build machine learning models to accurately classify depression and normal people,providing an effective means for depression screening,which can save a lot of medical The social value and market benefits are huge,as it can save a lot of medical resources and reduce the risk of depression and serious illness.However,machine learning for depression identification has problems such as unclear information of depression EEG features,associated electrodes,large individual differences and low identification accuracy.The paper designs machine learning algorithm models based on resting state and evoked state EEG signals to study machine recognition methods for mild depression,aiming to explore depression-related EEG features and effective classification algorithms.The main research of the thesis includes:(1)To address the problem of unclear EEG features of resting-state depression and low accuracy of machine recognition,we extracted EEG time-domain,frequency-domain,and time-frequency nonlinear features closely related to depressive mood,and explored an effective method to improve the accuracy of machine recognition of mild depression.Based on the resting state depression EEG data,the EEG signals recorded from each lead electrode were preprocessed to calculate the EEG activity,mobility,and complexity of depressed patients and normal people;the Burg algorithm and wavelet transform were used to extract the frequency domain features and time-frequency nonlinear features of each EEG signal;the support vector machine machine learning algorithm was designed for depression EEG classification.The effects of different time windows,different feature combinations,EEG lead combinations,rhythm combinations and machine classification algorithms on depression recognition results were experimentally investigated.The experimental results show that the best time window is 20 s,the leads are O2,T5 combination,the features are EEG activity,mobility,wavelet energy entropy,wavelet singular entropy combination,and the rhythm is alpha,beta,gamma combination,the accuracy,recall and precision of depression recognition can reach 94.24%,92.35% and 96.23%.This method can effectively improve the correct classification rate and model generalization ability of resting state depression EEG.A convolutional neural network improvement model based on feature fusion and attention mechanism is proposed for the problems of electrode less,individual differences and evoked state depression EEG recognition.Based on the evoked depression EEG signal,a multicore CNN model is designed by embedding the attention mechanism to fuse the features extracted from the convolutional layers to enhance the diversity of features;the EEG spatiotemporal feature maps are extracted to reduce the influence of individual differences.Through model ablation experimental analysis,we optimize the network architecture and explore the effects of different number of convolutional layers,different feature fusion methods,attention mechanism and adaptive callback module on the network model performance.In addition,for wearable devices to monitor and screen depression need to reduce the number of EEG recording electrodes,the EEG electrodes were optimized using convolutional layer feature map visualization to obtain EEG differential electrodes for depressed and normal subjects,and effective identification of depression with fewer electrodes was achieved.The experimental results showed that the average recognition accuracy of the improved deep learning model for depression using EEG gamma rhythm reached 99.39%,which was 1.43% higher than the recognition accuracy without feature fusion;the classification accuracy,recall,and precision of depression recognition using electrode-less EEG were 91.41%,92.11%,and 92.27%,respectively,and the model test The variance of the results was 4.32E-05,which had good stability.This shows that the improved deep learning model can effectively identify and screen induced state mild depression.Based on the EEG signal feature extraction and machine learning algorithm model,the paper realizes the effective recognition of mild depression,and provides new ideas and methods for the diagnosis and screening of depression.The research results have important scientific and application value in the fields of mental health,disease diagnosis,brain cognition,artificial intelligence,etc. |
