Research On Depression Patients Recognition Algorithm Based On Deep Learning

As a common mental illness,major depressive disorder(MDD)has had a serious negative impact on people’s daily life.Early detection and treatment is critical to the cure of MDD.However,the current scale-based MDD diagnosis methods have strong subjectivity and low accuracy.Therefore,we seek an objective and efficient method to identify MDD patients.With the help of the powerful representation learning ability of deep learning(DL),this paper taken the electroencephalogram(EEG)signals of MDD patients and normal controls(HCs)as the research object,and carried out the identification of MDD patients from two aspects: the research of DL algorithm based on self-attention mechanism and the research of DL algorithm based on spatio-temporal convolutional neural network(CNN).The main work content is summarized as follows:1)The DL algorithm based on the self-attention mechanism was proposed to identify MDD.Considering that the self-attention mechanism has the ability to capture the correlation between data,this method firstly adopted the multi-head self-attention mechanism to automatically learn the potential connectivity relationships among EEG channels.Then,the learned connectivity matrices were input into the parallel two-branch CNN module for deeper and more complex feature extraction.Finally,the feature information learned by the two branches was fused and input into the dense layer for the MDD classification.Compared with the DL methods based on traditional brain connectivity matrices,the proposed method achieved better classification performance using the leave-one-subject-out(LOSO)cross-validation method,and its average classification accuracy was 91.06%.2)The DL algorithm based on spatio-temporal CNN was proposed to identify MDD.Based on the temporal and spatial differences of EEG signals between MDD patients and HCs,this method firstly used the designed end-to-end two-branch CNN model to automatically mine the spatio-temporal feature information of EEG signals in the temporal and spatial dimensions.Then,the spatio-temporal feature information was connected and input to the dense layer for MDD classification.The model also adopted the LOSO cross-validation method,and reached an average classification accuracy of 92.81%.In addition,this paper also explored the ability of five sub-bands(delta,theta,alpha,beta and gamma)to identify MDD patients on two DL models.The experimental results show that delta,beta and gamma bands have the potential to be biomarkers of MDD detection.In conclusion,the DL methods proposed in this paper provide effective solutions for MDD identification,and further verify the feasibility of DL application in this field.