Recognition Of Depression Disorders Based On Eeg And Speech Signals

EEG as an electrophysiological technology has been widely used in the diagnosis of mental diseases,but the traditional EEG technology needs professional equipment and professional technicians,the whole diagnosis process time is relatively long,it is not easy to carry out large-scale detection,so the research and development of portable and easy-to-use EEG equipment has become a key.In addition,speech signal can also be used as a behavioral indicator to evaluate the mental state of patients with depressive disorder.The combination of speech signal and EEG can realize the complementary information between modes and make the detection results more objective.Therefore,the study of EEG and speech evaluation method provides a theoretical basis for portable,rapid and large-scale depression screening.In this paper,we start from the direction of large-scale and easy-to-use depression detection.First,we study the EEG technology from the perspective of physiology,and then combined with the voice signal for fusion research.The main research work and achievements of this paper include:（1）First,we study physiological EEG.After preprocessing,four linear features and seven nonlinear features are obtained.Through single feature for depression recognition,it is found that the linear features of first-order differential mean,second-order differential mean,standard deviation and fractal dimension have high recognition accuracy.Further analysis of the overall differences between the depressive disorder group and the normal group shows that the four characteristics are significantly different in the prefrontal cortex（mainly E27,E23,E9 and E2）.In addition,by dividing subjects according to gender,age and education level,it is found that differences among different populations still generally appear in the prefrontal lobe region of the brain,which provides a reference for us to choose prefrontal lobe electrodes to simplify 128 guide devices.（2）Combined with the regional differences of prefrontal cortex found by single feature analysis,considering the convenience of wearing portable EEG equipment,we propose to simplify 128 channel EEG equipment by using 7 channels of prefrontal cortex（E34,E27,E23,E18,E15,E9,E2）.The 11 EEG features are fused serially,and the classification result is poor due to the redundancy between the features.In order to solve this problem,three feature selection methods are used to select the fused features.Among them,the accuracy of the feature subset selected by the maximum correlation maximum correlation distance method proposed in this paper is 90%,while the other two methods can only achieve the highest 80%.Considering the influence of the basic feature subset,we try to make the basic feature subset of all features in turn,and then select 19 features from the feature library to form 20 dimensional feature subsets for classification.The average classification accuracy of the feature subsets selected by the maximum correlation maximum Association distance method is 90%,and the maximum value is about 93%.At the same time,it has a larger F1 score.Based on the prefrontal electrode feature fusion and the proposed feature selection method,it provides a feasible scheme for the development of portable EEG devices.（3）The combination of EEG and speech can effectively detect depressive disorder.Mel frequency cepstral coefficient（MFCC）is used to change the scale of MFCC,which is arranged into a matrix with consistent dimensions,and then input into convolution neural network as image features for further feature extraction.The extracted highdimensional features and EEG features are fused serially,and then input into the longterm and short-term memory network,and finally output the results through the full connection layer.In this way,the classification accuracy of EEG and speech feature fusion model based on neural network is up to 87.5% after training.At the same time,artificial feature selection is avoided.It has more potential than single speech signal detection and single EEG signal detection.In conclusion,firstly,using the prefrontal EEG signal and feature subset selection method MRMCD can effectively recognize the patients with severe depressive disorder under the feature serial fusion;secondly,the feature fusion of EEG and speech signals based on deep neural network is more competitive in the recognition of severe depressive disorder.