Research On The Fusion Model Of EEG And Speech Signal Multi Classifiers For Depression Recognition

Depressive disorder is usually called melancholia or depression.Its clinical manifestations are low mood,decreased interest and persistent pessimism.Nowadays,depressive disorder has caused serious damage to patients’ psychological and physical health,which has become one of the main causes of mental illness and disability in the world.In the medical treatment,the diagnosis of depressive disorder mainly depends on the clinical observation and questionnaire of specialists.However,because these diagnosis methods are easy to be forged and concealed,the clinical diagnosis rate is closely related to the cooperation degree of patients.Therefore,to find an objective and effective method to detect depression has become a research hotspot.Patients with depression will be accompanied by changes in cognitive function and behavior of the brain.As an objective physiological index,EEG signal can reflect the state of the brain in different physiological or pathological conditions through its rich time-frequency information.At the same time,speech as a kind of undisturbed behavior index can be used to reflect the clinical mental state of patients.These two kinds of signals have the advantages of fast acquisition,convenience and no invasion of human body.Therefore,in this paper,we consider the study from two aspects of physiology and behavior at the same time and make use of the complementary information between physiology and behavior to make the detection of depression more objective,effective and convenient.In addition,some studies have pointed out that there is no general classifier model,but the performance of the model can be improved by complementing multiple classifiers.In order to improve the recognition accuracy of early depression and assist doctors in the diagnosis of early depression,a multi-modal data fusion model is constructed by using a multi classifier system.The main research work and achievements of this paper include:1)Construction of multimodal physiological and behavioral data set: First of all,based on the experimental paradigm designed in the early stage of the laboratory,this study collects the voice data with emotional tendency and the EEG data of the subjects in the state of resting and closed eyes,among which the control group was matched according to the clinical scale,gender,age,education and other sociological information,so as to reduce the interference of additional factors as much as possible,collect and construct The physiological and behavioral data set of depression subjects needed for the follow-up experiment was established.Finally,170 subjects(81 patients with depression and 89 normal controls)extracted the effective signal characteristics under the corresponding mode.2)This paper proposes a method of depressive disorder recognition based on dynamic classifier selection strategy: firstly,the collected multimodal features are fused in the feature layer,and a new feature vector is constructed through feature splicing to form a multimodal feature space.In this feature space,a multi classifier system fusion model is constructed,in which a dynamic classifier selection strategy is used to select a suitable set of sub classifiers for each sample to be tested to build a depression recognition model.After many experiments,the results show that the introduction of multi-modal information can improve the model’s recognition of depression,and the average accuracy of the multi classifier system fusion model is up to by 73.2 %,Moreover,the introduction of dynamic classifier selection strategy improves the accuracy of model recognition and improves the stability of depression recognition model.3)This paper proposes a method of depressive disorder recognition based on multiagent strategy: The sub classifiers are trained independently for the features of different modes.Analogy the sub classifiers as agents,and the multi-agent strategy is used to build the multi classifier system fusion model.In the interaction process,agents use decision co-occurrence tensor to adjust their classification results,to achieve the purpose of using low-order correlation information of sub classifier.After many experiments,the results show that the accuracy,F1 score and sensitivity of the proposed multi-agent fusion strategy are better than those of single-mode classifier or other traditional typical multi classifier system fusion strategies.The accuracy of the strategy is76 % and the sensitivity index is 71 %,which is significantly improved compared with other methods.The experimental results show that the classification performance of the model will be significantly different in the case of gender differences.In this paper,the multi-agent fusion strategy is used to identify multimodal depression.The sub classifier effectively alleviates the impact of gender and improves the generalization performance of the model through information exchange under the multi-agent strategy.In conclusion,through the collection and utilization of patients’ physiological and behavioral information,it is helpful to recognize and construct the recognition model of depression from multiple perspectives.In addition,this paper uses the multi classifier system to model the multi-modal data fusion,making full use of the difference between modes and the complementarity of sub classifiers,effectively improving the accuracy and robustness of the depression recognition model,and providing a new idea for the detection of depression.