Research On Recognition Method Of Depression Based On EEG And Graph Neural Network

The rapid development and fierce competition of today’s society have led to the increasing incidence rate of mental diseases such as depression.When patients do not receive timely diagnosis and treatment,their physical and mental health can be severely affected,and may even increase the risk of suicide.However,there is a general lack of scientific understanding of depression,and there are even fewer people seeking timely medical treatment and receiving accurate diagnosis and treatment interventions.For the diagnosis of depression,traditional methods often use scales combined with doctor inquiries,which are easily influenced by the doctor’s experience level and the patient’s emotions during medical treatment,leading to misdiagnosis.Therefore,seeking a timely,objective,and accurate diagnostic method for depression is an indispensable prerequisite for the formulation and effective implementation of treatment plans.The development of Electroencephalogram(EEG)signals provides a new approach for the diagnosis of depression,as a non-invasive diagnostic method with high time resolution,the rules contained in EEG signals can be used as a potential biomarker for diagnosing depression.Deep learning is a powerful tool that can accurately decode EEG signals.Through deep learning methods,real-time diagnosis of depression can be achieved,thereby monitoring the health status of patients.From the perspective of practical clinical applications,a high-performance independent subject based depression recognition model is a necessary condition for the widespread application of EEG signals in the real world.Therefore,cross subject dataset partitioning is more suitable for depression recognition tasks.The individual differences of the subjects pose great challenges.In response to the above background and issues,this article proposes two depression re recognition methods based on EEG signals and graph neural networks,and conducts experimental verification and analysis.The main work of the thesis includes:1.A depression recognition method based on Parallel Spatial Temporal Convolutional Network.This method performs binary classification on depression patients and healthy controls in the MODMA public dataset.Firstly,the dataset is preprocessed,differential entropy and wavelet entropy features are extracted,and stacked for enhancement;Then,the features are input into the spatiotemporal convolution model and hybrid multi-layer perceptron composed of convolution layer and graph convolution layer,and the depth features are extracted,and the test set distribution domain is modified through the maximum mean difference loss function to narrow the distance between the test set distribution domain and the training set data distribution domain,so as to overcome the problem of individual differences;Finally,the features extracted from the two models are spliced and fused,and input into the Multilayer-Perceptron for binary classification.The experiment uses the data set partition method based on cross subjects,and gets the experimental results through six fold cross validation.The experiment adopts a cross subject dataset partitioning method,and obtains experimental results through six fold cross validation.The recognition performance of the model is evaluated from five aspects,including accuracy and recall.The experimental results indicate that compared to the results of comparative experiments and ablation experiments,this method has advantages in accuracy,precision,and other indicators,and has better depression recognition performance.2.A depression recognition method based on Structural Entropy Guided Graph Hierarchical Pooling.Due to the excellent performance of the graph method and feature enhancement in the first method,this method improves the model from these two perspectives.On the one hand,the graph pooling strategy is applied to the graph to deeply mine the spatial information of EEG signals,and the structural entropy minimization algorithm is applied to construct the clustering allocation matrix to improve the performance of the graph pooling layer and alleviate the suboptimal and local information loss problems that have occurred in previous pooling models.This step coarsens the graph composed of each experiment to a feature vector;On the other hand,this method innovatively proposes a method of sharing information between different experiments through graph convolution.This step gathers all experiments in the form of nodes into the same graph,and achieves message transmission between experiments through graph convolution to achieve the effect of feature enhancement.The experimental results showthat this method outperforms the proposed comparative and ablation experiments in terms of accuracy,recall,and other aspects,demonstrating better depression recognition performance.