Self-adaptive graph convolution-based depression EEG (electroencephalogram) multi-level information processing method

Abstract

The invention discloses a depression EEG (electroencephalogram) multilevel information processing method based on adaptive graph convolution. Firstly, differential entropy features of five different frequency bands are extracted from electroencephalogram signals to serve as a feature matrix; constructing a functional connection matrix by calculating Pearson's correlation coefficients among different electrodes, calculating geodesic distances among the electrodes according to three-dimensional space coordinates of the electrodes to form a spatial structure matrix, performing single point multiplication on the functional connection matrix and the spatial structure matrix to obtain a shared mask parameter matrix, and combining the shared mask parameter matrix and the functional connection matrix to obtain an adjacent matrix; and jointly inputting the feature matrix and the adjacent matrix into a graph convolutional network model for training. According to the method, brain network structure information is enriched, a multi-level view angle is provided for detecting depression patients, more important connections for depression detection are recognized in training through the shared mask parameter matrix capable of being automatically updated, the interpretability of the model is enhanced, and the accuracy of depression detection is improved. And a more accurate and reliable solution is provided for identifying the depression patient based on electroencephalogram.

Description

Technical Field

[0001] The present invention relates to the technical field of medical signal processing, and specifically to a multi-level information processing method for depression EEG based on adaptive graph convolution. Background Art

[0002] Depression, as a common and serious mental health disorder, has profound negative impacts on individuals, families and society. It not only erodes the patient's mental health, causing persistent pessimism and loss of interest, but may also lead to a decline in cognitive function, such as distraction and memory loss, which in turn affects work and learning ability. More seriously, depression is a major risk factor for suicidal behavior, causing immeasurable losses to the patient's family and society. If depression is not treated promptly, it may turn into a chronic disease, leading to long-term functional impairment and social isolation, increasing the difficulty of treatment and the risk of relapse. Depression may also affect the next generation through family environment and genetic factors, and have a negative impact on the mental health and growth of children and adolescents. Therefore, early identification, diagnosis, treatment and prevention of depression are not only crucial to the patient's recovery, but also of great significance to reducing the social public health burden and promoting social stability and harmony.

[0003] Currently, the identification and diagnosis of depression mainly rely on communication interviews between doctors and patients and psychiatric questionnaire assessments. However, these methods all have certain limitations. The former depends on the professional knowledge of the physician and the cooperation of the patient, while the latter may also be affected by factors such as the patient's own reasons, subjective bias, denial or concealment. Therefore, the development of an objective assessment and detection method for depression, which is different from clinical subjective diagnosis, is crucial to improving the effectiveness of early identification of depression. In recent years, there are many objective diagnostic methods for depression, such as electroencephalography (EEG), positron emission tomography (PET), near-infrared spectroscopy (fNIRS) and functional magnetic resonance imaging (fMRI). EEG is often used for objective detection of depression due to its low cost and portability.

[0004] An electroencephalogram (EEG) is a non-invasive neurodiagnostic test used to record the brain's electrical activity. EEG captures weak electrical signals generated by neurons in the brain using a series of electrodes placed on the scalp. EEG has played a significant role in the diagnosis of many psychiatric disorders, particularly depression. In recent years, EEG-based depression diagnosis and identification has evolved from traditional observation of abnormal EEG indicators to machine learning and even deep learning. Numerous EEG-based artificial intelligence (AI) diagnostic models for depression have emerged.

[0005] However, many traditional machine learning diagnostic models, as well as some deep learning diagnostic models, have limitations. For example, some models treat only the electrode channels in EEG signals as isolated points. However, brain regions possess distinct connectivity and structural characteristics, interconnecting and interacting structurally or functionally. This connectivity is crucial for overall brain function, influencing cognition, emotion, behavior, and responses to external stimuli. Multiple studies have shown that individuals with depression exhibit significant differences in connectivity, including brain network connectivity strength and graph-theoretic properties, compared to controls. While some models consider interregional connectivity, using functional connectivity matrices or spatial distance matrices to construct interregional connectivity relationships, these relationships are multi-layered and multifactorial. Relying solely on functional connectivity matrices may overlook the spatial topological constraints of brain networks, while relying solely on spatial distance matrices may overlook the dynamic collaborative activities between distant brain regions. In addition, although many traditional machine learning diagnostic models and some deep learning diagnostic models have achieved relatively high performance, they ignore the interpretability of the models and do not conduct further exploration to discover interpretable biomarkers and reveal the neural mechanisms of the disease. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a multi-level information processing method for EEG of depression based on adaptive graph convolution, which provides multi-dimensional network structure information for EEG-based depression detection and can further explore and discover interpretable biomarkers based on the output of the model, thereby improving the interpretability of the model.

[0007] The purpose of the present invention is to achieve the following technical solution: a method for multi-level information processing of EEG for depression based on adaptive graph convolution, the specific steps of which are as follows:

[0008] S1. Construction of EEG signal feature matrix: Obtain EEG data and perform filtering processing, segment and amplify the filtered EEG data, and then extract the differential entropy features of each electrode channel in five different frequency bands: Delta frequency band, Theta frequency band, Alpha frequency band, Beta frequency band, and Gamma frequency band. The differential entropy features of the five frequency bands are integrated into the EEG signal feature matrix;

[0009] S2. Construct the EEG signal adjacency matrix. The specific steps are as follows:

[0010] S2.1. Construct a spatial topological structure matrix based on physical distance: Use the three-dimensional coordinates (x, y, z) of each electrode channel in the EEG signal acquisition device mapped to the unit sphere to calculate the geodesic distance between different channels. The calculation formula is:

[0011]

[0012] Among them, d ij represents the geodesic distance between electrodes i and j, r represents the unit distance 1, (x i ,y i ,z i ) and (x j ,y j ,z j ) represent the spatial coordinates of two different electrode points;

[0013] The calculated geodesic distance between each pair of channels is calculated using the following formula and converted into the weight representing the relationship between different channels in the spatial topological structure matrix;

[0014]

[0015] The value of α is 1;

[0016] The transformed relationship weights between different electrode channels are used as elements in the symmetric matrix to construct a fully connected matrix As the spatial topological structure matrix obtained for each subject;

[0017] S2.2. Construct a functional connectivity matrix based on the Pearson correlation coefficient: Use the EEG data after segmentation and amplification to calculate the Pearson correlation coefficient between the EEG signals at different electrodes of the subject;

[0018] The Pearson correlation coefficients calculated between all different channels are constructed into the final symmetric matrix. All elements in the symmetric matrix are absolute and the main diagonal elements are set to 0 to obtain the final functional connectivity matrix.

[0019] S2.3, construct the adaptive shared mask parameter matrix: construct the structure mask matrix W respectively Spatial and the function mask matrix W Functional , the values ​​of the elements in the matrix are randomly initialized using normal distribution;

[0020] S2-4. Construct the adjacency matrix of the input graph convolutional network. The specific calculation method is as follows:

[0021]

[0022] S3. Input the EEG signal feature matrix and adjacency matrix into the graph convolutional network for training, which includes the following steps:

[0023] S3.1. Initialize two adaptive shared mask parameter matrices in the adjacency matrix to generate an adjacency matrix;

[0024] S3.2. Input the EEG signal feature matrix and adjacency matrix into the graph convolutional network, first inputting it into the first graph convolutional layer, and using the LeakyReLU activation function to process the output of the first graph convolutional layer;

[0025] S3.3. Input the vector processed by the activation function into the second graph convolution layer and use LeakyReLU activation;

[0026] S3.4. Input the features obtained in S3.3 into a global pooling layer to generate a graph-level embedding vector;

[0027] S3.5. Dropout regularization is performed on the graph-level embedding vector, randomly dropping 20% ​​of the feature dimensions and scaling the remaining features by 1 / (1-p) to maintain the expected value.

[0028] S3.6. The regularized vector is input into the first fully connected layer for dimensionality reduction. The graph-level features are nonlinearly transformed and dimensionally adjusted through the fully connected layer with LeakyReLU activation.

[0029] S3.7. Input the transformed vectors from the first fully connected layer into the second fully connected layer for linear mapping to the unnormalized original prediction values, completing the end-to-end prediction from graph embedding to classification results.

[0030] S3.8. Convert the unnormalized original prediction value obtained in S3.7 into the final probability distribution through the Softmax function to obtain the final output prediction result of the model.

[0031] The beneficial effects of the present invention are: the method of the present invention first performs simple preprocessing on the acquired EEG signal, and then uses the sliding window method to segment and amplify the preprocessed EEG signal to obtain small segments of subject samples, and then extracts the DE differential entropy features of five different frequency bands for each subject signal sample, integrates them into a feature matrix, and then calculates the Pearson correlation coefficient between different electrode channels based on the sample signal to construct a functional connection matrix. In addition, the geodesic distance between different channels is calculated based on the three-dimensional coordinates of each electrode channel in the EEG acquisition device projected onto the unit sphere and converted into a spatial topological structure matrix between the channels through mathematical changes. The calculated spatial topological structure matrix and the functional connection matrix are respectively multiplied by a shared mask parameter matrix, and added and combined to obtain the final adjacency matrix. Finally, the adjacency matrix and the feature matrix are input together into the constructed graph convolutional network to complete depression detection. The method of the present invention introduces the functional connectivity relationship and spatial topological structure relationship between electrode channels, and provides a graph structure relationship of multiple dimensions for the graph convolutional network when performing aggregation channel node itself and its neighbor information, comprehensively reflecting the complexity of the brain network composed of each brain region of the brain, providing more hierarchical and rich network structure information between brain regions, better identifying the difference between depression patients and healthy people, and then improving the accuracy and robustness of depression detection. Not only that, the method of the present invention also multiplies the shared mask parameter matrix that can be automatically updated in the adjacency matrix, through continuous training and learning, to identify the brain connections that are important for depression detection, so as to further analyze important brain regions, brain connections, brain network indicators, etc., to find interpretable biomarkers and reveal the neural mechanism of the disease.

[0032] Compared with the existing EEG depression detection model, the innovation of the present invention lies in combining the spatial structural connectivity and functional connectivity between brain regions to provide multi-dimensional network structure information between brain regions. Spatial connectivity (based on the geodesic distance between electrodes) as stable structural prior information can help the model better capture the relationship between local brain regions, suppress noise in functional connectivity (such as short-term state fluctuations or artifacts), and avoid model overfitting; functional connectivity provides abnormal dynamic coordinated activities between distant brain regions caused by the disease. The complementarity of the two enhances the generalization and robustness of the model. Many other models that also use the connectivity between brain regions to detect depression mostly use a thresholding method to binarize the matrix when constructing the adjacency matrix between brain regions, so as to select important connections for depression detection. However, there has been no unified standard for the selection of thresholds, and the selected threshold will have a great impact on the accuracy and generalization ability of the model. In addition, this method may also ignore some connections with relatively small calculated connectivity values ​​but that have an important impact on depression detection. The present invention uses a fully connected matrix to construct the adjacency matrix, and multiplies the spatial topology matrix and the functional connectivity matrix by a shared mask parameter matrix respectively. Through training, the weights of the connections between channels are dynamically adjusted to identify the brain region connections that are truly important for depression detection. In addition, the parameter matrix finally output by the model can be used to further perform brain region analysis, brain connection analysis, etc., to further explore explainable biomarkers and enhance the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the method for multi-level information processing of depression EEG based on adaptive graph convolution of the present invention;

[0034] Figure 2 This is a flow chart of pretreatment in the method of the present invention;

[0035] Figure 3 It is a feature extraction module and an adjacency matrix construction module in the method of the present invention;

[0036] Figure 4 Schematic diagram of the structure of the graph convolutional network in the method of the present invention. DETAILED DESCRIPTION

[0037] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0038] like Figure 1 As shown, the present invention provides a method for multi-level information processing of EEG for depression based on adaptive graph convolution, and the specific steps are as follows:

[0039] S1, EEG signal feature matrix construction;

[0040] First, electroencephalogram (EEG) data is obtained. This embodiment uses the Lanzhou University public depression EEG dataset MODMA as the dataset for this method.

[0041] Then the data is preprocessed. For the MODMA data set obtained, the original data is collected by an EEG device with 128 electrodes. Since there are interfering electrodes in the data set, the interfering electrodes E8, E14, E21, E25, E43, E48, E49, E56, E63, E68, E73, E81, E88, E94, E99, E107, E113, E119, E120, E125, E126, E127, and E128 are removed. After removing 23 electrodes, the EEG data corresponding to the remaining 105 electrodes are obtained. A simple preprocessing method is used to select a 50Hz notch filter to filter the signal to remove the notch interference; then a 1-40Hz fourth-order Butterworth bandpass filter is selected to filter the signal in the second stage. The filtering process is as follows: Figure 2 shown.

[0042] Because the original MODMA dataset only included 53 subjects, we segmented and amplified each subject's EEG data using a sliding window approach based on the preprocessed EEG data. We selected a 2-second window length and set the overlap between adjacent windows to 1 second, ensuring 50% overlap between consecutive windows and continuous coverage of brain activity throughout the time series. Each segmented EEG data segment was treated as a separate subject. The amount of data from patients with depression and healthy controls after segmentation and amplification is shown in Table 1.

[0043] Table 1

[0044] MODMA dataset MDD (depressive disorder) HC (healthy people) Subject 24 29 Subject Windows 7176 8671

[0045] For the EEG data after filtering preprocessing, the differential entropy features of each electrode channel in five different frequency bands, namely Delta (1-3 Hz), Theta (4-8 Hz), Alpha (8-13 Hz), Beta (13-30 Hz), and Gamma (30-40 Hz) were extracted respectively. The calculation formula of the differential entropy features in each frequency band is:

[0046]

[0047] The time series τ follows a Gaussian distribution (μ,σ 2 ).

[0048] The differential entropy features of the five frequency bands extracted from each electrode channel of the subject are integrated into the EEG signal feature matrix of the subject; Figure 3 For each subject, the dimension DEi ∈R C×F The feature matrix of , where i is the i-th subject, C is the number of electrode channels in the EEG dataset, and F is the number of extracted features. Since there are 105 electrodes remaining in the subject window of the preprocessed dataset, and each electrode has features of 5 frequency bands, the dimension of the feature matrix in this example is DE i ∈R 105×5 .

[0049] S2. Construct the EEG signal adjacency matrix. The process is as follows: Figure 4 The second and third branches are shown in the dotted box on the left. The specific steps are as follows:

[0050] S2.1. Construct a spatial topological structure matrix based on physical distance: Use the three-dimensional coordinates (x, y, z) of each electrode channel in the EEG signal acquisition device mapped to the unit sphere to calculate the geodesic distance between different channels. The calculation formula is:

[0051]

[0052] Among them, d ij represents the geodesic distance between electrodes i and j, r represents the unit distance 1, (x i ,y i ,z i ) and (x j ,y j ,z j ) represent the spatial coordinates of two different electrode points i and j respectively; the spatial coordinates of the electrode points are the three-dimensional coordinates of each electrode channel in the EEG signal acquisition device mapped onto the unit sphere.

[0053] The calculated geodesic distance between each pair of channels is calculated using the following formula and converted into the weight representing the relationship between different channels in the spatial topological structure matrix;

[0054]

[0055] The parameter α has been verified through multiple experiments. When α = 1, the accuracy of the model is the highest. Therefore, in this example, α is set to 1.

[0056] The transformed relationship weights between different electrode channels are used as elements in the symmetric matrix to construct a fully connected matrix As the spatial topological structure matrix obtained by each subject; the dimension of the spatial topological structure matrix obtained by each subject is: Since the same set of EEG signal acquisition equipment is used when collecting EEG signals for both depression patients and healthy subjects, the spatial topological structure matrix here is the same for both groups of people.

[0057] S2.2. Construct a functional connectivity matrix based on the Pearson correlation coefficient: Use the EEG data after segmentation and amplification to calculate the Pearson correlation coefficient between the EEG signals at different electrodes of the subject;

[0058] The formula for calculating the Pearson correlation coefficient is as follows:

[0059]

[0060] Where Corr(x,y) represents the Pearson correlation coefficient between electrode channels x and y, and its value range is [-1,1]. x ,σ y Represents the standard deviation of the electrode x and y signals, and Cov(x,y) represents the covariance between the electrode x and y signals to measure the degree of joint fluctuation between them. The formula for calculating covariance is:

[0061]

[0062] Where N represents the number of time points, μ x 、μ y Represents the mean of the signals of channels x and y.

[0063] The Pearson correlation coefficients calculated between all different channels are constructed into the final symmetric matrix. All elements in the symmetric matrix are absolute and the main diagonal elements are set to 0 to obtain the final functional connectivity matrix. Its dimensions are:

[0064] S2.3, construct the adaptive shared mask parameter matrix: construct the structure mask matrix W respectively Spatial and the function mask matrix W Functional , whose dimensions are all N×N, and the values ​​of the elements in the matrix are randomly initialized using a normal distribution; here W Spatial and W Functional There are two shared mask parameter matrices that can be continuously updated and optimized through back propagation during training.

[0065] S2-4. Construct the adjacency matrix of the final input graph convolutional network. The specific calculation method is as follows:

[0066]

[0067] The spatial topology matrix and the structure mask matrix W Spatial , functional connectivity matrix and the function mask matrix W Functional The corresponding elements in are multiplied element by element, and then the average is calculated by adding them.

[0068] Through the above-mentioned feature extraction and adjacency matrix construction, and because in this example, 7176 MDD subjects and 8671 HC subjects were finally obtained after segmentation and amplification, a corresponding feature matrix and adjacency matrix were obtained for both MDD subjects and HC subjects.

[0069] S3, input the EEG signal feature matrix and adjacency matrix into the graph convolutional network for processing, such as Figure 4 As shown, the following steps are included:

[0070] S3.1. Initialize two adaptive shared mask parameter matrices in the adjacency matrix to generate an adjacency matrix;

[0071] S3.2. Input the EEG signal feature matrix and adjacency matrix into the graph convolutional network. Set the batch size for training to 4. First, input the first graph convolutional layer (GCNCovn1). The input dimension of the first graph convolutional layer is [420, 5], where 420 is the product of the number of nodes and the batch size, and 5 is the number of input feature channels. The feature matrix updates the node feature representation by aggregating neighbor information based on the graph structure relationship network of the adjacency matrix. The final output dimension is [420, 64]. The output of the first graph convolutional layer is then processed using the LeakyReLU activation function.

[0072] S3.3. The activation function-processed vector is input into the second graph convolutional layer (GCNCovn2) with an input dimension of [420, 64]. The node's feature representation is updated a second time by aggregating neighbor information using the same adjacency matrix graph structure network, resulting in an output dimension of [420, 20]. The output vector is batch-normalized (BatchNorm) and then activated with LeakyReLU, leaving the output dimension unchanged at [420, 20]. Through two layers of graph convolution, a node feature representation is obtained that undergoes high-order neighborhood information aggregation and nonlinear transformation.

[0073] S3.4. Input the node feature representation obtained in S3.3 into a global pooling layer (Gobalpooling). Sum all node features belonging to the same graph along the node dimension to generate a graph-level embedding vector. The final output vector has a dimension of [4, 20].

[0074] S3.5. Dropout regularization is performed on the graph-level embedding vector, randomly dropping 20% ​​of the feature dimensions (setting them to zero), and scaling the remaining features by 1 / (1-p) to maintain the expected value;

[0075] S3.6. The regularized vector is input into the first fully connected layer (FC) for dimensionality reduction. The graph-level features are nonlinearly transformed and dimensionally adjusted through the fully connected layer with LeakyReLU activation. The output dimension is [4,10].

[0076] S3.7. Input the transformed vectors from the first fully connected layer into the second fully connected layer for linear mapping to the unnormalized original prediction values, completing the end-to-end prediction from graph embedding to classification results.

[0077] S3.8. Convert the unnormalized original prediction value obtained in S3.7 into the final probability distribution through the Softmax function to obtain the final output prediction result of the model.

[0078] The graph convolution layer is constructed using the first-order approximate propagation rule of spectral convolution. The graph convolution layer produces feature changes of its input, and the formula is as follows:

[0079]

[0080] Among them, H (l) is the node feature matrix of the lth layer in the graph convolution layer; σ is the activation function, which uses the LeakyReLU function; is the adjacency matrix A ij Adding the identity matrix, for The diagonal matrix, W (l) is the weight matrix.

[0081] The MODMA dataset was processed using the method of this invention, and the model performance was evaluated using the average results of 10-fold cross-validation. This experiment was compared with several basic EEG-based depression detection models and previous work on depression detection using the MODMA dataset (the basic machine learning model and deep learning model methods corresponding to the Basic Method column were all independently developed). The results are shown in Table 2.

[0082] Table 2

[0083]

[0084]

[0085] As can be seen from Table 2, the detection accuracy of the model of the present invention in EEG-based depression detection reached 96.86%. Compared with some basic depression detection models, the accuracy is significantly higher than the basic detection models. In addition, compared with EEG-based depression detection models in recent years, the detection performance is higher than these models.

[0086] The test results of this example on the MODMA dataset show that the present invention performs well in EEG-based depression detection.

[0087] In summary, the method of the present invention fully combines the multi-level graph structure, integrates the spatial topological structure information between electrodes with the dynamic collaborative activity information between electrodes, and provides a richer relationship representation for the neighbor information of the integrated nodes in the graph convolutional network. In addition, in order to avoid ignoring some connections with small connection values ​​but more important for depression detection, the form of a fully connected matrix is ​​selected, and then a shared mask parameter matrix that can be automatically updated is multiplied for the spatial topological structure matrix and the functional connection matrix respectively. Through training, the weights of the connections between channels are dynamically adjusted to identify the brain area connections that are really important for depression detection. By continuously training to select important connections, it is more objective to reflect the importance of connections between brain areas for depression detection than by using a custom threshold. On the other hand, the parameter matrix output by the trained model and the functional connection can be further analyzed for brain areas, brain networks, and brain connections, so as to further explore the symptom mechanism of depression and find objective biomarkers.

[0088] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A multi-level information processing method for depression EEG based on adaptive graph convolution, characterized by: The specific steps are as follows: S1. Construction of EEG signal feature matrix: Obtain EEG data and perform filtering processing, segment and amplify the filtered EEG data, and then extract the differential entropy features of each electrode channel in five different frequency bands: Delta frequency band, Theta frequency band, Alpha frequency band, Beta frequency band, and Gamma frequency band. The differential entropy features of the five frequency bands are integrated into the EEG signal feature matrix; S2. Construct the EEG signal adjacency matrix. The specific steps are as follows: S2.

1. Construct a spatial topological structure matrix based on physical distance: Use the three-dimensional coordinates (x, y, z) of each electrode channel in the EEG signal acquisition device mapped to the unit sphere to calculate the geodesic distance between different channels. The calculation formula is: Among them, d ij represents the geodesic distance between electrodes i and j, r represents the unit distance 1, (x i ,y i ,z i ) and (x j ,y j ,z j ) represent the spatial coordinates of two different electrode points; The calculated geodesic distance between each pair of channels is calculated using the following formula and converted into the weight representing the relationship between different channels in the spatial topological structure matrix; The value of α is 1; The transformed relationship weights between different electrode channels are used as elements in the symmetric matrix to construct a fully connected matrix As the spatial topological structure matrix obtained for each subject; S2.

2. Construct a functional connectivity matrix based on the Pearson correlation coefficient: Use the EEG data after segmentation and amplification to calculate the Pearson correlation coefficient between the EEG signals at different electrodes of the subject; The Pearson correlation coefficients calculated between all different channels are constructed into the final symmetric matrix. All elements in the symmetric matrix are absolute and the main diagonal elements are set to 0 to obtain the final functional connectivity matrix. S2.3, construct the adaptive shared mask parameter matrix: construct the structure mask matrix W respectively Spatial and the function mask matrix W Functional , the values ​​of the elements in the matrix are randomly initialized using normal distribution; S2-4. Construct the adjacency matrix of the input graph convolutional network. The specific calculation method is as follows: S3. Input the EEG signal feature matrix and adjacency matrix into the graph convolutional network for training, which includes the following steps: S3.

1. Initialize two adaptive shared mask parameter matrices in the adjacency matrix to generate an adjacency matrix; S3.

2. Input the EEG signal feature matrix and adjacency matrix into the graph convolutional network, first inputting it into the first graph convolutional layer, and using the LeakyReLU activation function to process the output of the first graph convolutional layer; S3.

3. Input the vector processed by the activation function into the second graph convolution layer and use LeakyReLU activation; S3.

4. Input the features obtained in S3.3 into a global pooling layer to generate a graph-level embedding vector; S3.

5. Dropout regularization is performed on the graph-level embedding vector, randomly dropping 20% ​​of the feature dimensions and scaling the remaining features by 1 / (1-p) to maintain the expected value. S3.

6. The regularized vector is input into the first fully connected layer for dimensionality reduction. The graph-level features are nonlinearly transformed and dimensionally adjusted through the fully connected layer with LeakyReLU activation. S3.

7. Input the transformed vectors from the first fully connected layer into the second fully connected layer for linear mapping to the unnormalized original prediction values, completing the end-to-end prediction from graph embedding to classification results. S3.

8. Convert the unnormalized original prediction value obtained in S3.7 into the final probability distribution through the Softmax function to obtain the final output prediction result of the model.

2. The method for multi-level information processing of depression EEG based on adaptive graph convolution according to claim 1 is characterized in that: The graph convolution layer is constructed using the first-order approximate propagation rule of spectral convolution. The graph convolution layer produces feature changes of its input, and the formula is as follows: Among them, H (l) is the node feature matrix of the lth layer, σ is the activation function, is the adjacency matrix A ij Adding the identity matrix, for The diagonal matrix, W (l) is the weight matrix.

Images