Method for recognizing depression of electroencephalogram signal based on multi-view capsule graph network

Abstract

The application discloses a depression recognition method based on a multi-view capsule graph network, belongs to the field of electroencephalogram analysis in the field of biological medicine, and comprises the following steps: preprocessing an electroencephalogram signal; extracting node features; constructing a topological graph, a functional graph and a causal graph according to the node features; adaptively fusing the topological graph, the functional graph and the causal graph to obtain a fused graph; a bidirectional adjustment mechanism of a capsule graph network; constructing a loss function and a total training target of the capsule graph network; and a depression recognition process based on a dynamic routing mechanism of the capsule graph network.The application introduces the capsule graph network to structurally model topological structures and connection modes among electroencephalogram channels; the mutual relationship among the electroencephalogram channels is enhanced through adaptive fusion, so that the capsule graph network can more accurately capture the interaction and coupling features of brain regions; and the bidirectional adjustment mechanism is introduced to flexibly adjust the contribution degree of features, so as to enhance the information transmission capacity of the capsule graph network in a global context and the sensitivity and robustness of the capsule graph network to depression-related features.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) analysis technology in the biomedical field, specifically relating to a method for depression recognition based on EEG signals using a multi-view capsule graph network. Background Technology

[0002] Depression is a common mental health problem. Traditional identification methods rely on scale assessments and clinical interviews, which are highly subjective and prone to misdiagnosis or missed diagnosis. Electroencephalography (EEG), as a non-invasive and objective neurophysiological indicator with millisecond-level temporal resolution, can capture brain electrical activity in real time, providing new possibilities for the auxiliary diagnosis of depression. Significant differences exist in the brain electrical activity of patients with depression compared to healthy individuals in EEG signals, particularly in the asymmetry of activity in the prefrontal cortex. Therefore, EEG-based depression identification technology has broad application prospects and can provide a new approach for accurate and efficient auxiliary diagnosis of depression.

[0003] Currently, deep learning-based depression recognition methods mainly focus on modeling physiological signals such as multi-channel EEG. These methods typically represent EEG signals as Euclidean tensors and utilize convolutional networks, recurrent networks, or attention mechanisms for end-to-end feature learning to improve the representational ability and stability of depression-related patterns. Among these, graph neural network-based depression recognition methods treat EEG channels or brain regions as nodes and the connections between channels as edges. They characterize the brain network structure based on a specific connection feature (such as topological or functional connectivity) and perform feature propagation and aggregation on the graph to capture the correlation information between channels.

[0004] The closest existing depression identification method to this invention is a deep learning approach based on EEG, which employs the TSF-MDD model and combines spatiotemporal frequency features for major depressive disorder (MDD) identification. Specifically, this method first transforms the EEG signal into a four-dimensional spatiotemporal frequency domain representation through a data reconstruction scheme, capturing the features of the EEG signal in the time, space, and frequency dimensions. Subsequently, the TSF-MDD model uses a 3D convolutional neural network (3D-CNN) and a capsule network (CapsNet) for feature extraction. The 3D-CNN processes the four-dimensional data, preserving the spatiotemporal features of the signal, while CapsNet further captures higher-order spatial features. Although this method improves the feature learning ability and robustness of the TSF-MDD model to some extent, it still has the following significant shortcomings:

[0005] First, simply treating multi-channel signals as Euclidean tensors ignores the topological interactions between depression-related brain regions and fails to effectively quantify channel coupling strength or adapt to nonlinear embeddings, resulting in insufficient modeling of the constraints on the overall brain network.

[0006] Secondly, it only considers a single connectivity feature (such as topological connectivity, functional connectivity, or causal connectivity) without considering all connectivity features, thus failing to effectively integrate different connectivity features and making full use of the correlation information between EEG channels.

[0007] Third, capsule graph networks only have a bottom-up feedback mechanism. Although they can effectively extract local brain connectivity patterns, they ignore the dynamic adjustment of the global context. This makes the aggregation of the bottom capsule graphs unconstrained by the task semantics, and they are prone to learning statistically stable patterns that are unrelated to depression. As a result, the model's sensitivity to depression-related features decreases in noisy environments, ultimately affecting the overall classification accuracy and robustness. Summary of the Invention

[0008] To address the problems existing in current depression recognition methods, this invention provides a depression recognition method based on EEG signals using a multi-view capsule graph network.

[0009] The technical solution adopted by this invention to solve the technical problem is as follows:

[0010] This invention provides a method for identifying depression based on electroencephalogram (EEG) signals using a multi-view capsule graph network, comprising the following steps:

[0011] Step 1: Preprocess the EEG signals;

[0012] Step 2: Extract node features from the preprocessed EEG signals;

[0013] Step 3: Construct topological graphs, functional graphs, and causal graphs based on node features to represent different brain connectivity mechanisms; adaptively fuse the topological graphs, functional graphs, and causal graphs to obtain a fused graph;

[0014] Step 4: The bidirectional regulatory mechanism of capsule diagram networks;

[0015] Step 5: Construct the loss function and overall training objective for the capsule graph network;

[0016] Step 6: Depression identification process based on capsule graph network dynamic routing mechanism.

[0017] Furthermore, in step one, the channel names and numbers are unified according to the standard electrode layout; the sampling rate is unified, and bandpass filtering is used to remove high and low frequency noise; 50 Hz power frequency interference is suppressed by notch filtering; the waveforms, power spectra, and scalp distribution of independent components in the EEG signal are observed, and components related to eye movement or electromyography are removed; all EEG signals are calibrated to a unified benchmark by using the whole brain average rereference.

[0018] Furthermore, in step two, EEG channels or brain regions are regarded as nodes, and the connections between channels are regarded as edges. The preprocessed EEG fragments are used as input to perform convolution modeling on the time series of each node. By sliding the convolution kernel in the time dimension to capture local patterns, and combining pooling operations and nonlinear activation to extract time structure features, each node is mapped to a fixed-length time-domain embedding vector, i.e., node features.

[0019] Furthermore, in step three, when constructing the topology graph, the Euclidean distance between any two electrodes is first calculated, and the Euclidean distance is mapped to a similarity between (0,1) using a Gaussian kernel function to construct a basic topological adjacency matrix. Then, the topological relationships of the basic topological adjacency matrix are initialized using radial basis functions to obtain the topology graph.

[0020] Furthermore, in step three, when constructing the functional graph, mutual information is first used as a measure of functional connectivity. By estimating the joint probability distribution and marginal distribution of the time series of two nodes within an EEG segment, the statistical dependency strength is calculated. The mutual information of all node pairs is organized into a matrix and normalized to obtain the functional graph.

[0021] Furthermore, in step three, when constructing the causal graph, the residual difference between using only its own historical information and using the historical information of two nodes simultaneously to predict the target node is compared. GC test is used to obtain the causal relationship between nodes, and the prediction error between nodes is calculated to construct the causal graph in the spatiotemporal domain.

[0022] Furthermore, in step three, a dynamic weighted approach is used to adaptively fuse the topology graph, functional graph, and causal graph: first, the three adjacency matrices of topology, function, and causality are combined with the temporal embedding vectors of the nodes, and the temporal embedding vectors of each node pair are concatenated. Then, trainable parameter vectors are defined for the topology graph, functional graph, and causal graph respectively. The view attention weights of the node pairs are obtained through softmax. The relative importance of different views is adaptively adjusted at the edge level using these view attention weights, and a fused graph of multi-view fusion is obtained.

[0023] Furthermore, in step four, the global information of the classification capsules is passed back to the primary capsules through the higher-level capsules to achieve top-down feature adjustment; first, all classification capsule vectors are... Aggregate to form a global class vector global class vector As a conditional input, the adjustment parameter is obtained through linear transformation. and Used for primary capsules in the lower layers Perform feature scaling and offset.

[0024] Furthermore, in step five, the loss function includes: margin loss. Reconstruction loss and attention entropy regularization term The overall training objective is , and These are all hyperparameters used to balance the relative importance of classification accuracy, reconstruction capability, and attention entropy regularization.

[0025] The beneficial effects of this invention are:

[0026] This invention introduces a capsule diagram network to structurally model the topological structure and connection patterns between brain electrical channels. It enhances the interrelationships between these channels through multi-view adaptive fusion, enabling the capsule diagram network to more accurately capture the interaction and coupling features of brain regions. Furthermore, by introducing a bidirectional adjustment mechanism into the capsule diagram network, the contribution of features can be flexibly adjusted, solving the problem that existing capsule diagram networks rely solely on bottom-up feedback mechanisms. This enhances the information transmission capability of the capsule diagram network in the global context and its sensitivity to depression-related features, while also improving its robustness, especially maintaining high accuracy and stability in environments with significant noise interference. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of a depression recognition method based on multi-view capsule graph network for EEG signals provided by the present invention.

[0028] Figure 2 This is a diagram of the capsule diagram network model proposed in this invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings.

[0030] To more comprehensively characterize brain network structures, this invention proposes a modeling method that combines multi-view capsule diagram networks to achieve comprehensive modeling of brain regions in EEG signals. Specifically, by adaptively fusing topological, functional, and causal views, it can simultaneously capture global information of the brain and detailed features of local brain regions, significantly improving the recognition ability of depression-related brain networks and making the classification of EEG signals by capsule diagram networks more accurate.

[0031] like Figure 1As shown, this invention provides a method for depression recognition based on EEG signals using a multi-view capsule graph network. First, the raw multi-channel EEG signals are preprocessed, and node features are extracted. Then, based on the extracted node features, a topology graph, a functional graph, and a causal graph are constructed to represent different brain connectivity mechanisms. Next, an adaptive fusion module fuses the topology graph, functional graph, and causal graph into a single, information-rich fusion graph. Subsequently, the structural information of the fusion graph is encoded into a set of primary capsules containing spatial hierarchical relationships. This bidirectional adjustment mechanism generates advanced capsules representing spatiotemporal information. Next, a top-down contextual feedback loop is constructed, which performs feature scaling and translation on the underlying graph capsule vectors based on the initially generated classification capsules, thereby obtaining the calibrated primary capsules. The calibrated features are then processed again through dynamic routing calculations, ultimately outputting the vector length of the target classification capsule. This vector length determines whether a sample is classified as depressed or healthy. The specific judgment rule is as follows: the capsule graph network model ultimately outputs two classification capsules representing depression and health, respectively. The vector length of these two classification capsules is then calculated. The norm (i.e., vector length) serves as the probability distribution of an EEG signal sample belonging to its corresponding category (depression or health). If the vector length of the depression category capsule is greater than that of the health category capsule, the EEG signal sample is determined to originate from a patient with depression; otherwise, it is determined to be from a healthy individual.

[0032] This invention provides a method for identifying depression based on electroencephalogram (EEG) signals using a multi-view capsule graph network. The specific implementation process is as follows:

[0033] Step 1: Preprocess the EEG signals;

[0034] Step S1.1: Based on the standard electrode layout, unify the channel names and numbers to ensure that the channel topology of all subjects remains consistent.

[0035] Step S1.2: Resample the sampling rate to 250Hz to ensure consistency in subsequent feature extraction. Simultaneously, use bandpass filtering to remove high and low frequency noise outside the 1–40 Hz range.

[0036] Step S1.3: Suppress 50 Hz power frequency interference by notch filtering to maintain a good signal-to-noise ratio for EEG signals in the normal neural oscillation frequency band (δ wave, θ wave, α wave, β wave, etc.).

[0037] Step S1.4: Observe the waveform, power spectrum and scalp distribution of independent components in the EEG signal, and remove components related to eye movement or electromyography to obtain a cleaner EEG signal.

[0038] Step S1.5: All EEG signals are calibrated to a unified benchmark through a whole-brain average rereference to reduce the impact of overall potential drift on the relative relationships between channels.

[0039] Step 2: Extract node features from the EEG signal;

[0040] Treating EEG channels or brain regions as nodes and connections between channels as edges, the preprocessed EEG fragments... As input, where The total number of nodes. This represents the number of sampling points corresponding to the length of the EEG fragment. Let be the set of real numbers. For node ... time series Convolution modeling is performed using a multi-layer one-dimensional convolutional neural network (1D-CNN), where... By capturing local patterns through the sliding motion of convolutional kernels in the temporal dimension, and then combining pooling operations with nonlinear activation to extract temporal structural features, the mathematical expression of this process is as follows:

[0041]

[0042] in, This represents a mapping consisting of several layers of one-dimensional convolutions, pooling operations, and non-linear activations. These are trainable parameters. Ultimately, each node is mapped to a fixed-length temporal embedding vector. , that is, node features, where is the dimension of the temporal embedding vector.

[0043] Step 3: Construction and fusion of multi-view brain networks;

[0044] Step S3.1: Topology graph construction;

[0045] A topological graph is constructed based on the node features of the extracted EEG signals. Specifically, the Euclidean distance between any two electrodes can be calculated using their three-dimensional coordinates in a standard layout. This Euclidean distance is then mapped to a similarity between (0,1) using a Gaussian kernel function, thus constructing a basic topological adjacency matrix. In this matrix, electrode pairs with closer Euclidean distances have higher connection weights. The mathematical expression for the topological graph is obtained by initializing the topological relationships of the basic adjacency matrix using radial basis functions (RBFs).

[0046]

[0047] in, For nodes in the adjacency matrix of the topological graph and nodes Connection weights between them Represents a node and nodes The Euclidean distance between them; It is a constant and is mainly used to control the radial range of action.

[0048] Step S3.2: Functional diagram construction;

[0049] Functional connectivity describes the degree of synchronization between different brain regions in time or frequency, reflecting the brain's collaborative working patterns in specific tasks or states. This invention uses mutual information as a measure of functional connectivity by estimating the time series of two nodes within an EEG segment (here, the time series are treated as discrete random variables). and ) are given their joint probability distribution and marginal distribution, and their statistical dependence strength is calculated.

[0050] Specifically, mutual information can be defined as:

[0051]

[0052] in, For mutual information, For random variables and The values ​​are respectively and Joint probability distribution at time, and They are random variables and The marginal probability distribution. The larger the mutual information value, the tighter the functional coupling between the two nodes. The mutual information of all node pairs is organized into a matrix and normalized to obtain the mathematical expression of the functional graph as follows:

[0053]

[0054] in, This is a functional graph adjacency matrix. This represents the maximum value of mutual information among all node pairs, used to normalize the functional connection strength.

[0055] Step S3.3: Constructing the cause-effect graph;

[0056] Unlike functional connectivity, which emphasizes simultaneity, causal graphs describe the influence of activity in one brain region on future activity in another, i.e., temporal directionality and driving force. This invention uses Granger causality to construct causal graphs.

[0057] Specifically, by comparing the residual difference when predicting a target node using only its own historical information versus using the historical information of two nodes simultaneously, the GC test (Granger Causality Test) is used to obtain the causal relationship between nodes, and the prediction error between nodes is calculated, thereby constructing a causal graph in the spatiotemporal domain.

[0058] The mathematical expression for a cause-effect graph is defined as follows:

[0059]

[0060] in, It is a causal graph adjacency matrix. For nodes The prediction error For nodes With nodes The prediction error between them For nodes The prediction error For nodes With nodes The prediction error between them; , and All causal factors were calculated using Granger causality. Represents a node To the node causal relationship Represents a node To the node causal relationship Represents a node To the node The causal relationship. Represents a node For nodes Granger causality strength node For nodes Granger causality strength.

[0061] Step S3.4: Adaptive fusion;

[0062] This invention employs a dynamic weighting method to adaptively fuse three views: topological graph, functional graph, and causal graph. Specifically, it first combines the three adjacency matrices (topological, functional, and causal) with the temporal embedding vectors of the nodes, and then applies a dynamic weighting method to each node. Then concatenate their time-domain embedding vectors:

[0063]

[0064] in, For node pairs The splicing characteristics, and They represent the first Each node and Temporal embedding vectors of nodes, This represents the vector concatenation operation. Subsequently, trainable parameter vectors are defined for the three views: the topological graph, the functional graph, and the causal graph. The softmax method is used to obtain the node pair. The view attention weights are as follows:

[0065]

[0066] in, , , These represent the type labels for topology views, functional views, and cause-and-effect views, respectively. View attention weights represent the attention weights on node pairs. Upper The relative importance of each view satisfies This adaptively adjusts the relative importance of different views at the edge level, resulting in a fused graph of multiple views.

[0067]

[0068] in, Represents the nodes in the adjacency matrix of the fused graph. With nodes The final connection weight between them For the first The original adjacency matrix of each view is then used. Subsequently, a graph attention network is applied to this fused graph to linearly map node features, learning a representation space more suitable for attention modeling. For any node... Its temporal embedding vector is re-represented as:

[0069]

[0070] in, This is a learnable linear transformation matrix. Finally, through node... The node is obtained by weighted aggregation of the features of its neighbor nodes. The primary capsules are as follows:

[0071]

[0072] in, It is a primary capsule. It is a non-linear activation function. Adjacency matrix for multi-view fusion The defined set of neighboring nodes will include nodes Consider as a node The neighboring nodes, For neighboring nodes For nodes The contribution of splicing features It was obtained by calculation using the attention mechanism and normalization.

[0073] Finally, a set of node-level primary capsules is obtained. ,in Compared to the scalar node representation of traditional GNNs, capsule-shaped node vectors can carry richer structural information and directional features, providing a more flexible representation space for subsequent high-order dynamic routing and classification.

[0074] Step 4: The bidirectional regulatory mechanism of capsule diagram networks;

[0075] In traditional capsule graph networks, feature propagation is unidirectional, a bottom-up process from node capsules (including primary and advanced capsules) to classification capsules. However, this capsule graph network structure struggles to fully utilize learned global information, especially when dealing with multi-level features. To overcome this limitation, this invention proposes to pass global information from classification capsules back to primary capsules via advanced capsules, thereby achieving top-down feature tuning.

[0076] Specifically, firstly, for all categorical capsule vectors Aggregate to form a global class vector :

[0077]

[0078] Then, the global class vector As a conditional input, the adjustment parameter is obtained through linear transformation. and :

[0079]

[0080] in, and All of these are trainable parameters. and These are the scaling and offset parameters that are being adjusted. Next, we will adjust the parameters. and Used for weighted regulation of the characteristics of lower-level primary capsules, i.e., for lower-level primary capsules. Feature scaling and offset are performed, and the mathematical expression for this process is as follows:

[0081]

[0082] in, This is the primary calibration capsule obtained after scaling and offsetting. This represents element-wise multiplication. Through this bidirectional information exchange, the capsule graph network can dynamically adjust the contribution of each layer during feature extraction, enhancing the flexibility and expressive power of brain network structure modeling. This helps to obtain more stable and discriminative feature representations, thereby significantly improving classification performance.

[0083] Step 5: Loss Function and Training Objective;

[0084] Step S5.1: As the core classification loss of the capsule graph network, the margin loss is used to constrain the consistency between the length of the classification capsule vector and the true class of the sample. Let... For the first The output vector of each classification capsule, For the first The vector length of each classification capsule, For the first If the true label of each categorized capsule indicates that the margin loss is defined as follows:

[0085]

[0086] in, The target lower boundary for the classification capsule module length corresponding to the correct category. The target upper boundary of the classification capsule module length corresponding to the error category; The weight coefficients are used to reduce the penalty for incorrect classes, preventing the length of all capsules from being compressed to 0 in the early stages of training. Margin loss encourages the length of the classification capsule vector for the correct class to be close to 1, while the length of the classification capsule vector for the incorrect class is close to 0, giving the length of the classification capsule vector a probabilistic meaning.

[0087] Step S5.2: Using the classification capsule vector corresponding to the predicted category as conditional input, the original graph-level features or node embeddings are reconstructed through a decoder composed of several fully connected layers. The reconstruction error is measured in the form of mean squared error (MSE). The mathematical expression for this process is as follows:

[0088]

[0089] in, These are the original graph-level features. This is the decoder output. The reconstruction loss encourages the classification capsule to retain effective information related to the input features, which helps improve the interpretability of the representation and the stability of training.

[0090] Step S5.3: Introduce an attention entropy regularization term to appropriately constrain the attention distribution. The attention entropy regularization term is defined as:

[0091]

[0092] in, This indicates that the view is assigned to the first view during the multi-view fusion phase. Attention weights for various views (i.e., topology graph, functional graph, and cause-effect graph).

[0093] This invention encourages more selective attention distribution by adding an attention entropy regularization term to the overall loss and assigning it a smaller weight. This means that more weight is allocated to a few more important views or branches, thereby improving the interpretability of the capsule graph network model.

[0094] Step S5.4: The overall training objective of the capsule diagram network model can be expressed as:

[0095]

[0096] in, and These are all hyperparameters used to balance the relative importance of classification accuracy, reconstruction capability, and attention entropy regularization. In actual training, this invention achieves a better trade-off between classification performance and representation stability in the capsule graph network model by adjusting the hyperparameters on the validation set.

[0097] Through the design of the loss function and training objective described above, the capsule graph network model can maintain high classification accuracy while taking into account the interpretability of multi-view fusion and the structural constraints of capsule representation, providing a solid foundation for subsequent experimental results and visualization analysis.

[0098] Step Six: Depression Identification Process Based on Capsule Graph Network Dynamic Routing Mechanism;

[0099] like Figure 2 As shown, firstly, the initial input nodes and graph topology features are denoted as... The capsule graph network model extracts features layer by layer through a multi-layer graph neural network (GNN). They represent the first to the second level, respectively. The node and graph topology features extracted by the multilayer GNN are shown below. The final local node features output by the GNN are represented as follows: Subsequently, through feature aggregation operations, as shown in the diagram... Local node features are reorganized into primary capsules, with an overall dimension of [missing value]. ,in The number of primary capsules, Let be the vector dimension of a single primary capsule. Next, the primary capsule enters the "bidirectional adjustment mechanism + routing mechanism" module and is mapped into a series of higher-level capsules, denoted as . ,common Finally, the advanced capsule, through a bidirectional regulation and routing mechanism, outputs a classification capsule ultimately used for EEG signal depression recognition, denoted as [missing information]. These correspond to two different categories: depression and health.

[0100] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying depression based on EEG signals using a multi-view capsule diagram network, characterized in that, Includes the following steps: Step 1: Preprocess the EEG signals; Step 2: Extract node features from the preprocessed EEG signals; Step 3: Construct topological graphs, functional graphs, and causal graphs based on node features to represent different brain connectivity mechanisms; adaptively fuse the topological graphs, functional graphs, and causal graphs to obtain a fused graph; Step 4: The bidirectional regulatory mechanism of capsule diagram networks; Step 5: Construct the loss function and overall training objective for the capsule graph network; Step 6: Depression identification process based on capsule graph network dynamic routing mechanism.

2. The method for depression recognition based on EEG signals using a multi-view capsule graph network according to claim 1, characterized in that, In step one, the channel names and numbers are standardized according to the standard electrode layout; the sampling rate is standardized, and bandpass filtering is used to remove high and low frequency noise; notch filtering is used to suppress 50 Hz power frequency interference. Observe the waveforms, power spectra, and scalp distribution of independent components in the electroencephalogram (EEG) signal, and remove components related to eye movements or electromyography (EMG). All EEG signals are calibrated to a uniform benchmark using a whole-brain average rereference.

3. The method for depression recognition based on EEG signals using a multi-view capsule network according to claim 1, characterized in that, In step two, EEG channels or brain regions are regarded as nodes and the connections between channels are regarded as edges. The preprocessed EEG fragments are used as input to perform convolution modeling on the time series of each node. The local patterns are captured by the sliding of the convolution kernel in the time dimension. The temporal structure features are extracted by combining pooling operations and nonlinear activation. Each node is mapped to a fixed-length temporal embedding vector, i.e., node features.

4. The method for depression recognition based on EEG signals using a multi-view capsule network according to claim 1, characterized in that, In step three, when constructing the topology graph, the Euclidean distance between any two electrodes is first calculated. The Euclidean distance is then mapped to a similarity between (0,1) using a Gaussian kernel function to construct the basic topological adjacency matrix. Finally, the topological relationships of the basic topological adjacency matrix are initialized using radial basis functions to obtain the topology graph.

5. The method for depression recognition based on EEG signals using a multi-view capsule network according to claim 1, characterized in that, In step three, when constructing the functional graph, mutual information is first used as a measure of functional connectivity. By estimating the joint probability distribution and marginal distribution of the time series of two nodes within an EEG segment, the statistical dependency strength is calculated. The mutual information of all node pairs is organized into a matrix and normalized to obtain the functional graph.

6. The method for depression recognition based on EEG signals using a multi-view capsule graph network according to claim 1, characterized in that, In step three, when constructing the causal graph, the residual difference between using only its own historical information and using the historical information of two nodes to predict the target node is compared. GC test is used to obtain the causal relationship between nodes and to calculate the prediction error between nodes, thus constructing a causal graph in the spatiotemporal domain.

7. The method for depression recognition based on EEG signals using a multi-view capsule graph network according to claim 1, characterized in that, In step three, a dynamic weighted approach is used to adaptively fuse the topology graph, functional graph, and causal graph: First, the three adjacency matrices of topology, function, and causality are combined with the temporal embedding vectors of the nodes. The temporal embedding vectors of each node pair are concatenated. Then, trainable parameter vectors are defined for the topology graph, functional graph, and causal graph respectively. The view attention weights of the node pairs are obtained through softmax. The relative importance of different views is adaptively adjusted at the edge level using these view attention weights, and a fused graph of multi-view fusion is obtained.

8. The method for depression recognition based on EEG signals using a multi-view capsule network according to claim 1, characterized in that, In step four, the global information of the classification capsules is passed back to the primary capsules through the higher-level capsules to achieve top-down feature adjustment; first, all classification capsule vectors are processed. Aggregate to form a global class vector global class vector As a conditional input, the adjustment parameter is obtained through linear transformation. and Used for primary capsules in the lower layers Perform feature scaling and offset.

9. The method for depression recognition based on EEG signals using a multi-view capsule network according to claim 1, characterized in that, In step five, the loss function includes: margin loss. Reconstruction loss and attention entropy regularization term The overall training objective is , and These are all hyperparameters used to balance the relative importance of classification accuracy, reconstruction capability, and attention entropy regularization.

Images