Electroencephalogram signal multi-dimensional characterization learning method for depression detection

Abstract

The invention discloses an electroencephalogram signal multi-dimensional characterization learning method for depression detection, which comprises the following steps: S1, acquiring an electroencephalogram signal data set of a patient, the data set comprising electroencephalogram signals and corresponding depression labels; s2, preprocessing the electroencephalogram signals, wherein the preprocessing is to segment the electroencephalogram signals into fixed lengths; s3, extracting structural correlation features of the electroencephalogram signals through a graph feature extraction and structure learning module; s4, extracting time dynamic characteristics of the electroencephalogram signals by adopting a gating Transform mechanism, and optimizing and representing a time sequence state through a gating mechanism; s5, fusing and optimizing the structural correlation features and the time dynamic features into a gating mechanism by using a Mama-based state space model; and S6, performing dimension reduction on the fused features by using a UMAP technology, and generating a visual visualization result. According to the method, the most advanced effect is realized on the depression data set.

Description

Technical Field

[0001] The present invention relates to the technical fields of multi-dimensional representation learning and depression detection. More specifically, it particularly relates to a method for multi-dimensional representation learning of electroencephalogram (EEG) signals for depression detection. Background Art

[0002] Depression is a common but complex mental disorder, characterized by low mood, loss of interest, and cognitive function decline, which has a serious impact on the quality of personal life and social and economic development. Early and accurate depression detection is crucial for timely intervention and treatment. However, traditional depression detection methods mainly rely on questionnaire assessment and clinical interviews, which are often greatly affected by subjective factors and are difficult to achieve objective and rapid diagnosis. Therefore, objective detection methods based on biomarkers have gradually become a research hotspot.

[0003] Electroencephalogram (EEG) signals, as a non-invasive neurobiological biomarker, can capture subtle changes in brain activities and are important tools in depression detection. EEG signals can not only reflect the real-time dynamic activities of the brain but also provide rich temporal, spatial, and state information through multi-channel acquisition. However, due to the characteristics of multi-channel and high sampling rate of the acquisition devices, EEG data presents the characteristics of high dimension and long time series, bringing huge challenges to data processing and feature extraction.

[0004] Currently, most studies only focus on single-dimensional feature extraction, such as temporal features, spatial structure features, or state features, and it is difficult to comprehensively capture the correlations of these features among multiple channels. At the same time, when dealing with long time series data, traditional methods are prone to loss of key information. In addition, due to individual differences in the patient's condition, the inconsistency of detection time further increases the complexity of data processing, which poses higher requirements for efficient and accurate depression detection. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for multi-dimensional representation learning of EEG signals for depression detection to overcome the defects of the prior art.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A method for multi-dimensional representation learning of EEG signals for depression detection, comprising the following steps:

[0008] S1. Obtain a dataset of EEG signals of patients, where the dataset includes EEG signals and corresponding depression labels;

[0009] S2. Preprocess the EEG signals, where the preprocessing is to segment the EEG signals into a fixed length;

[0010] S3. Extract the structural correlation features of the EEG signals through the graph feature extraction and structure learning module;

[0011] S4. Adopt the gated Transformer mechanism to extract the temporal dynamic features of the EEG signals, and optimize and characterize the temporal states through the gating mechanism;

[0012] S5. Utilize the Mamba-based state space model to fuse the structural correlation features and the temporal dynamic features and optimize them into the gating mechanism;

[0013] S6. Use the UMAP technique to reduce the dimension of the fused features and generate an intuitive visualization result.

[0014] Furthermore, the formula for the patient EEG signal dataset in step S1 is d eeg ={(X 1 , Y 1 ), (X 2 , Y 2 ),..., (X n , Y n )}, where X is the EEG signal, Y is the corresponding depression label, and n represents the total number of samples, where N e represents the number of electrodes, represents the electrode set, and col represents the length of the signal.

[0015] Furthermore, in step S2, the EEG signal is divided into T non-overlapping time sliding windows, where T = signal length / time window size, and the time windows are indexed as {t 1 , t 2 ,..., t T}.

[0016] Furthermore, in step S3, the structural correlation features of the EEG signal are defined as M e , and the electrode v is used as a node in the graph structure G e , and the initial adjacency matrix is formed according to the spatial relationship connection Then, the symmetric adjacency matrix is constructed by connecting the electrodes with cross-hemisphere symmetry The comprehensive adjacency matrix

[0017] Furthermore, for each node v in the graph structure G e =(V e , E e ), calculate the average feature of its neighbors:

[0018] h N(v) =mean({eu |u ∈ N(v)})

[0019] Connect the feature e of each node v with the average feature h of its neighbors N(v) to form a combined feature:

[0020] h v = e v || h N(v)

[0021] The combined feature is transformed through a linear layer and processed by a ReLU activation function defined by weights W and bias b:

[0022] h' v = ReLU(W · h v + b)

[0023] The transformed node feature h' v is integrated into a matrix as the initial feature matrix for subsequent neural network processing.

[0024] Furthermore, a graph convolutional network is used to capture the relationships between nodes. The basic operation of the graph convolutional network is expressed as:

[0025]

[0026] where H (l) is the node feature representation of the l-th layer, H (l+1) is the node feature representation of the (l + 1)-th layer, A e is the adjacency matrix of the graph, D is the degree matrix, W (l) is the weight matrix of the l-th layer, σ is the activation function. In the formula, is used to normalize the adjacency matrix to maintain numerical stability when aggregating adjacent node features, H (l) W (l) represents feature transformation, that is, mapping the current features to a new feature space through the weight matrix; the activation function σ provides a non-linear mapping.

[0027] Furthermore, it also includes adopting a structure learning mechanism to optimize the graph structure after pooling and encode the potential pairwise relationships between nodes, specifically including:

[0028] The structure learning mechanism adopts a sparse attention method, where each subgraph is generated from the graph G i after the k-th layer of pooling and is used as the input. The adjacency matrix and the node feature matrix are input into a single-layer neural network, which consists of a weight vector The similarity score between the parameterized neural network computing node v p and v q is as follows:

[0029]

[0030] where σ(·) is the activation function, λ is the parameter for balancing directly connected nodes and indirectly connected nodes, and for the entire vector apply the sparsemax function:

[0031]

[0032] The sparsemax function transforms the input scores into a sparse probability distribution to enhance interpretability and reduce noise, and its definition is:

[0033]

[0034] where is a (K - 1)-dimensional simplex;

[0035] After optimizing the subgraph structure, create a fixed-size graph-level embedding by concatenating the results of mean pooling and max pooling for each subgraph:

[0036]

[0037] Sum the embeddings of different subgraphs to form a comprehensive graph neural network embedding:

[0038]

[0039] Furthermore, explore the interaction relationships between features through a multi-head attention layer:

[0040]

[0041] In the formula, and represent the query, key, and value weights of the i-th head respectively. For the t-th record, the initial input, apply a per-token gating layer where is a trainable weight vector used to control the magnitude of each token embedding before linear projection, and the formula is as follows:

[0042]

[0043] where ⊙ represents element-wise multiplication, represents vector concatenation;

[0044] Obtain the Transformer outputs for all time steps:

[0045]

[0046] Furthermore, the formula for optimizing and characterizing the timing state using a gating mechanism in step S4 is:

[0047]

[0048] The output vector is:

[0049]

[0050] State vector c t+1 Is updated through the gating mechanism:

[0051] f t = σ(W f p t + b f )

[0052] i t = σ(W i p t + b i )

[0053]

[0054] In the formula, W f , W i and W h are trainable weight matrices, and b f , b i and b h are bias vectors.

[0055] Furthermore, in step S5:

[0056] Add the structure-related feature stru t and the time-related feature temp t as the input x t to the state space model. The state space model dynamically converts the input sequence into the output sequence whose dynamics are described by the following formula:

[0057]

[0058] y t = Ch t ,

[0059] where is the hidden state, is the input at time t, is the output.

[0060] Compared with the prior art, the advantages of the present invention are as follows: In view of the complexity of high-dimensional and long-sequence EEG data, the present invention utilizes multi-dimensional representation learning to capture the temporal, structural, and state correlations between multiple channels. By segmenting EEG signals into adjustable time segments and employing multiple graph convolutional networks, the structural relationships between different brain regions are effectively extracted. Meanwhile, in combination with the LSTM network and the attention mechanism, the historical states of EEG segments are modeled to ensure the retention of key information in long sequences. In addition, by adopting the Mamba-based state space model and selection mechanism, the model's ability to recognize important brain activity patterns related to depression detection is further enhanced, and the present invention achieves state-of-the-art results on the depression dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0062] Figure 1 It is a schematic flowchart of the multi-dimensional representation learning method for EEG signals for depression detection according to the present invention.

[0063] Figure 2 It is an overall framework diagram of the multi-dimensional representation learning method for EEG signals for depression detection according to the present invention.

[0064] Figure 3 It is a result visualization diagram of the multi-dimensional representation learning method for EEG signals for depression detection according to the present invention on the MODMA dataset.

[0065] Figure 4 It is a result visualization diagram of the multi-dimensional representation learning method for EEG signals for depression detection according to the present invention on the Xiangya dataset. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] The following will elaborate on the preferred embodiments of the present invention in conjunction with the drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0067] Refer to Figure 1-2 As shown, this embodiment discloses a multi-dimensional representation learning method for EEG signals for depression detection, including the following steps:

[0068] Step S1: Obtain the EEG signal dataset of patients, which includes EEG signals and corresponding depression labels. Among them, the EEG signals of the MODMA dataset are recorded using a 128-channel HydroCel Geodesic sensor net and Net Station acquisition software. The sampling frequency is 250Hz. The Xiangya dataset is collected using a 64-channel cap (Easy-cap 2.0) based on the 10-20 system, and the sampling frequency is 500Hz.

[0069] Step S2: Preprocess the EEG signals, which is to segment the EEG signals into a fixed length to meet the requirements of subsequent feature extraction and model analysis.

[0070] Step S3: Extract the structural correlation features (graph structure) of the EEG signals through the graph feature extraction and structure learning module, including the symmetric adjacency matrix and the GraphSAGE algorithm, to capture the spatial relationship between electrodes and the cross-hemisphere neural activity pattern.

[0071] Step S4: Adopt the gated Transformer mechanism to extract the temporal dynamic features of the EEG signals, and optimize and characterize the temporal state through the gating mechanism.

[0072] Step S5: Use the Mamba-based state space model to fuse the structural correlation features and the temporal dynamic features into the gating mechanism and optimize them. Adopt the dynamic gating mechanism to efficiently process long-range dependence information and optimize the context awareness ability of the model;

[0073] Step S6: Use UMAP technology to reduce the dimension of the fused features and generate an intuitive visualization result to facilitate the evaluation of the learning effect and feature distribution of the model in the depression detection task.

[0074] In this embodiment, the EEG dataset d eeg is defined as a set composed of EEG data X and the corresponding depression label Y, and is formally expressed as d eeg ={(X 1 , Y 1 ), (X 2 , Y 2 ),..., (X n , Y n )}, where n represents the total number of samples, where N e represents the number of electrodes, represents the electrode set, and col represents the length of the signal.

[0075] In this embodiment, to ensure that the model can adapt to dynamic input lengths, the original EEG signals are divided into T non-overlapping time-sliding windows based on the signal length, where T = signal length / time window size, and the time windows are indexed as {t 1 , t 2 ,..., t T}.

[0076] In this embodiment, a preliminary electroencephalogram structure M e is defined. In this structure, the electrodes v are used as nodes in the graph G e , and are connected according to their spatial relationships to form an initial adjacency matrix . Then, a symmetric adjacency matrix is constructed by symmetrically connecting the electrodes across hemispheres . This enhances the ability to detect cross-hemispheric neural activity patterns by exploring inter-hemispheric connections. The comprehensive adjacency matrix A e is calculated as follows:

[0077]

[0078] In this embodiment, to solve the high-dimensional problem while retaining all channel information, the GraphSAGE algorithm is adopted. This algorithm learns node embeddings by subsampling and aggregating the features of neighbor nodes, thereby reducing data complexity and computational requirements. For each node v in the graph structure G e = (V e , E e ), the average feature of its neighbors is calculated:

[0079] h N(v) = mean({e u | u ∈ N(v)})

[0080] Next, the feature e v of each node is concatenated with h N(v) to form a combined feature:

[0081] h v = e v ‖ h N(v)

[0082] This combined feature is transformed through a linear layer and processed by a ReLU activation function defined by the weights W and bias b:

[0083] h′ v = ReLU(W · h v + b)

[0084] Finally, the transformed node features h′ v are integrated into a matrix as the initial feature matrix for subsequent neural network processing.

[0085] In this embodiment, a graph convolutional network is used to capture the relationships between nodes. The basic operation of the graph convolutional network can be expressed as:

[0086]

[0087] where H (l) is the node feature representation of the l-th layer, and H (l+1) is the node feature representation of the (l + 1)-th layer. A e is the adjacency matrix of the graph, D is the degree matrix, W (l) is the weight matrix of the l-th layer, and σ is the activation function. In the formula, is used to normalize the adjacency matrix to maintain numerical stability when aggregating adjacent node features. The term H (l) W (l) represents feature transformation, that is, mapping the current features to a new feature space through the weight matrix. The activation function σ provides a non-linear mapping, enabling the network to learn complex patterns.

[0088] Considering that the graph pooling process may disconnect the connections of closely related nodes in the subgraph, weaken the integrity of the graph structure, and hinder the message passing process. This embodiment introduces a structure learning mechanism for optimizing the graph structure after pooling and encoding the potential pairwise relationships between nodes. This mechanism adopts a sparse attention method, where each subgraph

[0089] is generated after the k-th layer of pooling from the graph G is used as the input. The adjacency matrix i and the node feature matrix are input into a single-layer neural network. This network is parameterized by a weight vector . The neural network calculates the similarity score between nodes v and v p and v q as follows:

[0090]

[0091] where σ(·) is the activation function and λ is a parameter for balancing directly connected nodes and indirectly connected nodes. To standardize these similarity scores among different nodes, the sparse maximization function is applied to the entire vector :

[0092]

[0093] The sparse maximization function transforms the input scores into a sparse probability distribution, thereby enhancing interpretability and reducing noise. Its definition is:

[0094]

[0095] Among them, is a (K - 1)-dimensional simplex.

[0096] After optimizing the subgraph structure, by concatenating the results of mean pooling and max pooling for each subgraph, a graph-level embedding of a fixed size is created:

[0097]

[0098] The embeddings of different subgraphs are summed to form a comprehensive graph neural network embedding:

[0099]

[0100] This structure learning method ensures the integrity of the pooled graph, minimizes data noise, and enables clearer and more efficient information propagation through alternating operations of the graph convolutional network and the pooling layer.

[0101] In this embodiment, to capture the temporal state representation, first, the interaction relationship between features is explored through a multi-head attention layer:

[0102]

[0103] Among them, and represent the query, key, and value weights of the i-th head respectively. In particular, for the t-th record (time window), the initial input applies a token-wise gating layer Among them, is a trainable weight vector used to control the magnitude of each token embedding before linear projection. The formula is as follows:

[0104]

[0105] Among them, ⊙ represents element-wise multiplication, represents vector concatenation.

[0106] Finally, the Transformer outputs for all time steps are obtained:

[0107]

[0108] This embodiment introduces a Transformer-based temporal state representation method (TSR). TSR constructs and maintains a fixed-size state vector that summarizes the information the model has received so far. It takes the output of the Transformer and the current state vector c tPerform self-attention and dot-attention operations, and generate p through a linear projection layer after combining the outputs t . This process is summarized as follows:

[0109]

[0110] The output vector is:

[0111]

[0112] The state vector c t+1 is updated through a gated mechanism:

[0113] f t = σ(W f p t + b f )

[0114] i t = σ(W i p t + b i )

[0115]

[0116] where W f , W i and W h are trainable weight matrices, and b f , b i and b h are bias vectors.

[0117] In this embodiment, in step S5, the structure-related feature stru t and the time-related feature temp t are added as the input x t to the state space model. The state space model dynamically converts the input sequence into an output sequence whose dynamics are described by the following formula:

[0118]

[0119] y t = Ch t ,

[0120] where is the hidden state, is the input at time t, is the output.

[0121] To combine structural and temporal dependencies, this embodiment introduces a gating mechanism that enables the model to selectively update its hidden state based on the importance of new input information. Specifically, the gating vector is defined as:

[0122] g t = σ(W g (stru t + temp t ) + b g )

[0123] where and are trainable parameters, and σ is the Sigmoid activation function.

[0124] The hidden state is updated through the gating mechanism as:

[0125]

[0126] where ⊙ denotes element-wise multiplication, is the candidate hidden state, and the calculation formula is:

[0127]

[0128] By integrating structural and temporal features into the gating mechanism, the model can dynamically filter and emphasize relevant contexts, thereby effectively handling long-range dependencies while maintaining computational efficiency.

[0129] To verify the effectiveness of the present invention, Table 1 presents the comparison results of the model constructed in this embodiment with other state-of-the-art methods on two datasets. Using accuracy (ACC), F1-score, recall (REC), and precision (PRE) as evaluation metrics, and comparing with the previous model results, it can be seen that the performance of the model trained using the present invention is significantly better than the existing model methods in these two datasets. Figure 3 and Figure 4 are the clustering visualizations of the features extracted by the present invention on the MODMA and Xiangya datasets, clearly showing the separability of the healthy class (HC) and the depressive class (MDD). Among them, the left side is the training set and the right side is the test set.

[0130] Table 1 Performance of Different Models on MODMA and Xiangya Datasets

[0131]

[0132] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, the patent owner may make various deformations or modifications within the scope of the appended claims. As long as they do not exceed the protection scope described in the claims of the present invention, they should be within the protection scope of the present invention.

Claims

1. A method for learning multidimensional representation of EEG signals for depression detection, characterized in that: The following steps are involved: S1. Obtain a patient's EEG signal dataset, which includes EEG signals and corresponding depression labels; S2, preprocessing the EEG signal, wherein the preprocessing is to segment the EEG signal into segments of fixed length; S3, extracting structural correlation features of EEG signals through graph feature extraction and structure learning modules; S4, using the gated Transformer mechanism to extract the temporal dynamic characteristics of EEG signals, and optimizing and characterizing the timing state through the gating mechanism; S5. Using the Mamba-based state-space model, the structural correlation features and temporal dynamic features are fused and optimized into the gating mechanism; S6. Use UMAP technology to reduce the dimension of the fused features and generate intuitive visualization results.

2. The method for learning multidimensional representation of EEG signals for depression detection according to claim 1, characterized in that: The formula of the patient's EEG signal data set in step S1 is d eeg ={(X1, Y1), (X2, Y2),..., (X n , Y n )}, X is the EEG signal, Y is the corresponding depression label, n is the total number of samples, Where N e represents the number of electrodes, represents the electrode set, and col represents the length of the signal.

3. The method for learning multidimensional representation of EEG signals for depression detection according to claim 1, characterized in that: In step S2, the EEG signal is divided into T non-overlapping time sliding windows, where T = signal length / time window size, and the time windows are {t1, t2, ..., t T } for indexing.

4. The method for learning multidimensional representation of EEG signals for depression detection according to claim 1, characterized in that: In step S3, the structural correlation feature of the EEG signal is defined as M e , electrode v as the graph structure G e The nodes in are connected according to the spatial relationship to form the initial adjacency matrix Then, a symmetric adjacency matrix is ​​constructed by connecting electrodes symmetrically across the hemisphere. Comprehensive adjacency matrix 5. The method for learning multidimensional representation of EEG signals for depression detection according to claim 4, characterized in that: For the graph structure G e =(V e , E e ), calculate the average feature of its neighbors: h N(v) =mean({e u |u∈N(v)}) The feature e of each node v The average feature h of its neighbors N(v) Connect them to form a combined feature: h v =e v ||h N(v) The combined features are transformed by the linear layer and processed by the ReLU activation function defined by the weight W and bias b: h' v =ReLU(W·h v +b) The transformed node feature h′ v is integrated into a matrix As the initial feature matrix for subsequent neural network processing.

6. The method for learning multidimensional representation of EEG signals for depression detection according to claim 5, characterized in that: Graph convolutional networks are used to capture the relationship between nodes. The basic operation of graph convolutional networks is expressed as: in, H (l) is the node feature representation of the lth layer, H (l+1) is the node feature representation of the l+1th layer, A e is the adjacency matrix of the graph, D is the degree matrix, and W (l) is the weight matrix of the lth layer, σ is the activation function, in the formula, Used to normalize the adjacency matrix to maintain numerical stability when aggregating adjacent node features, H (l) W (l) Represents feature transformation, that is, mapping the current feature to the new feature space through the weight matrix; the activation function σ provides nonlinear mapping.

7. The method for learning multidimensional representation of EEG signals for depression detection according to claim 6, characterized in that: It also includes adopting a structural learning mechanism to optimize the graph structure after pooling and encode the potential pairwise relationships between nodes, including: The structure learning mechanism adopts a sparse attention method, where each subgraph From Figure G i After the kth layer of pooling, the adjacency matrix is ​​generated as input. and the node feature matrix Input a single-layer neural network, which consists of a weight vector Parameterized, neural network computing node v p and v q The similarity score between them is as follows: Among them, σ(·) is the activation function, λ is the parameter that balances directly connected nodes and indirectly connected nodes, and for the entire vector Apply the sparse maximization function: The sparse maximization function transforms the input scores into a sparse probability distribution to enhance interpretability and reduce noise, which is defined as: in, is a K-1 dimensional simplex; After optimizing the subgraph structure, a fixed-size graph-level embedding is created by concatenating the results of mean pooling and max pooling for each subgraph: The embeddings of different subgraphs are summed to form a comprehensive graph neural network embedding:

8. The method for learning multidimensional representation of EEG signals for depression detection according to claim 6, characterized in that: Explore the interaction between features through multi-head attention layers: In the formula, and Represent the query, key, and value weights of the i-th header, respectively. For the t-th record, the initial input, the token-by-token gating layer is applied. Where is a trainable weight vector used to control the magnitude of each token embedding before linear projection, and the formula is as follows: Among them, ⊙ represents element-wise multiplication, Represents vector concatenation; Get the Transformer output for all time steps:

9. The method for learning multidimensional representation of EEG signals for depression detection according to claim 6, characterized in that: The formula for optimizing and characterizing the timing state using the gating mechanism in step S4 is: The output vector is: State vector c t+1 Updates via a gated mechanism: f t =σ(W f p t +b f ) I t =σ(W i p t +b i ) Where W f , W i and W h is a trainable weight matrix, b f , b i and b h is the bias vector.

10. The method for learning multidimensional representation of EEG signals for depression detection according to claim 9, characterized in that: In step S5: The structure-related features stru t and time-related features temp t Added together, as the input x of the state-space model t , the state-space model transforms the input sequence into Convert to output sequence Its dynamics is described by the following formula: the t =Ch t , in, is the hidden state, is the input at time t, is the output.

Images