Depression assessment method based on multi-view subgraph construction

Abstract

The invention discloses a depression assessment method based on multi-view subgraph construction, which is applied to the field of electroencephalogram data processing and aims at solving the problem that the accuracy and reliability of emotion recognition are limited due to the fact that comprehensive brain function structure features cannot be effectively extracted from multiple dimensions in the prior art. According to the method, firstly, a graph construction mode based on Euclidean distance, time-frequency analysis and nonlinear features is introduced, and the model can perform deep abstraction on signals of the brain from multiple angles, so that feature loss possibly caused by a single-view-angle analysis method is effectively avoided, and the adaptability to the complexity and diversity of brain activities is improved. Wherein the nonlinear analysis helps to excavate a potential complex dynamic relationship in the signal, and helps to reveal the nonlinear change of the brain in a specific mental state. Finally, the construction of a multi-view image not only enhances the recognition of functional connection between brain areas, but also improves the sensitivity to different electroencephalogram activity states.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) data processing, and specifically relates to a depression assessment technology. Background Technology

[0002] Depression is a common and serious mood disorder, typically characterized by persistent low mood, loss of interest in daily activities, fatigue, difficulty concentrating, sleep disturbances, and negative self-esteem. This mood state not only affects a patient's mental health but can also lead to significant functional impairment, affecting multiple aspects of work, social life, and personal life. Symptom presentation can vary greatly from individual to individual, and the course of the illness can manifest as acute onset, chronicity, or relapse. Therefore, depression assessment is particularly important in this context. It not only helps to clarify the severity and clinical characteristics of depression but also helps to develop precise treatment plans, monitor treatment effectiveness, optimize the treatment process, and thus provide a scientific basis for patient recovery.

[0003] Current clinical methods for assessing depression primarily include clinical symptom observation and standardized scale assessment. In terms of clinical symptom observation, physicians observe patients' symptoms in various aspects, such as sadness, interests, attention, physical condition, and sleep patterns, to determine depressive symptoms. Regarding standardized scale assessment, there are relatively applicable assessment tools, such as the Montgomery Depression Rating Scale, the Hamilton Depression Rating Scale, and the Beck Depression Rating Scale. However, the accuracy of these assessments depends on factors such as the assessor's clinical experience and the patient's cooperation, which may introduce potential errors during the diagnostic process. Therefore, researchers are exploring the use of multidimensional information, including signals, speech, brain signals, facial features, and social media text, i.e., assessment through machine learning and deep learning models, to improve the accuracy and objectivity of identifying brain depression.

[0004] Chinese patent CN115670463B, "A Depression Detection System Based on EEG Emotional Neurofeedback Signals," extracts emotional features from preprocessed EEG signal fragments using short-time Fourier transform and differential entropy, and then assesses the user's depression level using a Gaussian kernel Naive Bayes classification model. By utilizing the emotional characteristics or symptoms reflected in the EEG emotional neurofeedback signals, depression detection can be performed on individuals requiring screening, assisting medical professionals in diagnosis and treatment.

[0005] Chinese patent "CN112971781B A Depression Comparison Device Based on Brain Functional Networks" collects multi-channel EEG data from normal individuals and patients with depression while they are listening to music at rest. Then, using brain functional network modeling methods, an association matrix is ​​established to extract brain functional network features reflecting neural activity. Based on this, the correlation between brain functional network features such as average clustering coefficient, average path length, and degree distribution and the mechanism of depression is analyzed. Order relation analysis is used to determine the weights of relevant features, yielding the depression level results. The invention considers the characteristic that EEG signals can translate the state of brain functional connectivity and uses the Pearson correlation coefficient to calculate the correlation between various EEG channels.

[0006] The above research methods provide a new approach to depression assessment, but they have the following limitations:

[0007] Each electrode in an EEG corresponds to a specific brain region, capturing the unique neuronal activity of that region. Existing machine learning and deep learning methods often focus on the strong temporal and frequency characteristics of signals, without considering the relationship between individual electrode stability and local functional state and the intensity of depression.

[0008] Existing depression assessment studies based on signal data typically focus on the time domain, frequency domain, and time-frequency features of signals during feature extraction; they fail to fully capture and consider the spatiotemporal information and nonlinear features in EEG signals.

[0009] Existing graph construction methods can capture features from corresponding angles to some extent, but such one-dimensional analysis leads to the loss of a large amount of original EEG signal feature information, making it impossible to effectively extract comprehensive brain functional structure features from multiple dimensions, thus limiting the accuracy and reliability of emotion recognition. Summary of the Invention

[0010] To address the aforementioned technical issues, this invention proposes a depression assessment method based on multi-view subgraph construction. This method comprehensively considers single-electrode signals and local brain region information, combining the spatial location relationship, time-frequency characteristics, and nonlinear characteristics of EEG signals to construct multiple subgraph models. By fusing multi-dimensional features, it improves the assessment effect of depressive states, providing a more scientific and practical basis for the diagnosis and intervention of depression.

[0011] The technical solution adopted in this invention is: a depression assessment method based on multi-view subgraph construction, comprising:

[0012] S1. Preprocess the collected EEG data from each channel;

[0013] S2. Based on the preprocessed EEG data, three sub-graphs are constructed: a graph based on two-dimensional location distance, a graph based on time-frequency similarity, and a graph based on nonlinear analysis.

[0014] S3. Input the three sub-graphs into the hybrid graph neural network to obtain brain region fusion features;

[0015] S4. Train the emotion assessment model based on the brain region fusion features obtained in step S3.

[0016] S5. Input the brain region fusion features obtained after processing the subject in steps S1 to S3 into the trained emotion assessment model to obtain the predicted depression score of the subject.

[0017] The beneficial effects of this invention are as follows: By integrating EEG signal features from multiple dimensions, this invention comprehensively enhances the model's ability to capture brain functional structures. Firstly, by introducing a graph construction method based on Euclidean distance, time-frequency analysis, and nonlinear features, the model can deeply abstract brain signals from multiple perspectives, effectively avoiding feature loss that may occur with single-view analysis methods and improving its adaptability to the complexity and diversity of brain activity. Nonlinear chaotic analysis helps to uncover the potential complex dynamic relationships within the signals, contributing to the revelation of nonlinear changes in the brain under specific mental states. Finally, the construction of multi-view graphs not only strengthens the identification of functional connections between brain regions but also improves the sensitivity to different EEG activity states. Attached Figure Description

[0018] Figure 1 This is a flowchart of the solution of the present invention;

[0019] Figure 2 This is a schematic diagram showing the location of the brain electrodes according to the present invention;

[0020] Figure 3 This is a schematic diagram of the two-dimensional embedding matrix of the present invention;

[0021] Figure 4 This is a block diagram of the present invention. Detailed Implementation

[0022] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0023] To improve the accuracy and reliability of depression assessment, this invention uses a multi-view graph construction module to more comprehensively abstract brain relationships from three different perspectives: Euclidean distance, time-frequency linearity, and nonlinearity.

[0024] 1) Pseudo-fully connected graphs based on traditional Euclidean distance. This type of graph has a fixed structure and cannot dynamically capture functional changes in brain activity. It is based on considerations of the location characteristics of various electrodes in the brain.

[0025] 2) Construction of a time-frequency similarity map based on the signal. First, time-frequency analysis (wavelet transform) is performed on the preprocessed signal to extract fine-grained and coarse representations based on the electrode signals. Then, Pearson similarity maps between electrodes are constructed based on considerations of the linear features and functional connectivity of various electrodes in the brain.

[0026] 3) Construction graph based on nonlinear characteristics. Nonlinear analysis is performed on each electrode, and its reconstructed phase matrix is ​​obtained. The optimal embedding dimension m and optimal time delay are then determined. If the similarity exceeds a set threshold, the two channels are considered related, based on the brain's nonlinear characteristics (m) and dynamic temporal characteristics (m). (This is a consideration.)

[0027] like Figure 1 As shown, the method of the present invention includes the following steps:

[0028] 1: Data Preprocessing

[0029] In this embodiment, EEG data from the SEED dataset is used, and the electrode distribution is set according to the 62-channel EEG settings of the international 10-20 system.

[0030] 11. Preprocessing of EEG data using EEGlab:

[0031] EEG signals ,in EEG data representing the i-th channel, , Representing the Time points within the channel The EEG signal values ​​are represented by N, where N represents the length of the EEG signal. Preprocessing includes bandpass filtering, artifact removal, and independent component analysis (ICA) denoising.

[0032] 12. According to Figure 2 The physiological brain regions, as shown in the neuroanatomical diagram, were divided into 62 electrode zones to obtain the corresponding electrodes for each brain region:

[0033] Left parietal lobe region (Parietal Lobe-L):

[0034] fc5,fc3,fc1,fcz, c5,c3,c1,cz, cp5,cp3,cp1

[0035] Parietal Lobe-R (Right):

[0036] fc2,fc4,fc6, c2,c4,c6,cz, cp2,cp4,p6,cpz

[0037] Left frontal lobe (Central Lobe-L):

[0038] F7,f5,f3,f1,fz, af7,af3, fp1,fpz,

[0039] Right frontal lobe (Central Lobe-R):

[0040] f2,f4,f6,f8; afz,af4,af8; fpz,fp2

[0041] Left temporal lobe (L):

[0042] ft7, t7, tp7, t9

[0043] Right temporal lobe (R):

[0044] ft8, t8, tp8, t10

[0045] Left occipital lobe region (L):

[0046] p7,p5,p3,p1,pz, poz,po7,po3,o1,oz,iz

[0047] Right occipital lobe region (R):

[0048] pz,p2,p4,p6,p8, poz,po4,po8, oz,o2

[0049] 2: Multi-view graph construction:

[0050] Existing graph construction methods can capture features from corresponding angles to some extent, but single-dimensional analysis leads to the loss of a large amount of original EEG signal feature information, making it impossible to effectively extract comprehensive brain functional structural features from multiple dimensions, thus limiting the accuracy and reliability of emotion recognition. Therefore, this embodiment abstracts brain relationships more comprehensively from three different angles: Euclidean distance, time-frequency linearity, and nonlinearity.

[0051] 21. Graph construction based on two-dimensional location distance:

[0052] First, a two-dimensional matrix embedding is performed on the EEG signals, with each electrode serving as a graph node, and its position is... Each graph node connects only to its four nearest other nodes. In practical applications, the number of nearest other nodes can be adjusted according to specific circumstances. A schematic diagram of the two-dimensional embedding matrix of brain electrodes is shown below. Figure 3As shown, the formula for determining the distance between two different nodes is as follows:

[0053]

[0054] in, Indicates the relationship between two different nodes Distance between axes Indicates the relationship between two different nodes Distance between axes;

[0055] This method avoids noise interference from fully connected graphs while preserving important local spatial features, resulting in node feature maps. Its corresponding adjacency matrix .

[0056] 22. Graph construction based on time-frequency similarity:

[0057] First, analyze the original signal. The goal of performing DWT (Discrete Wavelet Transform) wavelet decomposition is to decompose the EEG signal of each channel. Decomposed into several layers of high-frequency detail coefficients and low-frequency approximation coefficient This allows for the effective representation of the time-frequency characteristics of a signal.

[0058]

[0059] Where t is the time index of the original signal of a channel, j is the scale (or level), and k is the frequency (or position), which represents the index of the wavelet coefficients (the transformed position, the index after downsampling). Let t be the signal value at the t-th time point of the original signal sequence. For high-pass filters, used to generate high-frequency detail coefficients, g It is a low-pass filter used to generate low-frequency approximation coefficients.

[0060] Next, the high-frequency detail coefficients were analyzed. Soft thresholding is performed to remove noise. For the high-frequency components of the signal, smaller detail coefficients, such as noise, are removed while larger signal components are retained. This yields the detail coefficients after noise removal. The formula is as follows:

[0061]

[0062] in, It is a threshold parameter used to control the size of the threshold, which determines the detail coefficients. The degree of weakening, λ= , It is the standard deviation of noise. It is the signal length after wavelet decomposition (i.e., the total number of wavelet detail coefficients). It is the natural logarithm.

[0063] Then, using the processed detail coefficients and approximation coefficients Perform inverse DWT restoration to reconstruct the signal ,in Let represent the reconstructed signal of the i-th channel. The reconstructed signal process is as follows:

[0064]

[0065] Finally, the Pearson Correlation Coefficient (PCC) is used to calculate the similarity between each denoised and reconstructed signal and all other signals. For example, the denoised and reconstructed signal of channel i is calculated using Equation (4). With channel The signal after denoising and reconstruction as shown in formula (4) The formula for calculating the PCC coefficient is as follows:

[0066]

[0067] in, It is the length of the reconstructed signal; It is a passage and channels Pearson correlation coefficient between reconstructed signals; It is the first Reconstruction signals of each channel In the The value at each time point; It is the first Reconstruction signals of each channel In the The value at each time point; and They are the first The first channel and the first The average reconstructed signal of each channel is calculated using the following formula:

[0068]

[0069]

[0070] The four channels with the highest similarity are used as the connection objects during graph construction. The initial information of each node is the reconstructed signal, resulting in a node feature map. Its corresponding adjacency matrix .

[0071] 23. Graph construction based on nonlinear analysis:

[0072] 231. Channel-level nonlinear analysis is performed on each EEG channel by embedding the timing sequence of each EEG electrode into a high-dimensional space and reconstructing the phase space to reveal the complex nonlinear patterns and dynamic characteristics in the EEG signal, and to obtain two parameters that can represent the level of its nonlinear characteristics. and The phase space reconstruction steps are as follows:

[0073] The first step is to determine the time delay. EEG signals It is represented in two parts: and , , It is a predefined maximum delay value, not exceeding half of the signal. This is the first part of the signal, from time point 1 to time point 2. , The signal is from time... arrive The data, the length of both parts is , It is an electroencephalogram (EEG) signal. The length.

[0074] The second step involves calculating the mutual information of the signals using the discretized histogram method. Specifically, this is done using the matplotlib utility package in Python. Function calculation and The two-dimensional histogram is obtained. and joint probability distribution , , Then in , The above are respectively for and Summing along the axis to calculate the marginal distribution. and Then calculate the signal In time delay Mutual information below:

[0075] (8)

[0076] The third step is to calculate each delay. The mutual information values ​​under different conditions are compared, and the delay corresponding to the first local minimum mutual information value is selected. As the optimal delay for this channel The "time interval" between components is determined when reconstructing the phase space. Too small a value indicates high component correlation, i.e., information redundancy; conversely, a large value indicates high component correlation. If the value is too large, it indicates that the components are completely unrelated, that is, they lose their dynamic structure. The selection principle is as follows:

[0077]

[0078] The fourth step is to use the nearest neighbor method to determine the appropriate embedding dimension. .

[0079] First, the signal According to delay Phase space reconstruction is obtained , Each point is an m-dimensional vector, and each component of the vector is a value from the original time series at intervals of 1 / m. The point. Among them. The range of values ​​for is [1, (m-1)]. ], ensuring the last component It does not exceed the original signal length N.

[0080] Next, in the current m-dimensional space, we find the nearest neighbor for each point, that is, we traverse the phase space. Each point in To find another point in phase space This makes this point and The Euclidean distance is minimized, and the point is... Recorded as a point The nearest neighbor point, and from this, the minimum distance can be obtained. .

[0081] All the above calculations are performed in dimension m. Now, the system is embedded into dimension m+1 to examine distance changes. The same time series signal... Using the same However, if the dimension is m+1, then the reconstruction is obtained. That is, in An additional (m+1)th vector was added to the base vector. For every point in the m-dimensional phase space... and its nearest neighbor Find the corresponding new point in the m+1 dimensional phase space. and .

[0082] Calculate the proportion of spurious nearest neighbors:

[0083] Traverse all points The proportion of points with false nearest neighbors was calculated. :

[0084]

[0085]

[0086] in, Let be the total number of points in the m-dimensional phase space. Is the current point at Coordinates in the dimensional embedding space; It is its nearest neighbor point. Coordinates in the dimensional embedding space; It is the nearest neighbor Distance in 3D space. If Greater than a set threshold (Usually 10-15), indicating that when a dimension is added, these two previously close points suddenly become very far apart. This suggests that their "proximity" in m-dimensional space is false, caused by overlapping trajectories when projected into lower-dimensional space. Therefore, points There is a "false nearest neighbor".

[0087] The above operations all correspond to dimensions. Therefore, candidate dimensions are continuously increased. ( ), and repeat steps - to calculate each corresponding .along with As the number of false nearest neighbors increases, the proportion of false nearest neighbors will decrease. The first time it drops below the threshold (e.g., 0.05) corresponds to As the optimal embedding dimension.

[0088] Based on the determined time delay and optimal embedding dimension Reconstructing phase space:

[0089]

[0090] in, It is a reconstructed phase space trajectory matrix, where each row represents a point in phase space, containing a vector composed of sampling points of the time series at different time lags. Through this matrix, the original one-dimensional time series is mapped to a high-dimensional space, which can better reveal the dynamic behavior and complex structure of the signal.

[0091] 232. After obtaining the reconstructed phase space corresponding to each channel, three types of nonlinear features are extracted: (1) recursive quantization analysis features (vector): analyzing the complex dynamic characteristics of the time series; (2) multifractal spectrum features (vector): revealing multi-scale characteristics; (3) maximum Lyapunov exponent (single-value scalar): reflecting the sensitivity of the reconstructed phase space to the initial conditions. The three types of nonlinear features are concatenated into a vector as the initial value of each channel node in the nonlinear component diagram. The nonlinear feature of the i-th channel is concatenated as follows: . These are the recursive quantization analysis characteristics, multifractal spectrum characteristics, and maximum Lyapunov index of this channel, respectively.

[0092] 233. During the phase space reconstruction process, signals from each channel were obtained. Optimal Dimension of Nonlinear Embedding and optimal latency Among them, the embedding dimension This ensures that the reconstruction of the phase space can fully capture the complexity and degrees of freedom of the system. If If it's too small, it might lead to an imbalance in the reconfiguration space, failing to accurately represent the system's complex dynamics. If... Too large an amount will increase computational complexity and may lead to overfitting or data sparsity problems.

[0093] For appropriate delay The choice of time delay ensures that the reconstructed state vector captures the nonlinear dynamic characteristics of the system, including the time dependence and periodicity of the signal. Too short a time delay may lead to excessively high correlation between adjacent time points, resulting in an overly compact dimensional space for the reconstruction. Too long a time delay may lead to information loss or confusion, failing to accurately represent the system's long-term behavior.

[0094] Therefore, when constructing a graph from a non-linear perspective, the optimal embedding dimension is... and optimal latency It is a good parameter to help reveal the dynamic behavior and structural characteristics of each channel.

[0095] Due to the optimal dimension and optimal latency The dimensions are different, so first determine the optimal dimension. and optimal latency Standardize to the same scale:

[0096]

[0097] in, and These are the overall mean values ​​for the embedding dimension and the optimal latency, respectively. and The embedding dimension and optimal latency standard deviation are calculated using the following formula, where, For channel indexing.

[0098]

[0099] 234. Calculate the similarity of nonlinear parameters between standardized channels. Treat each channel as a two-dimensional point, using the optimal dimension. Using the optimal time delay as the two-dimensional coordinate system, the coordinates of any channel i and channel z are: and Then, calculate the distance between each point using Euclidean distance:

[0100]

[0101] The four channels with the highest similarity are used as the objects for edge connections during graph neural network construction. The initial information of each node is the non-linear features extracted in step 232. , obtain node feature map Its corresponding adjacency matrix .

[0102] 3: Hybrid Graph Neural Network Module

[0103] Each electrode in an electroencephalogram (EEG) corresponds to a specific brain region and captures neuronal activity specific to that region. Therefore, data from a single electrode can provide unique information about a particular brain region, which can have a profound impact on understanding the connectivity between local and overall brain function.

[0104] First, we use a basic Graph Neural Network (GNN) layer to... The nodes undergo initial feature learning and propagation. This layer helps capture graph structure information and node positional features, thus enabling the first layer to learn and transmit these features. The GNN output corresponding to each brain region is:

[0105]

[0106] Next, the time-frequency similarity graph is constructed. and nonlinear composition diagram Meanwhile, the input graph convolutional network layer GCN further processes the features of nodes through convolution operations, aggregating information from local neighborhoods to enhance the expression of globally consistent features.

[0107]

[0108]

[0109] Then, using the graph attention network layer GAT, the output is obtained, the first... The GAT output for each brain region is:

[0110]

[0111] This layer dynamically assigns attention to the output of the GCN layer. and By assigning different weights to neighboring nodes, graph attention networks can further refine feature learning and graph structure modeling. Graph attention networks can adaptively focus on information important to nodes, extracting more complex non-linear features. The time-frequency features in the model also contain nonlinear factors, allowing for the extraction of finer-grained features based on capturing the global structure, thus improving the model's expressiveness. Finally, and By splicing, we obtain the first... The feature fusion results corresponding to each brain region are

[0112]

[0113] 4: Training Module

[0114] This module aims to fully explore the relationships between different brain regions of patients without node-level labels, to achieve personalized assessment and final prediction of depression level. The total loss function is as follows:

[0115]

[0116] in,

[0117] Contrast loss is used to mine semantic similarity between brain regions;

[0118] : Weakly supervised regression loss of nodes within key brain regions;

[0119] : Based on the final regression loss of the overall depression score after significant electrode aggregation;

[0120] and These are weight parameters used to adjust... and The proportion of total losses.

[0121] First, construct positive and negative sample pairs for each subject. Let its brain region set be { , ,..., The similarity between two brain regions is calculated using cosine similarity, and pairs with a similarity greater than 0.7 are considered positive samples, while pairs with a similarity less than 0.3 are considered negative samples.

[0122] Subsequently, the feature vectors of each brain region were analyzed. d Add a projection head to obtain the mapping vector:

[0123]

[0124] Where ReLU is the activation function. and For learnable weights, and This is the bias value.

[0125] For each positive sample pair ( The contrastive loss is defined using the following loss function:

[0126]

[0127] in, Cosine similarity; It is a temperature parameter; Indicates and All relevant samples, including both positive and negative samples (excluding themselves). The average loss for all positive sample pairs is:

[0128]

[0129] in: It is the set of all positive sample pairs. .

[0130] Furthermore, this invention utilizes highly similar brain region pairs mined during the contrastive learning phase to construct a set of key brain regions. As prior structural information, it is assumed that these regions play a central role in task-related representations. To enhance the model's focus on these key regions, this invention designs a weakly supervised strategy based on key brain regions. By applying regression loss only to their internal electrode nodes, the model is guided to focus on regions highly correlated with depression levels, thereby effectively suppressing interference from non-task-related brain regions and improving the model's generalization ability and interpretability.

[0131] Specifically, each key brain region is identified. It contains several electrode nodes, and its original characteristics are: The corresponding brain region-level representation is a semantic vector after contrastive learning. Each electrode node in the same brain region They all share the same input Each node independently predicts its estimate of the overall depression score. While each node is trained independently, the training objective is to identify nodes that are intrinsically more salient to the task in key brain regions.

[0132]

[0133] To guide this process, this embodiment uses the patient's overall depression score Y as a pseudo-label to construct a node-level weakly supervised regression loss:

[0134] in Indicates key brain regions The set of electrode nodes included This indicates the number of key brain regions in the set. Through training, this regression module can identify which electrodes in key brain region K are more closely associated with overall depression levels, i.e., by comparing key brain regions... By analyzing the regression results of each electrode node, the electrodes that have a more significant impact on the overall level of depression are identified, enabling more granular and personalized quantification of brain region task participation.

[0135] Finally, the regression output of significant electrodes in all key brain regions K was calculated. The data is aggregated and input into a fully connected neural network (MLP) to predict the patient's overall depression score.

[0136]

[0137] in, The regression output of the i-th significant electrode representing a key brain region indicates the involvement or contribution of this brain region in the degree of depression, and s represents the number of key electrodes. It is a learnable nonlinear mapping function. The regression loss for the depression score is as follows:

[0138]

[0139] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A depression assessment method based on multi-view subgraph construction, characterized in that, include: S1. Preprocess the collected EEG data from each channel; S2. Based on the preprocessed EEG data, three sub-graphs are constructed: a graph based on two-dimensional location distance, a graph based on time-frequency similarity, and a graph based on nonlinear analysis. S3. Input the three sub-graphs into the hybrid graph neural network to obtain brain region fusion features; S4. Train the emotion assessment model based on the brain region fusion features obtained in step S3. S5. Input the brain region fusion features obtained after processing the subject in steps S1 to S3 into the trained emotion assessment model to obtain the predicted depression score of the subject.

2. The depression assessment method based on multi-view subgraph construction according to claim 1, characterized in that, The preprocessing in step S1 includes: performing bandpass filtering, artifact removal, and independent component analysis on the collected EEG data from each channel in sequence.

3. The depression assessment method based on multi-view subgraph construction according to claim 2, characterized in that, Step S1 also includes dividing the distribution of acquisition electrodes according to the physiological brain regions defined by brain neuroanatomy, and obtaining the electrodes corresponding to each brain region.

4. The depression assessment method based on multi-view subgraph construction according to claim 3, characterized in that, The graph construction process based on two-dimensional location distance in step S2 is as follows: Two-dimensional matrix embedding is performed on the EEG data preprocessed in step S1, with each electrode as a graph node. Each graph node is connected only to its four nearest other nodes, resulting in a node feature map. Its corresponding adjacency matrix .

5. The depression assessment method based on multi-view subgraph construction according to claim 4, characterized in that, The graph construction process based on time-frequency similarity in step S2 is as follows: A1. Perform DWT wavelet decomposition on the EEG data of all channels after preprocessing in step S1. The goal is to decompose the EEG signal of each channel into several layers of high-frequency detail coefficients. and low-frequency approximation coefficient ; A2. High-frequency detail coefficients Soft thresholding is performed to obtain the detail coefficients after noise removal. ; A3. Utilization and Perform inverse DWT restoration to reconstruct the signal; A4. For each reconstructed signal, the similarity between it and all other channel signals is calculated using the Pearson correlation coefficient. A5. Connect the four channels with the highest similarity to this channel during graph construction. The initial information of the nodes is the reconstructed signal, resulting in the node feature map. Its corresponding adjacency matrix .

6. The depression assessment method based on multi-view subgraph construction according to claim 5, characterized in that, The graph construction process based on nonlinear analysis in step S2 is as follows: B1. Perform channel-level nonlinear analysis on the EEG data of each channel after preprocessing in step S1. This is done by embedding the timing sequence of each EEG electrode into a high-dimensional space and reconstructing the phase space, and obtaining two parameters that can represent the level of its nonlinear characteristics: optimal dimension. and optimal latency ; B2. After obtaining the reconstructed phase space corresponding to each channel, extract three types of nonlinear features: recursive quantization analysis features, multifractal spectrum features, and the maximum Lyapunov exponent. These three types of nonlinear features are concatenated into a vector and used as the initial value for each channel node in the nonlinear construction graph; B3, Optimal Dimensions and optimal latency Standardize to the same scale; B4. Calculate the similarity of nonlinear parameters between standardized channels. Use the four channels with the highest similarity as edge connections in the graph neural network during graph construction. The initial information for each node is the vector concatenated in step B2, thus obtaining the node feature map. Its corresponding adjacency matrix .

7. The depression assessment method based on multi-view subgraph construction according to claim 6, characterized in that, The process of reconstructing the phase space is as follows: C1. Based on delay EEG signals It is represented in two parts: and , , It is a predefined maximum delay value, not exceeding half of the signal. This is the first part of the signal, from time point 1 to time point 2. Data, The signal is from time... arrive The data, the length of both parts is , It is an electroencephalogram (EEG) signal. Length; C2. Calculate the mutual information of the signal using the discretized histogram method, specifically by using the matplotlib utility package in Python. Function calculation and The two-dimensional histogram is obtained. and joint probability distribution , , Then in , The above are respectively for and Summing along the axis to calculate the marginal distribution. and Then calculate the signal. In time delay Mutual information below: ； C3. Calculate each delay The mutual information values ​​under different conditions are compared, and the delay corresponding to the first local minimum mutual information value is selected. As the optimal delay for this channel; C4. Use the nearest neighbor method to determine the optimal embedding dimension. ; C41, First, transfer the brainwave signals As determined in step C3 Phase space reconstruction is obtained , Each point is an m-dimensional vector, and each component of the vector is a value from the original time series at intervals of [missing value]. The point, among which The range of values ​​for is [1, (m-1)]. ], ensuring the last component Not exceeding the original signal length N; C42. Next, in the current m-dimensional space, find the nearest neighbor for each point, that is, traverse the phase space. Each point in To find another point in phase space This makes this point and The Euclidean distance is minimized, and the point is... Recorded as a point The nearest neighbor point, and from this, the minimum distance can be obtained. ; C43. Embed the system into the m+1 dimension and examine distance changes; use the same time series signal... Using the same However, if the dimension is m+1, then the reconstruction is obtained. That is, in An (m+1)th vector was added to the base vector. ; For every point in the m-dimensional phase space and its nearest neighbor Find the corresponding new point in the m+1 dimensional phase space. and ; C44. Calculate the proportion of spurious nearest neighbors: Traverse all points The proportion of points with false nearest neighbors was calculated. : ； ； in, Let be the total number of points in the m-dimensional phase space. Is the current point at Coordinates in the dimensional embedding space; It is its nearest neighbor point. Coordinates in the dimensional embedding space; It is the nearest neighbor Distance in 3D space; if Greater than a set threshold This explains the point. There is a "false nearest neighbor"; C45. Continuously increasing candidate dimensions And repeat steps C41-C44 to calculate each corresponding ;along with As the number of false nearest neighbors increases, the proportion of false nearest neighbors decreases; selection The first time it drops below the threshold As the optimal embedding dimension.

8. The depression assessment method based on multi-view subgraph construction according to claim 7, characterized in that, Step S3 includes the following sub-steps: S31, will Inputting a graph neural network layer GNN, the resulting layer is the first... The output corresponding to each brain region is denoted as ; S32, will and Input a graph convolutional network layer (GCN); the output is denoted as... and ; S33, will and Input the graph attention network layer GAT, and obtain the first... The output corresponding to each brain region is denoted as ; S34. Perform feature fusion. The fused features corresponding to each brain region are represented as follows: ； in, This indicates the characteristics after fusion.

9. A depression assessment method based on multi-view subgraph construction according to claim 8, characterized in that, Step S4 specifically includes: S41. Construct positive and negative sample pairs: Calculate the similarity between the fusion features of two brain regions of the same subject based on cosine similarity. Take the fusion features of two brain regions with a similarity greater than 0.7 as positive sample pairs and the fusion features of two brain regions with a similarity less than 0.3 as negative sample pairs. S42. Add a projection head to the fusion features of each brain region to obtain the mapping vector; S43. Calculate the contrast loss for each positive sample pair based on the mapping vector; S44. Calculate the average loss of positive sample pairs based on the contrastive loss of each positive sample pair; S45. Construct a set of key brain regions based on the brain region pairs corresponding to the positive sample pairs; S46. Construct a node-level weakly supervised regression loss based on the key brain region set. Through training, compare the regression performance of each electrode node in the key brain region set to identify significant electrodes. S47. Aggregate the regression outputs of significant electrodes in all key brain regions and input them into a fully connected neural network to obtain the predicted depression score.

Images