Cognitive state classification and key feature recognition method based on community graph convolutional neural network

Abstract

The invention discloses a cognitive state classification and key feature recognition method based on a community graph convolutional neural network, and belongs to the technical field of machine learning technologies. The method comprises the following steps: constructing a multi-modal neural network, inputting preprocessed data into the multi-modal neural network, and obtaining node information of all subjects; carrying out community detection on the multi-modal neural networks of all subjects by adopting a community detection method; and constructing a community graph convolutional neural network Com-GCN, and completing a cognitive state classification task and key feature recognition. According to the method, the multi-modal neural network is constructed and community detection is carried out, so that division of each community is preliminarily obtained, and the relation among the data is analyzed from the perspective of communities; in combination with an affinity aggregation model in the community graph convolutional neural network, the cognitive state classification task performance is enhanced, and key feature factors causing cognitive decline are accurately identified; classification is carried out through a full connection layer, so that the classification performance is further improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning, and particularly relates to a cognitive state classification and key feature identification method based on a community graph convolutional neural network. BACKGROUND

[0002] The pre-cognitive decline stage of the brain is between normal aging and significant cognitive decline, and individuals show mild cognitive decline, but the ability of daily life is not significantly affected, and it is easy to be mistaken for natural aging. The classification of cognitive state and the identification of key features help to identify the early changes in brain function, which is of great significance for improving the cognitive decline of individuals and delaying the process of brain dysfunction.

[0003] Currently, clinicians usually use individual clinical data and medical imaging to detect early symptoms of cognitive decline. In the clinical data, genetic data, cerebrospinal fluid biomarkers and cognitive function assessment scales are all ways to detect cognitive decline. Through genetic monitoring data, doctors can determine whether individuals are at higher risk of cognitive decline. The concentrations of beta-amyloid protein, tau protein and total tau protein in the cerebrospinal fluid of individuals with cognitive decline often change abnormally, and these cerebrospinal fluid biomarkers provide a basis for early classification of cognitive state. Cognitive function assessment scales (such as MMSE, MoCA, ADAS-Cog, Global CDR) test the individual's memory, language, attention, spatial cognition and other abilities, which can help clinicians determine the degree of cognitive impairment of individuals. In imaging technology, magnetic resonance imaging (MRI) technology can provide accurate brain structure images to help detect related changes in brain dysfunction. MRI includes structural MRI (sMRI) and functional MRI (fMRI). Among them, fMRI is used to observe the activity of the brain when performing different tasks. In the early stage of cognitive decline, when the individual's cognitive function declines slightly, the functional connectivity of the brain has already begun to change, and fMRI can detect this functional change at this stage.

[0004] However, these traditional cognitive state classification methods have certain limitations. First, the traditional method relies too much on the experience and subjective judgment of doctors, for example, in the process of interpreting medical images, only relying on the doctor to make personal judgments based on the images. Second, the traditional method is difficult to effectively integrate and analyze the complex relationships that may exist between these data, thereby affecting the final diagnosis result.

[0005] Compared with traditional diagnostic methods, modern technology can more efficiently find diseases. With the progress of computer technology and medical information processing technology, deep learning technology has been widely used in the analysis of biomedical imaging data and non-imaging data in the past few years. Deep learning is good at processing and integrating multiple types of data, especially in the combination of different modalities of data such as images, sequences and table data, which makes cognitive state classification more effective. However, existing researches mainly focus on single use of brain imaging or gene data and do not consider the complex relationship between multiple data.

[0006] Therefore, how to solve the problem of cognitive state classification and recognition in the case of complex relationship between multiple data is a technical problem to be solved by the present application. SUMMARY

[0007] The present application aims to provide a cognitive state classification and key feature recognition method based on community graph convolutional neural network to solve the problems in the background art.

[0008] The present application is achieved in that a cognitive state classification and key feature recognition method based on community graph convolutional neural network is characterized in that the method comprises the following steps:

[0009] Step S1: Preprocessing the collected data;

[0010] Step S2: Constructing a multi-modal neural network, inputting the preprocessed data into the multi-modal neural network, and obtaining node information of all subjects;

[0011] Step S3: Using a community detection method to detect the community of the multi-modal neural network of all subjects;

[0012] Step S4: Constructing a community graph convolutional neural network Com-GCN to complete the cognitive state classification task and key feature recognition.

[0013] Preferably, the multi-modal neural network is a multi-modal weighted graph network structure, specifically:

[0014] Let , , ; , ;

[0015] wherein, represents a node of the multi-modal neural network, represents an edge of the multi-modal neural network; represents feature information in the node, represents the weight of the edge;

[0016] The node information is divided into four categories of data, and four types of scales are made, each node representing the test score of each type of cognitive assessment scale of the subject; the edges in the network represent the association between two nodes, and the weight of the edge reflects the strength of the correlation between the nodes;

[0017] The edge weight is obtained by calculating the Pearson correlation coefficient between the nodes, and the specific formula is as follows:

[0018] ;

[0019] Wherein, represents the length of the standardized feature sequence, and represent the first feature sequence of node and node , respectively.

[0020] Preferably, the node information of the multi-modal neural network is detected by the community detection method in step S3, specifically:

[0021] Step S3-1: define the centrality of the node , find the high node centrality; the high node centrality indicates that the node is important and closely connected with other nodes, and the node centrality calculation formula is as follows:

[0022] ;

[0023] Wherein, is the total number of nodes, represents the degree of node , and the value of is between 1 and the total number of nodes ;

[0024] Step S3-2: sort the nodes according to the centrality of the nodes in the multi-modal neural network, and the node with the largest centrality is the central node of the first community;

[0025] Step S3-3: introduce the modularity index to measure the quality of community detection;

[0026] Step S3-4: introduce the local modularity index to evaluate the community detection;

[0027] Step S3-5: find the node closest to the community in the remaining nodes , calculate the local modularity increment brought by the merging of node ;

[0028] Step S3-6: Repeat step 3-5 until adding any node will cause At this time, the community detection is completed.

[0029] Step S3-7: In the remaining nodes, select the node with the largest centrality as the center node of the next community, repeat steps 3-4 to 3-6 until all nodes in the multi-modal neural network are divided into communities, and obtain the node information of each subject after community detection.

[0030] Preferably, the modularity index introduced in step S3-3 is used to measure the quality of community detection, specifically:

[0031] Modularity index is defined as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] wherein, represents the number of edges in the network; represents the adjacency matrix of the network; and represent the communities in which nodes and nodes are located, respectively; When Q = 1, a complete community structure is obtained, and there is no connection between communities. When Q = 0, a network without obvious community structure is obtained, and the connections between nodes are uniform.

[0036] The local modularity index is defined as follows:

[0037] ;

[0038] wherein, and represent the number of edges within and outside the community, respectively.

[0039] Preferably, the community graph convolutional neural network Com-GCN includes an input layer, an affinity aggregation model, a fully connected layer, and an output layer, and the affinity aggregation model includes an intra-community information aggregation module, an inter-community information aggregation module, and a dynamic community graph convolutional layer.

[0040] The community-based information aggregation module aggregates feature information between nodes, the inter-community information aggregation module performs cross-community information interaction, the dynamic community graph convolutional layer updates community information through matrix operations, and the feature information output after multiple iterations is input into the fully connected layer for final classification.

[0041] The Community Graph Convolutional Neural Network (Com-GCN) includes an input layer, an affinity aggregation model, a fully connected layer, and an output layer. The affinity aggregation model includes an intra-community information aggregation module, an inter-community information aggregation module, and a dynamic community graph convolutional layer.

[0042] The affinity aggregation model is as follows: community information is represented as , Indicates the first The second affinity aggregation The situation of each community, through calculation of the community The global efficiency is obtained, and the calculation formula is as follows:

[0043] ;

[0044] in, and Indicates community The total number of all nodes and nodes within the community To the node Shortest path distance;

[0045] Node information is represented as , Indicates the process The matrix after secondary affinity aggregation, the initial feature information matrix is: ;

[0046] Community affinity is represented as , Indicates community and community Affinity between individuals, affinity between communities The calculation formula is as follows:

[0047] ;

[0048] in, For the community , For the community Nodes in two highly matched communities will have relatively high similarity. The average similarity between nodes in two communities is calculated as follows:

[0049] ;

[0050] in, and respectively represent the total number of nodes in the community and the community ;

[0051] The affinity between nodes is represented as , represents the affinity between node and node , which is calculated by the common neighbors between the two nodes, and the calculation formula is as follows:

[0052] ;

[0053] wherein, represents the neighbor node of node , represents the set of common neighbor nodes of node and node , represents the degree of node ;

[0054] The full connection layer includes input, hidden layer and output, the hidden layer uses as the activation function, and the non-linear transformation gradually abstracts the high-order features; the output uses as the activation function to calculate the classification result;

[0055] The last layer of the full connection layer uses as the activation function, and the remaining layers use .

[0056] Preferably, the community graph convolutional neural network Com-GCN is constructed in step S4 to complete the cognitive state classification task and key feature identification, specifically:

[0057] Step S4-1: Use the inter-community information aggregation module for inter-community convolution:

[0058] After the inter-community convolution after the first affinity aggregation, the order BGCAS-CN is obtained, and the community information change matrix is defined as follows:

[0059] ;

[0060] wherein, represents a matrix with all values being 1, is the Hadamard product, is the unit matrix; is a diagonal matrix, wherein the diagonal elements collect the information received by each community from its neighbor communities; represents the community the community information change amount of the community;

[0061] The inter-community affinity change is calculated as follows:

[0062]

[0063]

[0064] wherein, is the inter-community affinity matrix, is the inter-community affinity diagonal matrix, the element in the matrix is the sum of the affinities of all the inter-community edges connected to the community , and represents the ratio of the affinities of all the edges connected to the community to the sum of the affinities of all the edges of the community; the matrix quantitatively describes the influence of the community information change on the edge affinity; the element in the matrix corresponds to the element in the matrix , which represents the degree of influence of the community information change on the edge affinity;

[0065] Step S4-2: intra-community convolution is performed by using an intra-community information aggregation module;

[0066] Step S4-3: the community graph convolutional neural network Com-GCN is trained;

[0067] Step S4-4: classification is performed by using a fully connected layer;

[0068] Step S4-5: feature extraction is performed by using the community graph convolutional neural network Com-GCN;

[0069] Step S4-6: the nodes are arranged in descending order of importance, and a subset of nodes is obtained.

[0070] Preferably, in the step S4-2, the intra-community convolution is performed by using an intra-community information aggregation module, specifically as follows:

[0071] After the intra-community convolution of the first affinity aggregation, each node updates its own feature information; the new node feature information matrix is composed of the node feature information matrices of the communities, wherein the th column of the matrix is composed of the node feature information matrices of the communities in the th column of the matrix .​​​​Node feature information matrix of each community The calculation is as follows:

[0072] ;

[0073] in, and Represent the weight matrix and the first Affinity matrix between community nodes; Indicates having A learnable filter with learnable parameters. Indicates the first The number of nodes in each community The length of the feature sequence, The first term used for normalization The node degree matrix of each community.

[0074] Preferably, the training of the community graph convolutional neural network Com-GCN in step S4-3 specifically involves:

[0075] The final weight matrix of the nodes is calculated as follows:

[0076] .

[0077] Preferably, in step S4-4, a fully connected layer is used for classification, specifically as follows:

[0078] exist A stable community graph convolutional neural network is obtained after the community graph convolution, which also generates the weight matrix of the corresponding community network. , will the node Fraction and participation coefficient By combining them, we obtain a one-dimensional feature vector. The input is fed into the fully connected layer;

[0079] For fully connected layers, the last layer uses... As the activation function, the remaining layers use The formula is expressed as follows:

[0080] ;

[0081] in, and These represent the first and second fully connected layers, respectively. Layer and first Weight matrices and biases between layers; and They represent the first Layer and first A one-dimensional vector of the layer; This indicates the number of layers in a fully connected layer.

[0082] Preferably, in steps S4-5, a community graph convolutional neural network (Com-GCN) is used for feature extraction, specifically as follows:

[0083] Accuracy score pass The layer's output is calculated as follows:

[0084] ;

[0085] in, and These are the predicted label and the true label of the sample, respectively.

[0086] The importance matrix corresponding to the parameter matrix is ​​calculated as follows:

[0087] ;

[0088] The obtained importance matrix and accuracy score Combined, the neuron importance score matrix is ​​obtained, calculated as follows:

[0089] ;

[0090] in, For the first An importance matrix, It is the first The number of columns in a learnable parameter matrix. ;, For the sample size, This represents the average accuracy score. Each element in the table represents an importance score for the community structure index;

[0091] The weight score of a node is derived from the weight score of its neuron. Weight score of each node The calculation is as follows:

[0092] ;

[0093] in, and They represent Fraction and participation coefficient Importance score.

[0094] Compared with the prior art, the present invention has the following improvements and advantages:

[0095] 1. By constructing a multi-modal neural network and performing community detection, the division of each community is obtained preliminarily, and the relationship between the data is analyzed from the community perspective; in combination with the affinity aggregation model in the community graph convolutional neural network, the performance of the cognitive state classification task is enhanced, and the key feature factors leading to cognitive decline are accurately identified.

[0096] 2. The information transmission within and between communities is described by the affinity aggregation model, the key features that contribute to classification are extracted, and the classification is performed through the full connection layer in combination with the full connection layer, thereby further improving the classification performance. BRIEF DESCRIPTION OF DRAWINGS

[0097] Figure 1 The application flowchart is shown.

[0098] Figure 2 The community detection flowchart of the application is shown.

[0099] Figure 3 The flowchart of the community graph convolutional neural network is shown. DETAILED DESCRIPTION

[0100] The application will be further described below in combination with the drawings.

[0101] As shown in the drawings, Figure 1 a cognitive state classification and key feature identification method based on a community graph convolutional neural network, the method comprising the following steps:

[0102] Step S1: pre-processing the collected data, specifically:

[0103] Step S1-1: using the resting state fMRI data processing assistant to perform slice time correction, head motion realignment, normalization, residual noise removal, and time filtering processing on the fMRI data;

[0104] Step S1-2: using the automatic anatomical labeling AAL template to divide the brain into 116 functional regions, and extracting the initial time series data of each brain region;

[0105] Step S1-3: inputting the initial time series data into the BrainLM model to generate time series data;

[0106] Step S1-4: using the PLINK tool to screen and process the genetic data, selecting 45 genes containing important single nucleotide polymorphism SNP sites related to cognitive function degradation for encoding; these SNP data are encoded in-1, 0, 1 encoding mode, TT encoding is 1, TC encoding is 0, and CC encoding is-1.

[0107] ​Steps S1-5: The preprocessed and encoded single nucleotide polymorphism data are further processed using the Nucleotide Transformer model to extract gene feature sequences;

[0108] Steps S1-6: For cerebrospinal fluid biomarker concentration samples and cognitive assessment scale data samples, remove data samples containing outliers and delete samples containing missing values.

[0109] Step S2: Construct a multimodal neural network. Input the preprocessed data into the multimodal neural network to obtain the node information of all subjects. Specifically:

[0110] Multimodal neural networks are multimodal weighted graph network structures, specifically:

[0111] make , , ; , ;

[0112] in, Representing nodes in a multimodal neural network, Represents the edges of a multimodal neural network; Represents the feature information in the node. Indicates the weight of the edge;

[0113] The node information is divided into four categories: brain region nodes, gene nodes, cerebrospinal fluid biomarker concentration nodes, and cognitive assessment scale nodes. Four types of scales are created, and each node represents the test score of the subject on each type of cognitive assessment scale. The edges in the network represent the relationship between two nodes, and the weight of the edge reflects the strength of the correlation between the nodes.

[0114] The edge weights are obtained by calculating the Pearson correlation coefficient between nodes, as shown in the following formula:

[0115] ;

[0116] in, This represents the length of the feature sequence after data standardization. and Representing nodes respectively and nodes The A number of characteristic sequences.

[0117] like Figure 2 As shown, in step S3, a community detection method is used to perform community detection on the node information of the multimodal neural network, specifically as follows:

[0118] Step S3-1: Define the centrality of nodes , find high node centrality; high node centrality indicates that the node is important and closely connected to other nodes, and the node centrality calculation formula is as follows:

[0119] ;

[0120] wherein, is the total number of nodes, represents the degree of node , the value of is between 1 and the total number of nodes ;

[0121] Step S3-2: Sort the nodes according to their centrality in the multi-modal neural network, and the node with the highest centrality is the center node of the first community;

[0122] Step S3-3: Introduce the modularity index to measure the quality of community detection;

[0123] The modularity index is defined as follows:

[0124] ;

[0125] ;

[0126] ;

[0127] wherein, represents the number of edges in the network; represents the adjacency matrix of the network; and represent the communities in which nodes and nodes are located, respectively; When , a complete community structure is obtained, and there is no connection between communities;

[0128] Step S3-4: Introduce the local modularity index to evaluate community detection;

[0129] The local modularity index is defined as follows:

[0130] ;

[0131] wherein, and represent the number of edges within and outside the community, respectively.

[0132] Step S3-5: Find the node closest to the community among the remaining nodes. compute nodes The local modularity increase brought about by merging into the community ;

[0133] Step S3-6: Repeat steps 3-5 until adding any node results in... At this point, community testing is complete;

[0134] Step S3-7: Among the remaining nodes, select the node with the highest centrality as the center node of the next community. Repeat steps 3-4 to 3-6 until all nodes in the multimodal neural network are divided into communities, and obtain the node information of each subject after community detection.

[0135] In step S4, a community graph convolutional neural network Com-GCN is constructed to complete the cognitive state classification task and key feature recognition, specifically as follows:

[0136] The Community Graph Convolutional Neural Network (Com-GCN) consists of an input layer, an affinity aggregation model, a fully connected layer, and an output layer. The affinity aggregation model includes an intra-community information aggregation module, an inter-community information aggregation module, and a dynamic community graph convolutional layer.

[0137] Community information is represented as , Indicates the first The second affinity aggregation The situation of each community, through calculation of the community The global efficiency is obtained, and the calculation formula is as follows:

[0138] ;

[0139] in, and Indicates community The total number of all nodes and nodes within the community To the node Shortest path distance;

[0140] Node information is represented as , Indicates the process The matrix after secondary affinity aggregation, the initial feature information matrix is: ;

[0141] Community affinity is represented as , Indicates community and community Affinity between individuals, affinity between communities The calculation formula is as follows:

[0142] ;

[0143] wherein, is the community , is the community ; the nodes in two communities with high matching degree will have relatively high similarity, and the average similarity between nodes in two communities is calculated as follows:

[0144] ;

[0145] wherein, and represent the total number of nodes in the community and the community ;

[0146] The affinity between nodes is represented as , represents the affinity between nodes and nodes , which is calculated by the common neighbors between the two nodes, and the calculation formula is as follows:

[0147] ;

[0148] wherein, represents the neighbor nodes of node , represents the set of common neighbor nodes of node and node , represents the degree of node ;

[0149] Each community gathers information of other communities under the guidance of inter-community affinity, and the information change calculation formula is as follows:

[0150] ;

[0151] The change of community information will change the affinity between communities, so as to update the community structure, and the change of inter-community affinity is calculated as follows:

[0152] ;

[0153] Each node aggregates the feature information of other nodes in the same community, and calculates the information of each node after the aggregation of the feature information of nodes in the community in the th time, and the calculation formula is as follows:

[0154] ;

[0155] ;

[0156] wherein, represents the community structure obtained in the th iteration;

[0157] After affinity aggregations, the intra-community affinity keeps increasing, while the inter-community affinity gradually decreases, forming a stable community structure. Based on the final affinity aggregation model, two indicators of community structure, i.e., the intra-community score and the participation coefficient , are calculated, which are used to measure the connectivity of a node within the community and the connectivity of other communities in the network, respectively, and the calculation methods are as follows:

[0158] ;

[0159] ;

[0160] Through the analysis of community structure indicators, the following two formulas provide the classification of normal control group and cognitive state degeneration subjects and the identification of key features.

[0161] ;

[0162] ;

[0163] wherein, represents the final classification result, represents the selected feature set, which contains the most useful features for the classification task.

[0164] The fully connected layer includes input, hidden layer and output, the hidden layer uses as the activation function, and the high-order features are gradually abstracted through the nonlinear transformation; the output uses as the activation function to calculate the classification result; the last layer of the fully connected layer uses as the activation function, and the other layers use .

[0165] As shown in Figure 3 , the cognitive state classification task and the key feature identification are completed, specifically:

[0166] Step S4-1: inter-community information aggregation module is used for inter-community convolution:

[0167] After the inter-community convolution after the th affinity aggregation, the order BGCAS-CN is obtained, and the community information change matrix is defined as follows:

[0168] ;

[0169] in, This represents a matrix where all values ​​are 1. For Hadama accumulation, It is the identity matrix; It is a diagonal matrix where the diagonal elements aggregate the information that each community receives from its neighboring communities; Indicates community The amount of change in community information;

[0170] The changes in inter-community affinity are calculated as follows:

[0171] ;

[0172] ;

[0173] in, A community affinity matrix, This is a diagonal matrix representing the affinity between communities. The elements in the matrix are connected to the community. The sum of affinity between all communities. This represents the ratio of the affinity of all edges connected to a community to the sum of the affinities of all edges in that community; matrix. middle Quantitatively described the community Information changes on the edge The influence of affinity; matrix elements in Corresponding matrix elements in This element represents the community. Community information changes on the border The degree of influence of affinity.

[0174] In step S4-2, the community-based information aggregation module is used to perform intra-community convolution, specifically as follows:

[0175] In the After the community convolution of second affinity aggregation, each node updates its own feature information; the new node feature information matrix Depend on The community's node feature information matrix is ​​composed of these elements, where the first element is... Node feature information matrix of each community The calculation is as follows:

[0176] ;

[0177] in, and respectively represent a weight matrix and an affinity matrix between nodes in the

[0178] In step S4-4, a full connection layer is used for classification, specifically:

[0179] After the

[0180]

[0181] For the full connection layer, the last layer uses as the activation function, and the remaining layers use , which is expressed as follows:

[0182]

[0183] where and represent the weight matrix and bias between the

[0184] In step S4-5, the community graph convolutional neural network Com-GCN is used for feature extraction, specifically:

[0185] The accuracy score is calculated through the output of the layer, and the calculation method is as follows:

[0186]

[0187] ​​​​​​​​​​​​​​​​​​​​​​​​in, and These are the predicted label and the true label of the sample, respectively.

[0188] The importance matrix corresponding to the parameter matrix is ​​calculated as follows:

[0189] ;

[0190] The obtained importance matrix and accuracy score Combined, the neuron importance score matrix is ​​obtained, calculated as follows:

[0191] ;

[0192] in, For the first An importance matrix, It is the first The number of columns in a learnable parameter matrix. ;, For the sample size, This represents the average accuracy score. Each element in the table represents an importance score for the community structure index;

[0193] The weight score of a node is derived from the weight score of its neuron. Weight score of each node The calculation is as follows:

[0194] ;

[0195] in, and They represent Fraction and participation coefficient Importance score.

[0196] In steps S4-6, the nodes are arranged in descending order of importance to obtain... A subset of nodes;

[0197] The subset is input into Com-GCN, the classification accuracy is recorded, the accuracy is compared, and the best feature subset is selected, namely abnormal brain regions, key genes, abnormal cerebrospinal fluid biomarker concentrations, and cognitive state abnormality scale scores.

[0198] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A community graph convolutional neural network-based cognitive state classification and key feature identification method, characterized in that: The method comprises the following steps: Step S1: preprocessing the collected data; Step S2: constructing a multi-modal neural network, inputting the preprocessed data into the multi-modal neural network, and obtaining node information of all subjects; Step S3: using a community detection method to detect the node information of the multi-modal neural network; Step S4: constructing a community graph convolutional neural network Com-GCN to complete the cognitive state classification task and key feature recognition.

2. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 1, characterized in that: The multi-modal neural network is a multi-modal weighted graph network structure, specifically: Let , , ; , ; wherein, represents a node of the multi-modal neural network, represents an edge of the multi-modal neural network; represents feature information in the node, represents a weight of the edge; The node information is divided into four types of data, and four types of scales are made, and each node represents the test score of each type of cognitive evaluation scale of the subject; the edge in the network represents the association between two nodes, and the weight of the edge reflects the strength of the correlation between the nodes; The edge weight is obtained by calculating the Pearson correlation coefficient between the nodes, and the specific formula is as follows: ； wherein, represents the length of the feature sequence after data standardization, and respectively represent the first and the second feature sequence of the node and the node .

3. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 1, characterized in that: In step S3, the node information of the multi-modal neural network is detected by using a community detection method, specifically: Step S3-1: defining the centrality of the node , find out the high node centrality; the high node centrality indicates that the node is important and closely connected with other nodes, and the formula for calculating the node centrality is as follows: ； wherein, is the total number of nodes, represents a node of degree, whose value is between 1 and the total number of nodes ; Step S3-2: Sort the nodes according to their centrality in the multi-modal neural network, and the node with the maximum centrality is the center node of the first community; Step S3-3: Introducing a modularity index for measuring the quality of community detection; Step S3-4: Introducing local modularity index , community detection evaluation; Step S3-5: finding the node closest to the community in the remaining nodes , computing nodes local modularity increment brought by merging into the community ; Step S3-6: Repeat step 3-5 until adding any node would result in At this time, the community detection is completed; Step S3-7: In the remaining nodes, select the node with the maximum centrality as the center node of the next community, and repeat steps 3-4 to 3-6 until all nodes in the multi-modal neural network are divided into communities, and the node information of each subject after community detection is obtained.

4. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 3, characterized in that: The step S3-3 introduces the modularity index for measuring the quality of community detection, in particular: Modularity index is defined as follows: ； ； ； wherein, denotes the number of edges in the network; denotes the adjacency matrix of the network; and denote the communities in which the nodes and are located, respectively; when is obtained a complete community structure, with no connections between communities; when is obtained a network without a clear community structure, with uniform connections between nodes; The local module index is defined as follows: ； wherein, and denote the number of edges within and outside the community, respectively.

5. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 1, characterized in that: The community graph convolutional neural network Com-GCN includes an input layer, an affinity aggregation model, a fully connected layer and an output layer, and the affinity aggregation model includes an intra-community information aggregation module, an inter-community information aggregation module and a dynamic community graph convolutional layer; The intra-community information aggregation module aggregates feature information between nodes, the inter-community information aggregation module interacts with cross-community information, and the dynamic community graph convolutional layer updates community information through matrix operation, and the output feature information after multiple iterations is input into the fully connected layer for final classification; The community graph convolutional neural network Com-GCN includes an input layer, an affinity aggregation model, a fully connected layer and an output layer, and the affinity aggregation model includes an intra-community information aggregation module, an inter-community information aggregation module and a dynamic community graph convolutional layer; The affinity aggregation model is as follows: community information is represented as , Indicates the first The second affinity aggregation The situation of each community, through calculation of the community The global efficiency is obtained, and the calculation formula is as follows: ； wherein, and denotes the number of all nodes in the community and the nodes within the community to the node shortest path distance; The node information is represented as , The matrix after the affinity aggregation is represented as The initial feature information matrix is ; Community affinity is expressed as , represents the affinity between communities and community Community affinity is calculated as follows: ； wherein, for the community , for the community ; nodes in two communities with high matching degree will have relatively high similarity, and the average similarity between nodes in two communities is calculated as follows: ； wherein, and respectively represent the total number of nodes in the community and the community respectively. The inter-node affinity is denoted as , denotes the inter-node affinity between nodes and nodes , which is computed by the common neighbors between the two nodes, and the formula is as follows: ； wherein, denotes a neighbor node of a node , denotes a set of common neighbor nodes of a node and a node , denotes a degree of a node ; The fully connected layer includes an input, a hidden layer, and an output, the hidden layer uses As an activation function, the hyperbolic tangent function gradually abstracts high-order features; the output uses As an activation function, the sigmoid function calculates a classification result; The last layer of the fully connected layer uses The remaining layers use as activation function .

6. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 5, characterized in that: In step S4, the community graph convolutional neural network Com-GCN is constructed to complete the cognitive state classification task and key feature recognition, specifically: Step S4-1: inter-community convolution is performed using the inter-community information aggregation module: After the community convolution after the second affinity aggregation, the community information change matrix is defined as follows: After the community convolution after the second affinity aggregation, the community information change matrix is defined as follows: The order BGCAS-CN, the community information change matrix is defined as follows: ； wherein, denotes a matrix of all values being 1, is the Hadamard product, is the identity matrix; is a diagonal matrix, wherein the diagonal elements collect the information received by each community from its neighboring communities; denotes the community information change amount; The inter-community affinity change is calculated as follows: ； ； in, A community affinity matrix, This is a diagonal matrix representing the affinity between communities. The elements in the matrix are connected to the community. The sum of affinity between all communities. This represents the ratio of the affinity of all edges connected to a community to the sum of the affinities of all edges in that community; matrix. middle Quantitatively described the community Information changes on the edge The influence of affinity; matrix elements in Corresponding matrix elements in This element represents the community. Community information changes on the border The degree of influence of affinity; Step S4-2: intra-community convolution is performed using the intra-community information aggregation module; Step S4-3: training the community graph convolutional neural network Com-GCN; Step S4-4: classification is performed using the fully connected layer; Step S4-5: feature extraction is performed using the community graph convolutional neural network Com-GCN; Step S4-6: Arrange the nodes in decreasing order of importance, obtaining a subset of the nodes. nodes.

7. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 1, characterized in that: In step S4-2, the intra-community convolution is performed using the intra-community information aggregation module, specifically: In the first After the convolution of the community of the second affinity aggregation, each node updates its own feature information; the new node feature information matrix By The node feature information matrix of the community is gathered together to form The node feature information matrix of the first community The calculation is as follows: ； wherein, and denote a weight matrix and an affinity matrix between nodes of the community, respectively; denotes a learning filter with learnable parameters, denotes the number of nodes of the community, is the length of the feature sequence, denotes a node degree matrix of the community used for normalization.

8. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 7, characterized in that: In step S4-3, the community graph convolutional neural network Com-GCN is trained, specifically: The calculation of the maximum weight matrix of the node is as follows: 。 9. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 7, characterized in that: The step S4-4 uses a full connection layer to classify, specifically as follows: In A stable community graph convolutional neural network is obtained after community graph convolution, and a weight matrix of the corresponding community network is also generated The node's Score And the participation coefficient Are combined to obtain a one-dimensional feature vector Input into the fully connected layer; For the fully connected layers, the last layer uses as the activation function, and the rest use , expressed in the formula as follows: ； wherein, and denote the weight matrix and bias between the layer and the layer in the fully connected layer, respectively; and denote one-dimensional vectors of the layer and the layer, respectively; denotes the number of layers in the fully connected layer.

10. The cognitive state classification and key feature identification method based on community graph convolutional neural network according to claim 7, characterized in that: The step S4-5 uses a community graph convolutional neural network Com-GCN to extract features, specifically as follows: Accuracy score By The output of the layer is computed as follows: ； wherein, and are the predicted and true labels of the sample, respectively; An importance matrix corresponding to the parameter matrix is calculated, and the calculation method is as follows: ； The resulting importance matrix is combined with the accuracy score In combination, the resulting matrix of neuron importance scores is calculated as follows: ； wherein, is the importance matrix of the th community structure index, is the number of columns of the th learnable parameter matrix, ;, is the number of samples, denotes the average accuracy score, each element of which represents an importance score of a community structure index; The weight score of a node is derived from the weight score of the neuron, the weight score of the node is calculated as follows: ； wherein, and respectively represent a score of importance of and a participation coefficient of the user.

Images