Schizophrenia classification method of fuzzy hypergraph neural network based on evidence theory

Abstract

The invention provides a schizophrenia classification method based on a fuzzy hypergraph neural network of an evidence theory, belongs to the technical field of hypergraph neural networks and evidence theories, and solves the technical problem of high heterogeneity in a schizophrenia identification task. According to the technical scheme, firstly, an excellent hyperedge granularity model is constructed through a sparse constraint function, and nodes with high heterogeneity are excluded from the hyperedge granularity model; then, in the construction process of the fuzzy hypergraph, fusing the correlation quality function and the distance quality function by applying an evidence theory; the node weight is described through the fuzzy membership degree, and the weight of the node with low heterogeneity is improved; and finally, through a fuzzy hypergraph convolution model, identifying labels of schizophrenia patients, improving classification precision of schizophrenia data and optimizing semantic interpretation. The schizophrenia diagnosis method has the beneficial effect that the schizophrenia diagnosis accuracy and interpretability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hypergraph neural networks and evidence theory, and in particular to a schizophrenia classification method based on a fuzzy hypergraph neural network based on evidence theory. Background Art

[0002] Schizophrenia is a serious mental disorder that causes hallucinations, delusions, and some extremely chaotic thinking and behavior, affecting the patient's daily life and possibly causing disability. The paper "Cognitive impairment in schizophrenia: aetiology, pathophysiology, and treatment" (Molecular Psychiatry, 2023) proposed that there are about 24 million cases of schizophrenia worldwide, and the paper "The current status of alexithymia in elderly patients with schizophrenia and its impact on mental symptoms" (Henan Medical Research, 2024) proposed that the pathogenesis of schizophrenia is a complex process with multiple factors and steps, which mainly depends on the doctor's experience diagnosis, with high uncertainty and poor treatment effect. At the same time, due to the lack of experienced doctors, professional equipment and other medical resources, the prediction accuracy of schizophrenia is low.

[0003] In order to improve the diagnosis of schizophrenia, the paper "Brain network analysis of schizophrenia patients based on hypergraph signal processing" (IEEE Transactions on Image Processing, 2023) uses hypergraph brain networks to explore the pathogenesis of schizophrenia, construct a real weighted brain hypergraph network, and convert the matrix into a weighted adjacency tensor to improve the classification accuracy of schizophrenia. However, since the multi-source heterogeneous schizophrenia data comes from different hospitals and institutions and has high heterogeneity, the model cannot model the high-order correlation between multimodal data, which hinders the subsequent diagnosis and classification of medical personnel.

[0004] The paper "Multi-Modal imaging genetics data fusion via a hypergraph-based manifold regularization: application to schizophrenia study" (IEEE Transactions on Medical Imaging, 2022) uses a hypergraph to fuse multimodal schizophrenia data. By fusing heterogeneous structural information between different modalities, it characterizes the high-order relationship between schizophrenia data and improves the prediction accuracy of schizophrenia. However, due to the irregular distribution and unstable quality of schizophrenia data, there is uncertainty in the hypergraph modeling in this model and an increase in the misdiagnosis rate. With the increasing maturity of deep learning technology, massive amounts of schizophrenia data have been processed and saved, providing technical guarantee and data support for the development and research of schizophrenia data. Summary of the invention

[0005] In order to make up for the deficiencies of the above-mentioned prior art, the present invention provides a schizophrenia classification method based on fuzzy hypergraph neural network of evidence theory, which reduces the heterogeneity of schizophrenia data, handles the uncertainty in the diagnosis process of schizophrenia patients, and improves the prediction accuracy of schizophrenia.

[0006] The inventive concept of the present invention is as follows: the method of the present invention constructs an excellent hyperedge granularity model through a sparse constraint function to process unevenly distributed schizophrenia data, excludes nodes with higher heterogeneity from the hyperedge granularity model, and reduces the heterogeneity of the data; then, in the process of constructing a fuzzy hypergraph, the evidence theory is applied to fuse the association quality function and the distance quality function; the node weight is characterized by the fuzzy membership, the weight of the node with lower heterogeneity is increased, and the uncertainty of the classification of the node with higher heterogeneity is characterized; finally, the labels of schizophrenia patients are identified through the fuzzy hypergraph convolution model, the classification accuracy of schizophrenia data is improved, and the semantic interpretation is optimized.

[0007] The technical solution adopted by the present invention is: a schizophrenia classification method based on fuzzy hypergraph neural network of evidence theory, comprising the following steps:

[0008] Step 1: Data of schizophrenia patients from 5 different sites were selected as the data set; the DPARSF tool was used to process the data, where the preprocessing methods mainly included inter-layer time correction, spatial normalization, global average intensity normalization, interference signal regression and bandpass filtering; the automatic anatomical labeling atlas template widely used in research was used to extract the time series of the region of interest (ROI), which divided the cerebral cortex into 90 functional brain areas with adjacent positions and similar functions, and the time series information of different functional brain areas was obtained;

[0009] Step 2, complete the hyperedge modeling task through the sparse constraint function: construct the association matrix of schizophrenia data nodes based on the sparse constraint function proposed in the present invention; characterize the optimal k value of different schizophrenia data nodes through the association matrix and the correlation threshold; construct a hyperedge granularity model based on the k nearest neighbor samples to characterize the high-order correlation between schizophrenia data;

[0010] Step 3, construct a fuzzy hypergraph through evidence theory: construct an association quality function based on the schizophrenia data node association matrix obtained in step 2; calculate the distance quality function through the distance between the schizophrenia data nodes; use the quality function to characterize the association coefficient of the schizophrenia data node to the hyperedge; apply the evidence theory, integrate the quality function, and calculate the fuzzy membership of the schizophrenia data node to the hyperedge; obtain the fuzzy adjacency matrix of the fuzzy hypergraph, and thus construct the fuzzy hypergraph;

[0011] Step 4: Based on the fuzzy adjacency matrix H obtained in step 3 F And the hypergraph neural network model, constructing a fuzzy hypergraph convolution model: Based on the fuzzy adjacency matrix H F The features of schizophrenia data nodes and hyperedges are aggregated to construct a fuzzy hypergraph convolution operator. Based on the fuzzy hypergraph convolution operator, a two-layer convolution structure is designed for schizophrenia data node label prediction to obtain the labels of schizophrenia subjects.

[0012] As a preferred technical solution of the present invention: the specific process of step 2 is as follows:

[0013] Step 2.1: Establish a hyperedge granularity model through a sparse constraint function, where the schizophrenia data node feature matrix X is expressed as n represents the number of schizophrenia data nodes, d represents the node feature dimension; by minimizing the reconstruction error of schizophrenia data nodes, the correlation matrix R between nodes is obtained; the sparse constraint function is calculated:

[0014]

[0015] Among them, ||·|| Frepresents the Frobenius norm of the matrix, ||·||2 is the l2-norm regularization term, L represents the Laplace matrix, ρ1 and ρ2 are adjustment parameters, Tr(·) represents the calculation function of the matrix trace, and R T With X T is the transpose of the matrices R and X;

[0016] Step 2.2: Based on the sparse constraint function of step 2.1, obtain the correlation matrix R between schizophrenia data nodes; based on the correlation threshold σ, determine the optimal k value for all schizophrenia data nodes; for schizophrenia data node x i , computing node x i The optimal k value k i :

[0017] k i =num(R(i)>σ) (17)

[0018] Where R(i) is the i-th column of the incidence matrix R, and num is the counting function;

[0019] Step 2.3: According to the optimal k value obtained in step 2.2, calculate the k nearest neighbor samples and use the k nearest neighbor samples as the schizophrenia data nodes within the hyperedge granularity; for the i-th hyperedge, calculate the hyperedge granularity HG k (x i ):

[0020]

[0021] Among them, k i For schizophrenia data node x i The optimal k value is d(x m ,x n ) are any two schizophrenia data nodes x m With x n The Euclidean distance function between .

[0022] As a preferred technical solution of the present invention: the specific process of step 3 is as follows:

[0023] Step 3.1: Get the hyperedge through step 2 x i is any node of schizophrenia data; calculate the schizophrenia data node x i With the central node x j The Euclidean distance d(x i ,x j ), get node x i With super edge e j The distance similarity re ij, calculate the schizophrenia data node x i For the distance quality function of hyperedge association:

[0024]

[0025]

[0026] Among them, m is the distance adjustment parameter, k j is the hyperedge e j Number of internal nodes;

[0027] Step 3.2: Based on the association vector R(j) obtained in step 2.2, obtain the hyperedge The correlation coefficient of all schizophrenia data nodes in the hyperedge; according to any schizophrenia data node x in the hyperedge i For the hyperedge e j Calculate the correlation coefficient of schizophrenia data node x i For the associated mass function of a hyperedge:

[0028]

[0029]

[0030] Among them, the hyperedge e j The number of nodes is k j , super edge e j The correlation coefficient between the ith node and the hyperedge is defined as r ij ;

[0031] Step 3.3: Based on the schizophrenia data node x in step 3.1 and step 3.2 i For the hyperedge e j The distance quality function and association quality function of the schizophrenia data node x are calculated. i For the hyperedge e j The fuzzy membership u ij :

[0032]

[0033]

[0034]

[0035]

[0036] Among them, X A and X B For node x containing schizophrenia data i The node set of , ρ is the evidence adjustment parameter;

[0037] Step 3.4: Obtain the schizophrenia data node x by calculating in step 3.3 i For the hyperedge e j Fuzzy membership u ij , which is the corresponding element of the jth column of the fuzzy adjacency matrix of the fuzzy hypergraph; repeat step 3.3 to calculate the fuzzy adjacency matrix H F :

[0038]

[0039] Where n is the number of rows in the node matrix X, and i and j are any number of rows and columns in the matrix.

[0040] As a preferred technical solution of the present invention: the specific process of step 4 is as follows:

[0041] Step 4.1: Based on step 3, we get the fuzzy adjacency matrix H. F , obtain the fuzzy membership of the schizophrenia data node to the hyperedge and the correlation coefficient of the hyperedge to the schizophrenia data node; based on the fuzzy membership and node-edge-node aggregation operation, aggregate the features of the hyperedge and the node to obtain the fuzzy convolution operator:

[0042]

[0043] Among them, X l With X (l+1) is the feature representation of the fuzzy hypergraph at the l and l+1 layers, σ(·) is the activation function, and H F T is the matrix H F The transpose of D v and D e For nodes and hyperedges based on H F The degree matrix, W is the identity matrix, Θ l is the trainable weight parameter matrix of layer l;

[0044] Step 4.2: Based on the fuzzy convolution operator obtained in step 4.1, a two-layer convolution structure is designed for node label prediction of schizophrenia data to obtain the label Z of the schizophrenia subject:

[0045] Z = Softmax(H(ReLU(HXΘ) (0) ))Θ (1) ) (29)

[0046]

[0047] Among them, Θ (0) is the weight matrix from the first layer to the hidden layer, Θ (1)is the weight matrix from the hidden layer to the last layer, Softmax(·) and ReLU(·) are the activation functions.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) Compared with traditional subjective treatment that relies on doctors' experience, the present invention can reduce the misdiagnosis rate caused by doctors' lack of experience and further assist doctors in completing the accurate treatment of schizophrenia.

[0050] (2) Compared with the traditional hypergraph neural network classification method for schizophrenia, the fuzzy hypergraph neural network proposed in the present invention can reduce the uncertainty caused by the characteristics of schizophrenia, such as multiple data types, irregular distribution, unstable quality, and large differences in data structures of different modalities, and provide interpretability. It is widely favored by medical staff and provides a strong model support for the prediction and identification of schizophrenia.

[0051] (3) Compared with the traditional schizophrenia classification method, the schizophrenia classification method based on the fuzzy hypergraph neural network of evidence theory proposed in the present invention uses the fuzzy hypergraph to reduce the heterogeneity caused by multi-source heterogeneous data, so that medical personnel can obtain more objective and accurate medical data and decision-making solutions. This is the important theoretical value and practical application of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0053] Figure 1 It is a schematic diagram of the overall framework of the schizophrenia classification method based on the fuzzy hypergraph neural network of the evidence theory of the present invention.

[0054] Figure 2 It is a schematic diagram of the specific structure of the schizophrenia classification method based on the fuzzy hypergraph neural network of the evidence theory of the present invention.

[0055] Figure 3 Schematic diagram of hyperedge modeling of the fuzzy hypergraph neural network based on evidence theory of the present invention.

[0056] Figure 4 It is a schematic diagram of fuzzy hypergraph convolution of the fuzzy hypergraph neural network based on evidence theory of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and experimental results. Of course, the specific experimental results described here are only used to explain the present invention and prove the superiority of the present invention, and are not used to limit the present invention.

[0058] Example 1

[0059] See also Figures 1 to 4 This embodiment provides a schizophrenia classification method based on a fuzzy hypergraph neural network based on evidence theory, comprising the following steps:

[0060] Step 1. This embodiment selects 8 nodes from schizophrenia data sets from five different sites as examples; uses the DPARSF tool to preprocess the data; uses the automatic anatomical labeling atlas template to extract the time series of the region of interest, divides the cerebral cortex into 90 functional brain areas with adjacent positions and similar functions, and obtains the time series information of different functional brain areas. The category labels of the 8 nodes are 0, 0, 0, 0, 0, 1, 1, 1; 0 represents a healthy subject and 1 represents a schizophrenia patient;

[0061] Step 2, this embodiment completes the hyperedge modeling task through the sparse constraint function: based on the sparse constraint function, construct the association matrix of 8 nodes; characterize the optimal k value of 8 nodes through the association matrix and the correlation threshold; construct a hyperedge granularity model based on the k nearest neighbor samples to characterize the high-order correlation between schizophrenia data;

[0062] Step 2.1: This embodiment establishes a hyperedge granularity model through a sparse constraint function, where the node feature matrix X is expressed as The number of nodes n is expressed as 8, and the node feature dimension d is 90. By minimizing the reconstruction error of the nodes, the association matrix R between the 8 nodes is obtained. The association matrix is ​​calculated and the diagonal is removed. The association matrix R is:

[0063]

[0064] Step 2.2, based on the association matrix R between the 8 nodes; set the association threshold σ to 0.1, determine the optimal k value of all nodes; calculate the optimal k value vector K of the 8 nodes through the absolute value of the value in the association matrix and the threshold:

[0065] K=[3 2 3 3 3 4 2 2] (2)

[0066] Step 2.3: According to the optimal k value obtained in step 2.2, calculate the k nearest neighbor samples and use the k nearest neighbor samples as nodes within the hyperedge granularity; for example, for the first hyperedge, it includes nodes x2, x3, and x4; for the eighth hyperedge, it includes nodes x3 and x4;

[0067] Step 3, this embodiment constructs a fuzzy hypergraph through evidence theory: construct an association quality function based on the association matrix of the 8 nodes obtained in step 2; calculate the distance quality function through the distance between the 8 nodes; apply the quality function to characterize the association coefficient of the 8 nodes for the hyperedge; apply the evidence theory, fuse the quality function, and calculate the fuzzy membership of the 8 nodes for the hyperedge; obtain the fuzzy adjacency matrix of the fuzzy hypergraph, thereby constructing the fuzzy hypergraph;

[0068] Step 3.1, this embodiment takes the first hyperedge as an example, through step 2, a hyperedge e1 = {x2, x3, x4} is obtained, and the distance quality functions of the three nodes x2, x3, x4 associated with the hyperedge e1 and the three nodes completely associated with the hyperedge e1 are: 0.332, 0.332, 0.286, 0.05 respectively;

[0069] Step 3.2, based on the association vector R(1) obtained in step 2.2, obtain the association coefficients of all nodes in the hyperedge e1 = {x2, x3, x4} to the hyperedge; the distance quality functions of the three nodes x2, x3, x4 associated with the hyperedge e1 and the three nodes completely associated with the hyperedge e1 are: 0.209, 0.532, 0.209, 0.05 respectively;

[0070] Step 3.3: Based on the distance quality function and association quality function of the three nodes x2, x3, and x4 for the hyperedge e1 in steps 3.1 and 3.2, calculate the fuzzy membership u of the three nodes x2, x3, and x4 for the hyperedge e1. 21 ,u 31 ,u 41 : 0.256, 0.409, 0.251;

[0071] Step 3.4: The fuzzy membership u of the three nodes x2, x3, and x4 to the hyperedge e1 in step 3.3 21 ,u 31 ,u 41 , which is the corresponding element of the first column of the fuzzy adjacency matrix of the fuzzy hypergraph; repeat step 3.3 to calculate the fuzzy adjacency matrix H F :

[0072]

[0073] Step 4: This embodiment is based on the fuzzy adjacency matrix H obtained in step 3 F And the hypergraph neural network model, constructing a fuzzy hypergraph convolution model: Based on the fuzzy adjacency matrix H F The features of nodes and hyperedges are aggregated to construct a fuzzy hypergraph convolution operator. Based on the fuzzy hypergraph convolution operator, a two-layer convolution structure for node label prediction is designed to obtain node labels.

[0074] Step 4.1: This embodiment takes the first and eighth hyperedges as examples and obtains the fuzzy adjacency matrix H based on step 3. F , obtain the fuzzy membership u of the three nodes x2, x3, and x4 to the hyperedge e1 21 ,u 31 ,u 41 and the fuzzy membership u of the two nodes x3 and x4 to the hyperedge e8 38 ,u 48 ; Based on fuzzy membership and node-edge-node aggregation operation; First, according to the nodes x2, x3, x4 and fuzzy membership u 21 ,u 31 ,u 41 Update the hyperedge e1 feature based on nodes x3, x4 and fuzzy membership u 38 ,u 48 Update the hyperedge e8 feature; secondly, according to the fuzzy membership u 31 ,u 38 The features of node x3 are updated with the features of the two hyperedges e1 and e8. The update process of node x4 is similar to that of x3, completing the node-edge-node aggregation operation;

[0075] Step 4.2. Based on the fuzzy convolution operator obtained in step 4.1, a two-layer convolution structure is designed for node label prediction. The label prediction is completed through the two-layer convolution structure and the updated node feature representation. The prediction result is 0, 0, 0, 0, 0, 1, 1, 1, which is consistent with the true label.

[0076] Example 2

[0077] See also Figures 1 to 4 This embodiment provides a schizophrenia classification method based on a fuzzy hypergraph neural network based on evidence theory, comprising the following steps:

[0078] Step 1. This embodiment selects 773 nodes from schizophrenia datasets from five different sites as an example for illustration; uses the DPARSF tool to preprocess the data; uses the automatic anatomical labeling atlas template to extract the time series of the region of interest, divides the cerebral cortex into 90 functional brain regions with adjacent positions and similar functions, and obtains the time series information of different functional brain regions, wherein this dataset has two labels, namely healthy subjects and schizophrenia patients;

[0079] Step 2, this embodiment completes the hyperedge modeling task through the sparse constraint function: based on the sparse constraint function, construct the association matrix of 773 nodes; characterize the optimal k value of 773 nodes through the association matrix and the correlation threshold; construct a hyperedge granularity model based on the k nearest neighbor samples to characterize the high-order correlation between schizophrenia data;

[0080] Step 2.1: This embodiment establishes a hyperedge granularity model through a sparse constraint function, where the node feature matrix X is expressed as The number of nodes n is expressed as 773, and the node feature dimension d is 90; by minimizing the reconstruction error of the nodes, the association matrix R between the 773 nodes is obtained; the sparse constraint function is calculated:

[0081]

[0082] Among them, ||·|| F represents the Frobenius norm of the matrix, ||·||2 is the l2-norm regularization term, L represents the Laplace matrix, ρ1 and ρ2 are adjustment parameters, defined as 0.1 and 0.01 respectively, Tr(·) represents the calculation function of the matrix trace, R T With X T is the transpose of the matrices R and X;

[0083] Step 2.2: Based on the sparse constraint function of step 2.1, obtain the correlation matrix R between the 773 schizophrenia data nodes; based on the correlation threshold σ = 0.1, determine the optimal k value of the 773 schizophrenia data nodes; for the schizophrenia data node x i , computing node x i The optimal k value k i :

[0084] k i =num(R(i)>σ) (32)

[0085] Where R(i) is the i-th column of the incidence matrix R, and num is the counting function;

[0086] Step 2.3: According to the optimal k value obtained in step 2.2, calculate the k nearest neighbor samples and use the k nearest neighbor samples as the schizophrenia data nodes within the hyperedge granularity; for the i-th hyperedge, calculate the hyperedge granularity HG k (x i ):

[0087]

[0088] Among them, k i For schizophrenia data node x i The optimal k value is d(x m ,x n ) are any two schizophrenia data nodes x m With x n The Euclidean distance function between .

[0089] Step 3, this embodiment constructs a fuzzy hypergraph through evidence theory: construct an association quality function based on the association matrix of 773 nodes obtained in step 2; calculate the distance quality function through the distance between 773 nodes; apply the quality function to characterize the association coefficient of 773 nodes for hyperedges; apply evidence theory, fuse the quality function, and calculate the fuzzy membership of 773 nodes for hyperedges; obtain the fuzzy adjacency matrix of the fuzzy hypergraph, thereby constructing the fuzzy hypergraph;

[0090] Step 3.1: Get the hyperedge through step 2 x i is any node of schizophrenia data; calculate the schizophrenia data node x i With the central node x j The Euclidean distance d(x i ,x j ), get node x i With super edge e j The distance similarity re ij , calculate the schizophrenia data node x i For the distance quality function of hyperedge association:

[0091]

[0092]

[0093] Among them, m is the distance adjustment parameter, defined as 2, k j is the hyperedge e j The number of internal nodes;

[0094] Step 3.2: Based on the association vector R(j) obtained in step 2.2, obtain the hyperedge The correlation coefficient of all schizophrenia data nodes in the hyperedge; according to any schizophrenia data node x in the hyperedge i For the hyperedge e j Calculate the correlation coefficient of schizophrenia data node x i For the associated mass function of a hyperedge:

[0095]

[0096]

[0097] Among them, the hyperedge e j The number of nodes is k j , super edge e j The correlation coefficient between the ith node and the hyperedge is defined as r ij ;

[0098] Step 3.3: Based on the schizophrenia data node x in step 3.1 and step 3.2 i For the hyperedge e j The distance quality function and association quality function of the schizophrenia data node x are calculated. i For the hyperedge e j The fuzzy membership u ij :

[0099]

[0100]

[0101]

[0102]

[0103] Among them, X A and X B For node x containing schizophrenia data i The node set of , ρ is the evidence adjustment parameter, defined as 0.01;

[0104] Step 3.4: Obtain the schizophrenia data node x by calculating in step 3.3 i For the hyperedge e j Fuzzy membership u ij , which is the corresponding element of the jth column of the fuzzy adjacency matrix of the fuzzy hypergraph; repeat step 3.3 to calculate the fuzzy adjacency matrix H F :

[0105]

[0106] Where n is the number of rows in the node matrix X, defined as 773, and i and j are any number of rows and columns in the matrix.

[0107] Step 4: This embodiment is based on the fuzzy adjacency matrix H obtained in step 3 F And the hypergraph neural network model, constructing a fuzzy hypergraph convolution model: Based on the fuzzy adjacency matrix H F The features of nodes and hyperedges are aggregated to construct a fuzzy hypergraph convolution operator. Based on the fuzzy hypergraph convolution operator, a two-layer convolution structure for node label prediction is designed to obtain node labels.

[0108] Step 4.1: Based on step 3, we get the fuzzy adjacency matrix H. F , obtain the fuzzy membership of the schizophrenia data node to the hyperedge and the correlation coefficient of the hyperedge to the schizophrenia data node; based on the fuzzy membership and node-edge-node aggregation operation, aggregate the features of the hyperedge and the node to obtain the fuzzy convolution operator:

[0109]

[0110] Among them, X l With X (l+1) is the feature representation of the fuzzy hypergraph at the l and l+1 layers, σ(·) is the activation function, and H F T is the matrix H F The transpose of D v and D e For nodes and hyperedges based on H F The degree matrix, W is the identity matrix, Θ l is the trainable weight parameter matrix of layer l;

[0111] Step 4.2: Based on the fuzzy convolution operator obtained in step 4.1, a two-layer convolution structure is designed for node label prediction of schizophrenia data to obtain the label Z of the schizophrenia subject:

[0112] Z = Softmax(H(ReLU(HXΘ) (0) ))Θ (1) ) (44)

[0113]

[0114] Among them, Θ (0) is the weight matrix from the first layer to the hidden layer, Θ (1) is the weight matrix from the hidden layer to the last layer, Softmax(·) and ReLU(·) are the activation functions.

[0115] In order to highlight the superiority of the model constructed in this embodiment, this embodiment selects 5 popular graph neural networks and hypergraph neural networks as comparison algorithms: GCN, HGNN, HyperGCN and HGNN+. This embodiment uses accuracy to characterize the performance of the model designed in this embodiment and the comparison algorithm.

[0116] GCN is the most widely used model in node classification tasks. It encodes schizophrenia data into a topological graph and inputs it into a neural network to complete the semi-supervised classification of schizophrenia patients. The HyperGCN model aims to build a hypergraph model based on the semi-supervised learning task of the hypergraph and convert it into a topological graph to further improve the prediction accuracy of schizophrenia patients. HGNN is used to encode high-order data correlations in the hypergraph structure, effectively process schizophrenia multimodal data, and apply the correlation of different modal data to improve the classification accuracy of schizophrenia. HGNN+ is an extension of the original HGNN model, which further integrates data of different modalities of schizophrenia and reduces the misdiagnosis rate of schizophrenia.

[0117] The model proposed in this embodiment is based on the hypergraph neural network model. Aiming at the heterogeneity in schizophrenia data, fuzzy membership is used to characterize the uncertainty between nodes and hyperedges. And different association quality functions are integrated by applying evidence theory to construct a fuzzy hypergraph neural network based on evidence theory. This embodiment designs a fuzzy hypergraph neural network classification method based on evidence theory, with a training set of 653 nodes and a test set of 120 nodes. The performance of different graphs and hypergraph neural network classification algorithms is shown in Table 1.

[0118] Table 1 Experimental results of hypergraph neural network on schizophrenia dataset

[0119]

[0120] The results of this embodiment and other hypergraph neural network classification methods are shown in Table 1. FHGNN is the model proposed by the present invention. Compared with other hypergraph neural networks GCN, HyperGCN, HGNN, and HGNN+, the classification accuracy of FGHNN is 60.5%, 59.6%, 63.0%, 61.5%, and 71.4%, respectively. The above experimental results prove that FHGNN can reduce the heterogeneity in schizophrenia data and improve the prediction accuracy of schizophrenia patients.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A schizophrenia classification method based on fuzzy hypergraph neural network based on evidence theory, characterized in that: The following steps are involved: Step 1: Select data of schizophrenia patients from 5 different sites as the data set; use the DPARSF tool to process the data, where the preprocessing methods include inter-layer time correction, spatial normalization, global average intensity normalization, interference signal regression and bandpass filtering; use the automatic anatomical labeling atlas template to extract the time series of the region of interest (ROI), which divides the cerebral cortex into 90 functional brain areas with adjacent positions and similar functions, and obtain the time series information of different functional brain areas; Step 2: Complete the hyperedge modeling task through the sparse constraint function: construct the association matrix of schizophrenia data nodes based on the sparse constraint function; characterize the optimal k value of different schizophrenia data nodes through the association matrix and the correlation threshold; construct a hyperedge granularity model based on the k nearest neighbor samples to characterize the high-order correlation between schizophrenia data; Step 3, construct a fuzzy hypergraph through evidence theory: construct an association quality function based on the schizophrenia data node association matrix obtained in step 2; calculate the distance quality function through the distance between the schizophrenia data nodes; use the quality function to characterize the association coefficient of the schizophrenia data node to the hyperedge; apply the evidence theory, integrate the quality function, and calculate the fuzzy membership of the schizophrenia data node to the hyperedge; obtain the fuzzy adjacency matrix of the fuzzy hypergraph, and thus construct the fuzzy hypergraph; Step 4: Based on the fuzzy adjacency matrix H obtained in step 3 F And hypergraph neural network model, construct fuzzy hypergraph convolution model: based on fuzzy adjacency matrix H F The features of schizophrenia data nodes and hyperedges are aggregated to construct a fuzzy hypergraph convolution operator. Based on the fuzzy hypergraph convolution operator, a two-layer convolution structure is designed for schizophrenia data node label prediction to obtain the labels of schizophrenia subjects.

2. The schizophrenia classification method based on fuzzy hypergraph neural network based on evidence theory according to claim 1 is characterized in that: The step 2 comprises the following steps: Step 2.1: Establish a hyperedge granularity model through a sparse constraint function, where the schizophrenia data node feature matrix X is expressed as n represents the number of schizophrenia data nodes, d represents the node feature dimension; by minimizing the reconstruction error of schizophrenia data nodes, the correlation matrix R between nodes is obtained; Compute the sparse constraint function: Among them, ||·|| F represents the Frobenius norm of the matrix, ||·||2 is the l2-norm regularization term, L represents the Laplace matrix, ρ1 and ρ2 are adjustment parameters, Tr(·) represents the calculation function of the matrix trace, and R T With X T is the transpose of the matrices R and X; Step 2.2: Based on the sparse constraint function of step 2.1, obtain the correlation matrix R between schizophrenia data nodes; based on the correlation threshold σ, determine the optimal k value for all schizophrenia data nodes; for schizophrenia data node x i , computing node x i The optimal k value k i : k i =num(R(i)>σ) (2) Where R(i) is the i-th column of the incidence matrix R, and num is the counting function; Step 2.3: According to the optimal k value obtained in step 2.2, calculate the k nearest neighbor samples and use the k nearest neighbor samples as the schizophrenia data nodes within the hyperedge granularity; for the i-th hyperedge, calculate the hyperedge granularity HG k (x i ): Among them, k i For schizophrenia data node x i The optimal k value, d(x i ,x1)≤d(x i ,x2)≤...≤d(x i ,x ki ), d(x m ,x n ) are any two schizophrenia data nodes x m With x n The Euclidean distance function between .

3. The schizophrenia classification method based on fuzzy hypergraph neural network based on evidence theory according to claim 1 is characterized in that: The step 3 comprises the following steps: Step 3.1: Through step 2, get the hyperedge e j ={x1,x2,..,x kj }, x i is any node of schizophrenia data; calculate the schizophrenia data node x i With the central node x j The Euclidean distance d(x i ,x j ), get node x i With super edge e j The distance similarity re ij , calculate the schizophrenia data node x i For the distance quality function of hyperedge association: Among them, m is the distance adjustment parameter, k j is the hyperedge e j Number of internal nodes; Step 3.2: Based on the association vector R(j) obtained in step 2.2, obtain the hyperedge The correlation coefficient of all schizophrenia data nodes in the hyperedge; according to any schizophrenia data node x in the hyperedge i For the hyperedge e j Calculate the correlation coefficient of schizophrenia data node x i For the associated mass function of a hyperedge: R(j)=[r 1j r 2j ...r ij ...r kjj ] (6) Among them, the hyperedge e j The number of nodes is k j , super edge e j The correlation coefficient between the ith node and the hyperedge is defined as r ij ; Step 3.3: Based on the schizophrenia data node x in step 3.1 and step 3.2 i For the hyperedge e j The distance quality function and association quality function of the schizophrenia data node x are calculated. i For the hyperedge e j The fuzzy membership u ij : Among them, X A and X B For node x containing schizophrenia data i The node set of , ρ is the evidence adjustment parameter; Step 3.4: Obtain the schizophrenia data node x by calculating in step 3.3 i For the hyperedge e j Fuzzy membership u ij , which is the corresponding element of the jth column of the fuzzy adjacency matrix of the fuzzy hypergraph; repeat step 3.3 to calculate the fuzzy adjacency matrix H F : Where n is the number of rows in the node matrix X, and i and j are any number of rows and columns in the matrix.

4. The schizophrenia classification method based on fuzzy hypergraph neural network based on evidence theory according to claim 1 is characterized in that: The step 4 comprises the following steps: Step 4.1: Based on step 3, we get the fuzzy adjacency matrix H. F , obtain the fuzzy membership of the schizophrenia data node to the hyperedge and the correlation coefficient of the hyperedge to the schizophrenia data node; based on the fuzzy membership and node-edge-node aggregation operation, aggregate the features of the hyperedge and the node to obtain the fuzzy convolution operator: Among them, X l With X (l+1) is the feature representation of the fuzzy hypergraph at the l and l+1 layers, σ(·) is the activation function, and H F T is the matrix H F The transpose of D v and D e For nodes and hyperedges based on H F The degree matrix, W is the identity matrix, Θ l is the trainable weight parameter matrix of layer l; Step 4.2: Based on the fuzzy convolution operator obtained in step 4.1, a two-layer convolution structure is designed for node label prediction of schizophrenia data to obtain the label Z of the schizophrenia subject: Z=Softmax(H(ReLU(HXΘ (0) ))Θ (1) ) (14) Among them, Θ (0) is the weight matrix from the first layer to the hidden layer, Θ (1) is the weight matrix from the hidden layer to the last layer, Softmax(·) and ReLU(·) are the activation functions.

Images