Autism identification method based on graph enhanced path convolutional network

Abstract

The invention discloses an autism identification method based on a graph enhanced path convolutional network, relates to the technical field of biomedical information, and provides a new GNN framework which is used for classifying brain diseases in an end-to-end rs-fMRI mode. According to the method, a path convolutional network is used for learning features of multiple paths, a Transform is used for capturing dynamic spatio-temporal information to strengthen the features, a relation between non-imaging information and a subject is brought into a framework, singular value decomposition, RES + SE enhanced data and channel influence of the enhanced data are provided, and the multi-center problem of ABIDE is alleviated. According to the method, the overall performance of the model is remarkably improved, a gradient method is proposed to find the importance of features and the important features learned by R-pool, and the two are mutually verified, so that the reliability of the method is proved.

Description

Technical Field

[0001] The present invention relates to the field of biomedical information technology, and in particular to an autism recognition method based on a graph-enhanced path convolutional network. Background Art

[0002] Autism Spectrum Disorder (ASD) is a common neurodevelopmental disorder characterized by difficulties in social interaction and communication, stereotyped behaviors and interests, and hypersensitivity to sensation. In recent years, the incidence of autism has continued to rise, becoming a public health issue of concern worldwide [1]. According to the latest research, the pathological mechanisms of autism involve a variety of factors, including complex interactions among genetics, environment, and neurodevelopment [2]. A large number of studies have explored the potential pathological mechanisms of brain diseases through the use of brain imaging techniques, such as functional magnetic resonance imaging (fMRI) [3]-[7], especially in terms of functional connectivity (FC) within and between brain networks. FC refers to the temporal correlation between neural activity in different brain regions and is believed to reflect the strength and efficiency of communication between these regions. Mental health and behavioral characteristics are closely related to changes in functional connectivity measured by rs-fMRI.

[0003] In the past few years, several statistical and machine learning methods have been used to predict autism using rs-fMRI and brain functional connectivity. Classification is performed based on the selected features of ridge classifiers, logistic regression, and support vector machines (SVM). Deep learning-based methods have shown good feature extraction capabilities in brain medical image analysis, such as convolutional neural networks (CNNs), which have superior capabilities to extract features from regularly structured data and have achieved great success in image-based classification and regression problems. More and more researchers are adopting these data-driven techniques to identify potential neuroimaging biomarkers for automatic brain disease recognition, but these methods all treat brain connectivity as a general matrix or independent vector without considering the specific spatial structure of the brain network, which limits their ability to extract features, classification, and regression tasks. Later, the graph convolutional neural network (GCN) was proposed to combine deep neural networks with graph representations by capturing the topological information of brain networks, so it is very suitable for analyzing the relationship between different brain regions. De Vlade et al. used Chebyshev polynomials to approximate the eigenvalue matrix of the graph Laplacian matrix. Kipf and Welling further simplified the calculation and developed a hierarchical convolution structure, making GCN easy to implement through matrix multiplication. Parisot et al. proposed a population graph and used GCN to predict autism and Alzheimer's disease.

[0004] However, existing research usually faces three challenges. First, due to the heterogeneity of data collected by different research centers, the multi-center problem becomes a challenge. For example, fMRI data from different research institutions may have problems such as differences in scanning parameters and instrument differences, which will have a negative impact on the reliability and generalization ability of the data. Secondly, static functional connectivity map analysis methods often ignore the dynamics of brain networks, and the brain activity of patients with autism is constantly changing, so it is important to capture and understand dynamic information. In addition, there are challenges in the interpretability of autism brain functional connectivity. When selecting important features based on GCN, using traditional methods to calculate the importance of nodes in the GCN model, such as class activation mapping (CAM) or gradient-weighted class activation mapping (Grad-CAM), may cause the adjacency matrix to be asymmetric, thereby violating the definition of the graph model. In previous studies, biomarkers can only be interpreted in combination with bioinformatics, which limits the progress of discovering biomarkers and understanding disease mechanisms. Summary of the invention

[0005] The purpose of the present invention is to provide an autism recognition method based on a graph-enhanced path convolutional network to solve the problems raised in the above background technology.

[0006] Our proposed SSR-GCN framework consists of four steps, including (step A) computation of slice-based dynamic functional connectivity matrix, (step B) SVD-SR module, (step C) multi-path feature convolution extraction, and (step D) transformer module to capture information.

[0007] We will introduce the construction of a dynamic brain functional connectivity matrix and its SVD-SR module in turn, then explain the strategy of the multi-path convolutional GCN framework, introduce the process of transformer capturing dynamic information, and finally use gradients to view important features.

[0008] A. Constructing a connectivity matrix

[0009] The input is the sequential rs-fMRI data based on brain regions, namely the blood oxygen level dependent (BOLD) signal. In order to utilize the dynamic functional connectivity information, the spatiotemporal dynamic changes between brain regions are obtained from multiple time points. By dividing the data into continuous sliding windows, we can capture the dynamic characteristics of the brain network and obtain multiple functional connectivity matrices, thereby more comprehensively describing the association pattern between brain regions. The time sliding window can be expressed as the following formula:

[0010] N=(TW) / I+1

[0011] Where N represents the number of windows, T represents the total time length of the data, W represents the window size, and I represents the window interval.

[0012] By calculating the number of windows N, we can determine how many consecutive sliding windows can be obtained given the total time length T of the data. These sliding windows can be used to capture the spatiotemporal dynamics between brain regions and obtain multiple functional connectivity matrices for subsequent analysis. Due to the heterogeneity of multi-center data resulting in different total time lengths, we adopted different window intervals to ensure that each participant has the same number of sliding windows. By adjusting the window interval, we can balance the number of data samples and keep the number of sliding windows for each participant consistent while keeping the sliding window size unchanged. Here we choose N as 3.

[0013] Then, the average time series of brain regions were extracted according to the automated anatomical atlas (AAL), and the functional connectivity matrix was calculated from the time series of brain regions based on the Pearson correlation coefficient. These matrices are also called dynamic functional connectivity (DFC) matrices.

[0014] B. SVD-SR module

[0015] Taking brain regions as nodes and their connections as edges, we can view brain connections as a graph. Based on the fully connected graph obtained in the previous section, a graph can be represented as G = (N, E), where N is the set of nodes and E is the set of edges. That is, if we set n as the number of nodes, then the number of edges is n 2 , that is, |N|=n and |E|=n 2 .

[0016] The ABIDE dataset is provided by multiple data centers, which may have different data acquisition devices, scanning parameters, and preprocessing methods. This heterogeneity may lead to differences between the DFC matrix data we obtain, affecting the performance and generalization ability of the model. To alleviate this limitation, we designed the SVD-SR module, which involves SVD technology, SENet, and ResNet technology.

[0017] The SVD technique can be expressed as the following formula:

[0018] A=UΣV T =u1σ1v1 T +u2σ2v2 T +...+u r σ r v r T

[0019] Among them, u1, u2, ..., u r is the column vector of U, v1, v2, ..., v r is the column vector of V, σ1, σ2, ..., σ ris the singular value on the diagonal of Σ, and A is the functional connectivity matrix obtained in the previous section. Larger singular values ​​correspond to the main information or important features in the matrix, while smaller singular values ​​correspond to secondary information or noise. By retaining larger singular values, we can choose to retain the most important features in the data. We select the first ten singular values ​​to extract the common features of the data, and then concatenate them with the original matrix and continue to perform matrix decomposition to retain the first 16 singular values ​​to form a new functional connectivity matrix.

[0020] SENet (Squeeze-and-Excitation Network) is an attention mechanism for enhancing deep convolutional neural networks, aiming to improve the representation ability and generalization performance of the network. SENet introduces an SE module to adaptively learn the relationship between channels and weight the channel features. The input feature map is reduced in dimension by a global average pooling layer, and the feature map of each channel is converted into a 1×1×C feature map (C is the number of channels), which is then learned through two fully connected layers to obtain channel weights. The first fully connected layer maps the input 1×1×C feature map to a smaller dimension, usually C / r (where r is a configurable scaling factor to reduce the number of parameters), and the second fully connected layer maps it back to the original 1×1×C dimension, normalized by the Sigmoid function, and obtains the weight of each channel. Finally, the output is element-wise multiplied with the original input feature map to obtain the weighted feature map. This weighting mechanism enables the network to adaptively adjust the importance of each channel, thereby improving the representation ability of the network.

[0021] Y=X*F(X,W)

[0022] Among them, X represents the input feature map, Y represents the weighted output feature map, F represents the operation of the SE module (including Squeeze and Excitation), and W represents the parameters of the SE module. * represents the element-level multiplication operation.

[0023] The basic unit of ResNet is the residual block, each of which consists of two or three convolutional layers. The input of the residual block is directly added to the output through a skip connection to form a residual connection. The formula of ResNet can be expressed as:

[0024] Y=F(X)+W*X

[0025] Among them, X represents the input feature map, Y represents the output feature map, F represents the transformation operation in the residual block, and W is a weight matrix. + represents the element-level addition operation. The entire network consists of multiple residual blocks, and the flow of information is achieved through residual connections. After these operations, new features are constructed, and then feature convolution is performed.

[0026] C. GPC multi-path convolution

[0027] Both direct and indirect contacts between brain regions play an important role in brain communication. For example, the communication between brain region i and brain region j not only has connection ij, but also indirect connections. We define direct connections as 0-order paths, indirect connections that only pass through another node as first-order paths, and so on for second-order paths and third-order paths. As mentioned earlier, these paths may include important information about the relationship between brain regions. In order to aggregate multi-order information from these paths, we use the GPC model to aggregate this information. Its formula is

[0028]

[0029] In the formula, represents the k between node i and node j th s-order path, GP i,j represents the aggregation of all paths between i and j. S restricts the highest order of the paths, is the number of paths of order s. s,k and θ s is a learnable parameter, i.e., w s,k It is k th The weight of the s-order path, θ s is an s-order path. GPC can aggregate the data in the 0-order path into the s-order path and extract meaningful features from these paths.

[0030] D. Variant Transformer

[0031] The Transformer model consists of an encoder and a decoder. The encoder is responsible for encoding the input sequence, and the decoder is responsible for generating the target sequence based on the encoder's output. A set of attention weights are calculated for each position in the input sequence, which represent the relative importance of different positions in the input sequence. The calculation of this attention weight can be obtained through three linear transformations: query, key, and value. The attention weight is obtained by multiplying the query with the transpose of the key and normalizing it. The attention weight is then multiplied by the value and weighted summed to obtain the self-attention output for each position. Multi-head attention is achieved by performing multiple parallel operations on the self-attention mechanism to obtain richer representation capabilities. Assuming there are H attention heads, the query, key, and value matrices of each head are linearly transformed through different weight matrices.

[0032] Multi-Head Attention Output=Concatenate(Head_1,Head_2,...,Head_H)*W_o

[0033] Among them, Head_i=Self-Attention Output_i

[0034] This is the main content of the traditional transformer, which realizes the modeling and feature extraction of sequence data. We have reformed the transformer. First, multiple functional connection matrices generated by a subject are convolved through GPC to generate multiple sequences. The sequences generated by the same subject are encoded and standardized to ensure that they have consistent formats and scales. Multiple attention heads are used to capture different relationships between sequences. Each input sequence can be used as an input channel of the model, and the interactions and dependencies between sequences are learned through self-attention mechanisms and multi-head attention mechanisms. In the case of multiple input sequences, a specific attention weight calculation method is designed to make full use of the correlation information between different input sequences. In the previous article, we designed that a subject would generate three sequences. The attention weight calculation formula is as follows:

[0035]

[0036] where d k is the scaling factor of the attention, which usually depends on the dimension of the encoding vector, ABC is the input sequence, and H- abc For the encoded representation, Attention(HB, HA, HC) and Attention(HC, HA, HB) are calculated in the same way. Using the calculated attention weights, the encoded representations of the other two sequences are weighted added to obtain the weighted representation of each sequence.

[0037] Weighted A =Attention(H A ,H B , H C )+H A

[0038] Similarly, we can also calculate the weighted representation of sequence B and sequence C. The spatiotemporal relationship between sequences can be captured and more accurate predictions can be made in subsequent tasks.

[0039] E. Interpretability

[0040] In our study, we focused on the interpretability of the model to better understand the relationship between functional connectivity maps and autism. To this end, we used a variety of interpretability methods to explain the model's predictions and the associations between brain regions.

[0041] First, to perform graph-level classification, a layer for dimensionality reduction is required, since the number of nodes and feature dimensions of each node are large. Recent findings suggest that some ROIs are more predictive of neurological diseases than others, suggesting that they should be kept in the dimensionality reduction step. We introduce a node (ROI) pooling layer (R-pool), which aims to retain the most representative ROIs and remove noisy nodes, thereby reducing the dimensionality of the entire graph. We adopt the method in reducing graph nodes. The choice of which nodes to remove is based on the number of nodes that are removed.

[0042] The features are projected into a learnable vector weight (W (l) ) to determine:

[0043]

[0044] Where N( l ) is the number of nodes in the hth layer.

[0045] Secondly, we took another approach to find important features. Our predicted output y and loss L (we use cross entropy as our loss) are compared with the true label. Then, by chain rule, the gradient can be calculated as:

[0046]

[0047] where x i is the i of input x th feature.

[0048] It can be calculated by the automatic back propagation (BP) algorithm of deep learning frameworks such as PyTorch. It is worth noting that is the gradient that increases L, and the inverse gradient is xi To the meaning of y.

[0049]

[0050] Among them, x i i represents the input x th The importance of the feature to the output y.

[0051] In the SSR-GPCsT model, a feature importance coefficient is assigned to each node feature, which reflects the contribution of the feature to the node classification task. These feature importance coefficients are calculated by averaging the gradient of each node. The formula is as follows:

[0052]

[0053] Among them, f i represents the feature importance coefficient of node i, N represents the number of neighbors of node i, L represents the loss function of the node classification task, The feature vector representing the edge between node j and node i.

[0054] Our idea is to use the trained model and the normal people in the data to find the gradient, and change the label of the normal person to the label of the autistic patient, that is, 1 becomes 0. The gradient actually means how the brain connection will change if the subject becomes ill. In this way, we can select the important features of the classification task.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. We propose a novel GNN framework for classifying brain diseases in an end-to-end rs-fMRI manner.

[0057] 2. It uses a path convolutional network to learn features of multiple paths and uses Transformer to capture dynamic spatiotemporal information to enhance features and incorporate non-imaging information and relationships between subjects into the framework.

[0058] 3. A singular value decomposition and RES+SE are proposed to enhance the data and channel influence of the enhanced data, and alleviate the multi-center problem of ABIDE. It significantly improves the overall performance of the model.

[0059] A method using gradients is proposed to find the importance of features, and the important features learned using R-pool are verified by each other, proving the reliability of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention.

[0061] Figure 2 4 is a schematic diagram of an SVD-SE module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] We used a public dataset ABIDE, which contains data from 505 ASD patients and 530 controls (typical controls, TC). The dataset collects data from 17 different imaging sites, including rs-fMRI images, T1 brain structural images, and phenotypic information for each patient. The graph-enhanced path convolutional network with singular spectral residual transform (SSR-GPCsT) was implemented on the PyTorch framework. The network optimization used the stochastic gradient descent (SGD) method with loss = 0.5, lr = 0.0005, momentum = 0.9, weight decay = 1e-5, and epoch = 2000. Each dataset was evaluated using 5-fold stratified cross-validation. The classification results were averaged over 10 tests.

[0064] Singular value decomposition is used to obtain the main information in the functional connection matrix, extract common features to reduce its multi-center, and larger singular values ​​correspond to the main information or important features in the matrix. By retaining larger singular values, we introduced the SE module to adaptively learn the relationship between channels and weighted channel features. This weighting mechanism enables the network to adaptively adjust the importance of each channel, thereby improving the network's representation ability, and use residual connections to form new features. Then, GPC can be used to aggregate the data in the s-order path into a 0-order path, and meaningful classification features can be extracted from these paths. In order to obtain more information, the information source is expanded to the time dimension, dynamic data is obtained, and Transformer is used to combine them, and finally more accurate predictions are made in the fully connected layer.

[0065] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0066] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for identifying autism based on graph-enhanced path convolutional networks, characterized in that: The steps include: (1) Constructing a connection matrix: Inputting sequential rs-fMRI data based on brain regions, namely, blood oxygen level dependent signal BOLD, in order to utilize dynamic functional connectivity information and obtain the spatiotemporal dynamic changes between brain regions from multiple time points, the data is divided into continuous sliding windows to capture the dynamic characteristics of the brain network and obtain multiple functional connectivity matrices, thereby more comprehensively describing the association pattern between brain regions. The time sliding window is expressed as the following formula: N=(TW) / I+1 Where N represents the number of windows, T represents the total time length of the data, W represents the window size, and I represents the window interval; By calculating the number of windows N, we determined how many continuous sliding windows we could obtain under a given total data time length T. These sliding windows were used to capture the spatiotemporal dynamics between brain regions and obtain multiple functional connectivity matrices for subsequent analysis. Due to the heterogeneity of multi-center data, the total time length was different, so different window intervals were used to ensure that each participant had the same number of sliding windows. By adjusting the window interval, we balanced the number of data samples and kept the sliding window size unchanged so that the number of sliding windows for each participant was consistent. We selected N as 3. Then, the average time series of brain regions were extracted according to the automated anatomical atlas AAL, and the dynamic functional connectivity matrix DFC was calculated from the time series of brain regions based on the Pearson correlation coefficient; (2) Constructing the SVD-SR module: The brain regions are regarded as nodes, their connections are regarded as edges, and the brain connections are regarded as a graph. According to the previous section, a fully connected graph is obtained. A graph is represented as G = (N, E), where N is the set of nodes and E is the set of edges. Let n be the number of nodes, then the number of edges is n 2 , that is, |N|=n and |E|=n 2 ; Design the SVD-SR module, using SVD technology, SENet and ResNet technology; The SVD technique is expressed as the following formula: A=UΣV T =u1σ1v1 T +u2σ2v2 T +...+u r s r v r T Among them, u1, u2, ..., u r is the column vector of U, v1, v2, ..., v r is the column vector of V, σ1, σ2, ..., σ r is the singular value on the diagonal of Σ, A is the functional connection matrix obtained in the previous section, the larger singular value corresponds to the main information or important features in the matrix, and the smaller singular value corresponds to the secondary information or noise. By retaining the larger singular value, the most important features in the data are retained, and the first ten singular values ​​are selected to extract the common features of the data. At the same time, after splicing with the original matrix, the first 16 singular values ​​are retained after matrix decomposition to form a new functional connection matrix. SENet technology: introduces an SE module to adaptively learn the relationship between channels and weight channel features. The input feature map is reduced in dimension through the global average pooling layer, and the feature map of each channel is converted into a 1×1×C feature map, where C is the number of channels. It is then learned through two fully connected layers to obtain channel weights. The first fully connected layer maps the input 1×1×C feature map to a smaller dimension, C / r, where r is a configurable scaling factor used to reduce the number of parameters. The second fully connected layer maps it back to the original 1×1×C dimension and normalizes it through the Sigmoid function to obtain the weight of each channel. Finally, the output is multiplied element-wise with the original input feature map to obtain the weighted feature map. The specific formula is as follows: Y = X * F (X, W) Among them, X represents the input feature map, Y represents the weighted output feature map, F represents the operation of the SE module, including Squeeze and Excitation, W represents the parameters of the SE module, and * represents the element-level multiplication operation; ResNet technology: The basic unit of ResNet is the residual block. Each residual block consists of two or three convolutional layers. The input of the residual block is directly added to the output through a skip connection to form a residual connection. The formula of ResNet is expressed as: Y=F(X)+W*X Among them, X represents the input feature map, Y represents the output feature map, F represents the transformation operation in the residual block, W is a weight matrix, + represents the element-level addition operation, and the entire network is composed of multiple residual blocks. The flow of information is realized through residual connections. New features are constructed through these operations, and then feature convolution is performed; (3) GPC multi-path convolution: The communication between brain area i and brain area j not only has connection ij, but also indirect connection. The direct connection is defined as the 0th order path, and the indirect connection through only another node is defined as the first order path. The second order path and the third order path are similar. In order to aggregate multi-order information from these paths, the GPC model is used to aggregate this information. The formula is: In the formula, represents the k between node i and node j th s-order path, GP i,j represents the aggregation of all paths between i and j, S limits the highest order of the path, is the number of s-order paths, w s,k and θ s is a learnable parameter, i.e., w s,k It is k th The weight of the s-order path, θ s It is an s-order path. GPC aggregates the data in the 0-order path into the s-order path and extracts meaningful features from these paths. D. Variant Transformer: The Transformer model consists of an encoder and a decoder r. The encoder is responsible for encoding the input sequence, and the decoder is responsible for generating the target sequence based on the output of the encoder. A set of attention weights are calculated for each position in the input sequence. These weights represent the relative importance of different positions in the input sequence. The attention weights are calculated through three linear transformations: query, key, and value. The attention weights are obtained by multiplying the query with the transpose of the key and normalizing it. The attention weights are then multiplied by the value and weighted summed to obtain the self-attention output for each position. Multi-head attention is achieved by performing multiple parallel operations on the self-attention mechanism to obtain richer representation capabilities. Assuming there are H attention heads, the query, key, and value matrices of each head are linearly transformed through different weight matrices: Multi-Head Attention Output=Concatenate(Head_1, Head_2,..., Head_H)*W_o Among them, Head_i=Self-Attention Output_i The transformer was reformed. First, multiple functional connection matrices generated by a subject were convolved with GPC to generate multiple sequences. The sequences generated by the same subject were encoded and standardized to ensure that they have consistent format and scale. Multiple attention heads were used to capture different relationships between sequences. Each input sequence was used as an input channel of the model. The interactions and dependencies between sequences were learned through self-attention mechanism and multi-head attention mechanism. In the case of multiple input sequences, a specific attention weight calculation method was designed to make full use of the correlation information between different input sequences. It was designed that one subject would generate three sequences. The attention weight calculation formula is as follows: where d k is the scaling factor of the attention, which depends on the dimension of the encoding vector, ABC is the input sequence, and H- abc The encoded representation is similarly calculated as Attention(HB, HA, HC) and Attention(HC, HA, HB). Using the calculated attention weights, the encoded representations of the other two sequences are weighted added to obtain the weighted representation of each sequence: Weighted A =Attention(H A ,H B ,H C )+H A Similarly, the weighted representations of sequence B and sequence C are calculated to capture the spatiotemporal relationship between the sequences and make more accurate predictions in subsequent tasks. (5) Interpretability: A variety of interpretability methods are used to explain the model's prediction results and the associations between brain regions. First, graph-level classification is performed, which requires a layer for dimensionality reduction. Because the number of nodes and feature dimensions of each node are large, a node (ROI) pooling layer (R-pool) is introduced to retain the most representative ROIs and remove noise nodes, thereby reducing the dimension of the entire graph. The method in is used to reduce the graph nodes. The choice of which nodes to delete is based on projecting the node features onto a learnable vector weight (W (l) ) to determine: Where N (l) is the number of nodes in the hth layer; Second, the predicted output y and loss L are compared with the true label, and then by the chain rule, the gradient is calculated as: where x i is the i of input x th feature; Calculated through the automatic back propagation (BP) algorithm of the PyTorch deep learning framework, is the gradient that increases L, and the inverse gradient is x i The meaning of y: Among them, x i i represents the input x th The importance of the feature to the output y; In the SSR-GPCsT model, a feature importance coefficient is assigned to each node feature, which reflects the contribution of the feature to the node classification task. These feature importance coefficients are calculated by averaging the gradient of each node. The formula is as follows: Among them, f i represents the feature importance coefficient of node i, N represents the number of neighbors of node i, L represents the loss function of the node classification task, The feature vector representing the edge between node j and node i; After the model is trained, the normal people in the data are used to find the gradient. The labels of normal people are changed to labels of autistic patients, that is, 1 becomes 0. The gradient actually means how the brain connections will change if the subject becomes ill, and the important features of the classification task are selected.

Images