An ASD identification method based on dynamic graph embedding

Abstract

The application discloses an ASD identification method based on dynamic graph embedding and belongs to the technical field of computer application. The ASD and typical control subjects' dynamic brain networks are constructed by using a sliding window and Pearson correlation, the brain region embedding of the dynamic brain networks is calculated by using dynamic graph embedding, all brain regions of each subject are clustered based on the brain region embedding, and the aggregate brain networks of each subject are constructed according to the clustering results, further independent sample T test and Cohen's D are used for inter-group analysis between the ASD group and the typical control group, thereby obtaining the aggregate brain network connection features with significant differences, and the ASD is identified. The method disclosed by the application can capture the dynamic interaction relationship between brain regions in the dynamic brain network, further obtain the significant dynamic brain network difference between the ASD and the typical control, and be used for identifying the ASD. The application provides a new method for ASD identification analysis.

Description

Technical Field

[0001] This invention relates to an ASD recognition method based on dynamic graph embedding, belonging to the field of computer application technology. Background Technology

[0002] Autism spectrum disorder (ASD) is a neurodevelopmental disorder. Early and accurate identification of ASD helps in effective intervention for patients and reduces the burden on their families and society. Currently, the diagnosis of ASD mainly relies on the experience and knowledge of specialist physicians, which consumes a significant amount of manpower and time. Therefore, research on computer-aided diagnosis of ASD is crucial. Recognizing ASD based on rs-fMRI brain functional connectivity is a commonly used method.

[0003] Reiter et al. ("Performance of machine learning classification models of autism using resting-state fMRI is contingent on sample heterogeneity". Neural Computing and Applications, 2021, 33(8), 3299-3310) used an improved random forest to learn static functional connectivity features associated with ASD and identified ASD. Wang et al. ("Identification of autism based on SVM-RFE and stacked sparse auto-encoder". IEEE Access, 2019, 7, 118030-118036) used the support vector machine recursive feature elimination algorithm to study significant static functional connectivity features and trained a stacked sparse autoencoder with two hidden layers to extract high-dimensional static functional connectivity features to improve the recognition accuracy of ASD. Grana et al. (Impact of machine learning pipeline choices in autism prediction from functional connectivity Data. International Journal of Neural Systems, 2021, 31(4), 2150009) investigated the impact of various feature extraction / selection methods and classifiers on the ability to identify ASD based on static functional connectivity. Chaitra et al. treated static functional connectivity as a static network, extracted features from it using graph theory metrics such as node degree, centrality, and clustering coefficients, and combined it with a support vector machine based on recursive clustering elimination to identify ASD. Similarly, Kazeminejad et al. (The importance of anti-correlations in graph theory based classification of autism spectrum disorder. Frontiers in Neuroscience, 2020, 14, 676) calculated the feature path length, clustering coefficient, and centrality of static brain networks and used a multilayer perceptron to learn deep network features to identify ASD. These studies on identifying ASD based on static functional connectivity neglected the changes in functional connectivity over time.

[0004] Zhao et al. ("Diagnosis of autism spectrum disorder using central-moment features from low-and high-order dynamic resting-state gunctional connectivity networks." Frontiers in Neuroscience, 2020, 14, 258) extracted high-order dynamic correlation features from dynamic functional connectivity based on the idea of ​​"correlation of correlations." They also introduced the central moment method to extract time-invariant features to address the dynamic mismatch in dynamic functional connectivity among participants, and applied it to identify ASD. Liu et al. ("Improved ASD classification using dynamic functional connectivity and multi-task feature selection." Pattern Recognition Letters, 2020, 138, 82-87) used an improved ensemble resilient network and streaming regularized multi-task feature selection method to identify ASD based on dynamic functional connectivity. These methods focus on learning the temporal features of dynamic functional connectivity, but are insufficient in learning spatial features. Cao et al. ("Modeling the dynamic brain network representation for autism spectrum disorder diagnosis." Medical & Biological Engineering & Computing, 2022, 60(7), 1897-1913) viewed dynamic functional connectivity as a dynamic brain network and proposed a graph structure-aware LSTM to learn the spatiotemporal features of dynamic functional connectivity and apply it to ASD identification. However, this method has difficulty simultaneously learning the temporal and spatial features of dynamic functional connectivity, and this limitation affects the ability to identify ASD.

[0005] To date, research on ASD identification methods based on functional connectivity remains an important issue. Improving ASD identification capabilities by better capturing temporal and spatial features from dynamic functional connectivity is a challenge. Summary of the Invention

[0006] To better learn the spatiotemporal features of dynamic functional connections and improve the ability to identify ASD, this invention provides an ASD identification method based on dynamic graph embedding, including:

[0007] Step 1: For the BOLD time series of ASD subjects and typical control subjects, a dynamic brain network was constructed for each subject using the sliding window technique and Pearson correlation.

[0008] Step 2: Calculate the brain region embedding representation of each subject's dynamic brain network one by one using dynamic graph embedding technology;

[0009] Step 3: Based on the brain region embedding representation, cluster all brain regions of each subject's whole brain, and construct a clustered brain network based on the clustering results;

[0010] Step 4: For the connections in the clustered brain network, perform intergroup analysis between the ASD group and the typical control group based on the independent samples t-test and Cohen's D, and select the connections with significant differences;

[0011] Step 5: Based on the significant difference connections selected in Step 4, select features from the cluster brain network as input, and use the trained classification model obtained by training with these features to identify ASD and typical controls; the trained classification model refers to a classification model based on support vector machine.

[0012] Optionally, step 1 includes:

[0013] Step 11: Apply the sliding window technique to the BOLD time series B of a subject. B consists of N time series of length T. B The time series composition, where N is the number of brain regions, is used to segment B using the sliding window technique. There are 1 time window, where w and s are the window length and sliding step size, respectively;

[0014] Step 12: Within the t-th time window, calculate the correlation between brain regions using the Pearson correlation coefficient.

[0015] Step 13: In the t-th time window, using brain regions as nodes, brain region v i With v j The boundary weights between them are used to construct a weighted undirected brain network G. t The brain network is the t-th frame of the dynamic brain network G;

[0016] Step 14: Process each time window according to steps 12 and 13 to obtain a dynamic brain network of the subject. Let V be the set of nodes in G. G is a dynamic network in which the nodes remain unchanged while the edges change over time.

[0017] Optionally, the process of dynamically embedding a subject's dynamic brain network in step 2 includes:

[0018] Step 21: Represent the dynamic brain network G as a set of random walk sequences:

[0019] In the t-th frame G of the dynamic brain network G t Based on the random walk strategy of node2vec guided by parameters p and q, walk_num random walks of length walk_length are performed on each brain region, resulting in walk_num random walk sequences. The set of these random walk sequences is denoted as R. t Perform the same random walk process on each frame of the dynamic brain network G to obtain a set of random walk sequences representing the dynamic brain network G.

[0020] Step 22: Treating brain regions in the random walk sequence as words, and combining the ideas of document embedding and the CBOW model, learn dynamic brain region embedding representations:

[0021] Assume v i ∈V is the selected central brain region to be predicted, from R t In a random walk sequence, a sliding window of fixed length ω is used to move from v i The context brain regions are obtained by sampling from both the left and right sides, and the set of context brain regions is denoted as . Based on document embedding, and considering G t and N t (v i The maximum optimization objective for learning brain region embeddings is: logp(v i |N t (v i ),G t );

[0022] The probability p(v) is calculated using the softmax function. i |N t (v i ),G t Therefore, the maximum optimization objective is written as:

[0023]

[0024] Where Φ is a mapping function that maps G... t and v i Mapped respectively to d-dimensional vectors Φ(G) t ) and Φ(v i We use stochastic gradient descent to update and optimize Φ, thereby maximizing the objective function;

[0025] Over the entire dynamic brain network G, considering the case where all brain regions are selected as central brain regions, the global maximum optimization objective is:

[0026]

[0027] Where, r∈R t Let Φ(v) represent a random walk sequence. After global optimization and updating Φ in G, we get Φ(v). i ) is brain region v i The dynamic embedding vectors, with the embedding vectors of all brain regions denoted as Node. Embed is an N×d dimensional matrix, where each row vector is an embedding representation of the corresponding brain region.

[0028] Optionally, step 3 includes:

[0029] Step 31: For each subject, cluster the dynamic brain region embeddings, grouping similar embedding vectors into the same class and distributing large-disparity embedding vectors into different classes, thus obtaining brain region clusters. The set of clusters is denoted as C = {C1, C2, ..., C}. m ,...,C M}, where M represents the number of clusters;

[0030] Step 32: For each subject, according to C = {C1, C2, ..., C...} m ,...,C M Constructing a convergent brain network G Agg The aggregated brain network G Agg The set of nodes is denoted as V. Agg The set of edges is equal to V; the set of edges is denoted as E. Agg If brain region v i and v j Within the same cluster, the right side between these two brain regions, i.e., for any...

[0031] Step 33: Construct a convergent brain network for all subjects according to steps 31 and 32.

[0032] Optionally, step 4 includes:

[0033] Step 41: Sum the convergent brain networks of the ASD group and the typical control group respectively to construct the convergent brain network G for the ASD group. ASD =(V ASD E ASD The brain-brain network G of the group compared with the typical control group TC =(V TC E TC );

[0034] Step 42: Normalize the two aggregated brain networks by dividing the two groups by the number of subjects in the corresponding groups;

[0035] Step 43: Subtract the normalized typical control group aggregated brain network from the normalized ASD group aggregated brain network to obtain the inter-group difference network; in the difference network, negative edge weights represent underconnections of ASD compared to typical control, and positive edge weights represent overconnections of ASD compared to typical control.

[0036] Step 44: Based on the edge weights of the differential network, pre-select the differential connections (including overconnections and underconnections) in the convergent brain network of the ASD group compared with the typical control group;

[0037] Step 45: Perform an independent samples t-test on the pre-selected differential connections to select those with significant differences;

[0038] Step 46: Perform Cohen's D statistical analysis on the differential connections with significant differences, and select the significant differential connections that pass the effect size test as the final selected significant differential connections.

[0039] Optionally, the training process in step 5 includes:

[0040] Step 51: Based on the significantly different connections selected in Step 4, select the corresponding connections from the clustered brain network of each subject and arrange them to form features for identifying ASD;

[0041] Step 52: Using this feature as input, train an RBF kernel support vector machine classifier, and use a grid search method to evaluate the effectiveness of the classification model with accuracy as the evaluation index to find the optimal support vector machine parameters; the parameters include the penalty coefficient C and the kernel function coefficient gamma;

[0042] Step 53: Set the C and gamma parameters of the support vector machine to the optimal values ​​obtained in step 52, thereby obtaining the trained classification model.

[0043] Optionally, during the dynamic graph embedding process, the hyperparameters walk_length, walk_num, p, q, ω, and d are set to 16, 10, 1, 0.5, 5, and 256, respectively.

[0044] Optionally, the window length and sliding step size are set to w = 20 and s = 2, respectively.

[0045] A second objective of the present invention is to provide an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the ASD identification method based on dynamic graph embedding as described above.

[0046] A third objective of the present invention is to provide a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the ASD identification method based on dynamic graph embedding described above.

[0047] The beneficial effects of this invention are:

[0048] This invention utilizes sliding windows and Pearson correlation to construct dynamic brain networks for ASD and typical control subjects. Dynamic graph embedding is used to learn the embedding representation of the dynamic brain network, simultaneously learning its temporal and spatial characteristics. Based on brain region embeddings, all brain regions of each subject are clustered, and aggregated brain networks are constructed for each subject according to the clustering results. Furthermore, independent samples t-tests and Cohen's D are used for inter-group analysis between the ASD group and the typical control group to obtain significantly different aggregated brain network connectivity features, thus obtaining important features related to ASD recognition, which are then used to identify ASD. The method described in this application can capture the dynamic interaction relationships between brain regions in the dynamic brain network, thereby learning the spatiotemporal characteristics of the dynamic brain network, further obtaining significant dynamic brain network differences between ASD and typical controls, and using them to identify ASD, providing a new method for ASD identification and analysis. In addition, this new method improves the accuracy of ASD identification and reduces the false negative rate of ASD identification. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a visualization of the significant differences in connectivity in the brain of selected ASDs compared to a typical control in an embodiment of the present invention, wherein (a) represents significant underconnection differences and (b) represents significant overconnection differences.

[0051] Figure 2 This is a chord graph of the significant differences in connectivity of the selected ASD compared to a typical control in an embodiment of the present invention, where (a) represents significant underconnection differences and (b) represents significant overconnection differences.

[0052] Figure 3 In this embodiment of the invention, the adjacency matrix diagram of the significantly different connections of the selected ASD compared with the typical control, where (a) represents significant underconnection differences and (b) represents significant overconnection differences. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0054] Example 1:

[0055] This embodiment provides an ASD recognition method based on dynamic graph embedding, the method comprising:

[0056] Step 1: For the BOLD time series of ASD and typical control subjects, a dynamic brain network was constructed for each subject using the sliding window technique and Pearson correlation.

[0057] In practical applications, BOLD time series data for ASD and typical control subjects can be obtained from the open-source ABIDE database or collected from other sources; the size and step size of the sliding window can be determined by the researchers.

[0058] Step 2: Using dynamic graph embedding technology, calculate the brain region embedding representation of the dynamic brain network of the subjects one by one;

[0059] Step 3: Based on the brain region embedding representation, cluster all brain regions of each subject's whole brain and construct a clustered brain network based on the clustering results;

[0060] In practical applications, the clustering method can be determined by researchers;

[0061] Step 4: For the connections in the clustered brain network, perform intergroup analysis between the ASD group and the typical control group based on the independent samples t test and Cohen's D, and select the connections with significant differences.

[0062] In practical applications, the significance value of the independent samples t-test and the threshold of Cohen's D will result in different differential connections; the significance value of the independent samples t-test and the threshold of Cohen's D can be determined by the researchers.

[0063] Step 5: Based on the significant difference connections selected in Step 4, select features from the cluster brain network as input, and use the trained classification model obtained by training with these features to identify ASD and typical controls; the trained classification model refers to a classification model based on support vector machine.

[0064] Example 2:

[0065] This embodiment provides an ASD identification method based on dynamic graph embedding. The method involves selecting data from 871 subjects from the ABIDE database, including 403 ASD subjects and 468 typical control subjects. A dynamic brain network is constructed for each subject using a sliding window and Pearson correlation. The brain region embedding representation of each subject's dynamic brain network is calculated using dynamic graph embedding. The brain region embeddings for each subject are clustered, and a clustered brain network is constructed based on the clustering results. For the connections in the clustered brain network, an independent samples t-test and Cohen's D are used to perform between-group analysis between the ASD group and the qualitative control group, selecting connections with significant differences. Furthermore, features from the clustered brain network are selected as input based on these significantly different connections to train a support vector machine classifier, resulting in a trained classification model used for ASD identification.

[0066] Specifically, it includes the following steps:

[0067] S1: Obtain data from 871 subjects from the ABIDE database, including 403 ASD subjects and 468 typical control subjects. The data for each subject is the BOLD time-series signal corresponding to 200 brain regions. For the BOLD data of each subject, construct a dynamic brain network according to the following procedure.

[0068] (1) Using the sliding window technique, B is composed of N time series B of length T for a subject. B The time series composition, N is 200, and B is segmented using the sliding window technique. There are 3 time windows, where w and s are the window length and sliding step size, respectively, and are set to w = 20 and s = 2;

[0069] (2) Within the t-th time window, the correlation between brain regions is calculated using the Pearson correlation coefficient.

[0070] (3) In the t-th time window, taking the brain region as the node, brain region v i With v j The boundary weights between them are used to construct a weighted undirected brain network G. t The brain network is the t-th frame of the dynamic brain network G;

[0071] (4) Process each time window according to (2) and (3) to obtain a dynamic brain network of the subject. Let V be the set of nodes in G. G is a dynamic network in which the nodes remain unchanged while the edges change over time.

[0072] S2: Using the dynamic graph embedding method, the brain region embedding representation of the dynamic brain network was calculated for each subject;

[0073] The parameters walk_length, walk_num, p, q, ω, and d of the dynamic graph embedding method are set to 16, 10, 1, 0.5, 5, and 256, respectively. The process of dynamic graph embedding of a subject's dynamic brain network includes:

[0074] (1) The dynamic brain network G is represented as a set of random walk sequences:

[0075] In the t-th frame of the dynamic brain network G t Based on the random walk strategy of node2vec guided by parameters p and q, walk_num random walks of length walk_length are performed on each brain region, resulting in walk_num random walk sequences. The set of these random walk sequences is denoted as R. t Perform the same random walk process on each frame of G to obtain a set of random walk sequences representing G.

[0076] (2) Treating brain regions in the random walk sequence as words, and learning dynamic brain region embedding representations by combining the ideas of document embedding and CBOW model:

[0077] Assume v i ∈V is the selected central brain region to be predicted, from R t In a random walk sequence, a sliding window of fixed length ω is used to move from v i The context brain regions are obtained by sampling from both the left and right sides, and the set of context brain regions is denoted as . Based on document embedding, and considering G t and N t (v i The maximum optimization objective for learning brain region embedding representations is:

[0078] logp(v i |N t (v i ),G t )

[0079] The probability p(v) is calculated using the softmax function. i |N t (v i ),G t Therefore, the maximum optimization objective can be written as:

[0080]

[0081] Where Φ is a mapping function that maps G... t and v i Mapped to d-dimensional vectors Φ(G) respectively t ) and Φ(vi We use stochastic gradient descent to update and optimize Φ, thereby maximizing the objective function.

[0082] Over the entire G region, considering all brain regions selected as central brain regions, the global maximum optimization objective is:

[0083]

[0084] Where, r∈R t Let v represent a random walk sequence. After globally optimizing and updating Φ in G, we obtain brain region v. i The dynamic embedding vector Φ(v i The embedding vectors of all brain regions are denoted as Node. Embed is an N×d dimensional matrix, where each row vector is an embedding representation of the corresponding brain region.

[0085] S3: Based on brain region embedding representation, all brain regions of each subject's whole brain are clustered, and a clustered brain network is constructed based on the clustering results;

[0086] (1) For each subject, the Birch clustering method was used to embed nodes into dynamic brain regions. Embed Clustering is performed to group similar embedding vectors into the same class and separate embedding vectors with large differences into different classes, thereby obtaining brain region clusters. The set of clusters is denoted as C = {C1, C2, ..., C6}. m ,...,C M}, where M represents the number of clusters;

[0087] (2) For each subject, according to C = {C1, C2, ..., C...} m ,...,C M Constructing a convergent brain network G Agg G Agg The set of nodes is denoted as V. Agg The set of edges is equal to V; the set of edges is denoted as E. Agg If brain region v i and v j Within the same cluster, the right side between these two brain regions, i.e., for any...

[0088] (3) Convergent brain networks were constructed for all subjects as described in (1) and (2).

[0089] S4: For the connections in the clustered brain network, the intergroup analysis between the ASD group and the typical control group was carried out according to the independent samples t test and Cohen's D, and the connections with significant differences were selected.

[0090] (1) The convergent brain networks of the ASD group and the typical control group were added together to construct the convergent brain network G for the ASD group. ASD =(V ASD E ASD The brain-brain network G of the group compared with the typical control TC =(V TC E TC );

[0091] (2) Normalize the two aggregated brain networks by dividing the two groups by the number of subjects in the corresponding groups;

[0092] (3) Subtract the normalized typical control group aggregated brain network from the normalized ASD group aggregated brain network to obtain the intergroup difference network; in the difference network, negative edge weights represent the underconnection of ASD compared to the typical control, and positive edge weights represent the overconnection of ASD compared to the typical control.

[0093] (4) Based on the edge weights of the differential network, pre-select the parts of the convergent brain network of the ASD group that are significantly different from the typical control group. The selected differential connections include 200 overconnections and 200 underconnections.

[0094] (5) An independent samples t-test with a significance value of 0.05 was performed on the pre-selected differential connections to select 361 differential connections with significant differences.

[0095] (6) For the differential connections with significant differences, Cohen's D statistical analysis was performed with an effect size threshold of 0.2. 93 significant differential connections that passed the effect size test were selected as the final significant differential connections. Among them, there were 43 underconnections and 50 overconnections, which are shown in Table 1 and Table 2, respectively.

[0096] Figure 1 In this study, BrainNet was used to visualize the selected significantly differentially connected and overconnected components in the brain for easy observation.

[0097] exist Figure 2 The chord diagrams of the selected significant differences between underconnected and overconnected systems are presented in the figure. Figure 2 In (a), brain regions 21, 28, 140, 47 and 96 are observed to be involved in many underconnections, indicating that these regions in ASD are significantly lacking in dynamic interactions. Figure 2 In (b), brain regions 111, 21, and 159 are observed to be involved in many overconnections, indicating that these regions in ASD have significant over-dynamic interactions.

[0098] Figure 3 The adjacency matrix diagram of the network composed of selected significantly different underconnected and overconnected networks is displayed. From Figure 3Some brain regions exhibiting under- or over-dynamic interactions can be observed.

[0099] Table 1 shows the significant differences in ASD compared to the typical control group.

[0100]

[0101]

[0102] Table 2 shows the significant differences in ASD compared to the typical control group.

[0103]

[0104]

[0105] S5: Based on the significantly different connections selected in S4, select the corresponding connections from the cluster brain network of each subject and arrange them to form features for identifying ASD; use these features as input to train a support vector machine classifier with RBF kernel. During the training process, use the grid search method to find the optimal penalty coefficient C and kernel function coefficient gamma. The values ​​of C and gamma are in [0.001, 0.01, 0.1, 1, 10, 100].

[0106] The classification model trained using the above process was used to identify ASD. The performance of the model, evaluated using 10-fold cross-validation, is shown in Table 3.

[0107] Table 3 Classification Model

[0108]

[0109] Among them, ACC, SEN, SPE, AUC, PRE, and F1 represent accuracy, sensitivity (also called true positive rate (TPR)), specificity, area under the ROC curve, and F1 score, respectively, and are indicators for measuring the classification performance of the model. These evaluation indicators are calculated based on four values: true positive (TP), false positive (FP), true negative (TN), and false negative (FN). TP represents the number of people who are actually ASD patients and are predicted to be ASD patients; FP represents the number of people who are actually typical controls and are predicted to be ASD patients; TN represents the number of people who are actually typical controls and are predicted to be typical controls; and FN represents the number of people who are actually ASD patients and are predicted to be TC.

[0110] Table 4 Evaluation metrics for model classification performance

[0111]

[0112] The evaluation metrics are defined as shown in Table 4. A higher ACC indicates higher accuracy in identifying ASD, a higher SEN indicates a lower rate of missed diagnoses of ASD, a higher SPE indicates a lower rate of false diagnoses of ASD, higher PRE and F1 indicate higher accuracy in identifying ASD, and a higher AUC indicates a more reliable classification model.

[0113] Existing methods for identifying ASD based on static brain functional connectivity generally have an ACC and AUC of around 65%, and a low SEN value, usually not exceeding 65%, resulting in a high false negative rate. Existing methods for identifying ASD based on dynamic brain functional connectivity have improved ACC and AUC, but most are close to 70%, and still suffer from a high false negative rate. As shown in the evaluation data in Table 3, the ASD identification method of this invention achieves an ACC and AUC of 78% and a SEN of 72%. Compared to existing ASD identification methods based on static / dynamic functional connectivity, this method demonstrates improved accuracy and precision, with a lower false negative rate, thus improving the overall accuracy and reliability of ASD identification.

[0114] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0115] 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 principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An ASD recognition method based on dynamic graph embedding, characterized in that, The method includes: Step 1: For the BOLD time series of ASD subjects and typical control subjects, a dynamic brain network was constructed for each subject using the sliding window technique and Pearson correlation. Step 2: Calculate the brain region embedding representation of each subject's dynamic brain network one by one using dynamic graph embedding technology; Step 3: Based on the brain region embedding representation, cluster all brain regions of each subject's whole brain, and construct a clustered brain network based on the clustering results; Step 4: For the connections in the clustered brain network, perform intergroup analysis between the ASD group and the typical control group based on the independent samples t-test and Cohen's D, and select the connections with significant differences; Step 5: Based on the significantly different connections selected in Step 4, select features from the clustered brain network as input, and use the trained classification model obtained by training with these features to identify ASD and typical controls; the trained classification model refers to a classification model based on support vector machines. Step 1 includes: Step 11: Analyze the BOLD time series of a subject. Using sliding window technology, Depend on The length of the strip is The time series composition, To determine the number of brain regions, a sliding window technique is used. Divided into There are several time windows, among which and These are the window length and the sliding step size, respectively. Step 12: In the Within a time window, the correlation between brain regions was calculated using the Pearson correlation coefficient. ; Step 13: In the Within a time window, with brain regions as nodes brain region and By considering the boundary weights between them, a weighted undirected brain network is constructed. This brain network serves as a dynamic brain network. No. frame; Step 14: Process each time window according to steps 12 and 13 to obtain a dynamic brain network of the subject. ,remember The set of nodes is , It is a dynamic network in which nodes remain unchanged while edges change over time; The process of dynamic graph embedding of a subject's dynamic brain network in step 2 includes: Step 21: Connect the dynamic brain network Represented as a set of random walk sequences: In the dynamic brain network The t-th frame Above, according to the random walk strategy of node2vec guided by parameters p and q, it performs a random walk on each brain region. The second length is The random walk yields There are n random walk sequences, and the set of these random walk sequences is denoted as . ; for the dynamic brain network Each frame undergoes the same random walk process to obtain the description of the dynamic brain network. The set of random walk sequences ; Step 22: Treating brain regions in the random walk sequence as words, and combining the ideas of document embedding and the CBOW model, learn dynamic brain region embedding representations: Assumption It is the selected central brain region to be predicted, from In a random walk sequence, a fixed length of The sliding window from The context brain regions are obtained by sampling from both the left and right sides, and the set of context brain regions is denoted as . Based on document embedding, and also considering and The maximum optimization objective for learning brain region embedding representations is: ; Use the softmax function to calculate the probability. Therefore, the maximum optimization objective should be written as follows: in, It is a mapping function that will and Mapped to d-dimensional vectors respectively and Use stochastic gradient descent to update and optimize This maximizes the objective function; Throughout the entire dynamic brain network Considering the case where all brain regions are selected as central brain regions, the global maximum optimization objective is: in, Describes a random walk sequence, in Perform global optimization and update After that, what was obtained It is a brain region The dynamic embedding vectors, denoted as the embedding vectors of all brain regions. It is A dimensional matrix, where each row vector is an embedding representation of the corresponding brain region.

2. The method according to claim 1, characterized in that, Step 3 includes: Step 31: For each subject, cluster the dynamic brain region embeddings, grouping similar embedding vectors into the same class and distributing large-disparity embedding vectors into different classes, thus obtaining brain region clusters. The set of clusters is denoted as . , Indicates the number of clusters; Step 32: For each subject, according to Building a convergent brain network The aggregated brain network The set of nodes is denoted as ,and Equal; the set of edges is denoted as If brain region and Within the same cluster, the right side between these two brain regions, i.e., for any... , ; Step 33: Construct a convergent brain network for all subjects according to steps 31 and 32.

3. The method according to claim 1, characterized in that, Step 4 includes: Step 41: Sum the convergent brain networks of the ASD group and the typical control group separately to construct the convergent brain networks of the ASD group. Brain-gathering network compared to typical control ; Step 42: Normalize the two aggregated brain networks by dividing the two groups by the number of subjects in the corresponding groups; Step 43: Subtract the normalized typical control group aggregated brain network from the normalized ASD group aggregated brain network to obtain the inter-group difference network; in the difference network, negative edge weights represent underconnections of ASD compared to typical control, and positive edge weights represent overconnections of ASD compared to typical control. Step 44: Based on the edge weights of the differential network, pre-select the differential connections in the aggregated brain network of the ASD group compared with the typical control group. The differential connections include overconnections and underconnections. Step 45: Perform an independent samples t-test on the pre-selected differential connections to select those with significant differences; Step 46: Perform Cohen's D statistical analysis on the differential connections with significant differences, and select the significant differential connections that pass the effect size test as the final selected significant differential connections.

4. The method according to claim 1, characterized in that, The training process in step 5 includes: Step 51: Based on the significantly different connections selected in Step 4, select the corresponding connections from the clustered brain network of each subject and arrange them to form features for identifying ASD; Step 52: Using this feature as input, train an RBF kernel support vector machine classifier, and use a grid search method to evaluate the effectiveness of the classification model with accuracy as the evaluation index to find the optimal support vector machine parameters; the parameters include the penalty coefficient C and the kernel function coefficient gamma; Step 53: Set the C and gamma parameters of the support vector machine to the optimal values ​​obtained in step 52, thereby obtaining the trained classification model.

5. The method according to claim 1, characterized in that, Hyperparameters during dynamic graph embedding , p, q d is set to 16, 10, 1, 0.5, 5, 256 respectively.

6. The method according to claim 1, characterized in that, The window length and sliding step size are respectively set as follows: =20 and =2.

7. An electronic device, comprising: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the ASD identification method based on dynamic graph embedding as described in any one of claims 1-6.

8. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the ASD identification method based on dynamic graph embedding as described in any one of claims 1-6.

Images