Autism auxiliary diagnosis method based on electroencephalogram micro-state analysis and graph convolution network

Abstract

The invention discloses an autism auxiliary diagnosis method based on electroencephalogram micro-state analysis and a graph convolution network, and the method comprises the following steps: S1, carrying out the model training stage: S1, carrying out the preprocessing of data in an electroencephalogram data set, and obtaining an electroencephalogram signal with a higher signal-to-noise ratio; s2, performing micro-state analysis on the preprocessed electroencephalogram signals, and constructing the electroencephalogram signals into a micro-state conversion mode pattern; s3, constructing and initializing a multi-scale feature extraction model, and inputting processed training data; and S4, extracting graph-level embedding as discriminant features of an autism identification task, and using a cross entropy loss optimization model to obtain a finally trained feature extraction model and classifier. S5, inputting the micro-state transition probability graph of the test data into the trained model to obtain an autism prediction result; and S6, calculating related indexes, and evaluating the performance of the model.

Description

Technical Field

[0001] The present invention belongs to the field of medical information processing and autism assisted diagnosis, and particularly relates to an autism assisted diagnosis method based on electroencephalogram microstate analysis and graph convolutional network. Background Art

[0002] In the field of neuroscience, graph convolutional neural networks are usually used to explore the brain functional connectivity patterns. By constructing functional connectivity networks between different brain regions through neuroimaging techniques such as electroencephalogram, functional magnetic resonance imaging or magnetoencephalogram, and using graph convolutional-based electroencephalogram modeling methods to learn brain topological information. However, most previous studies only focused on the static functional connectivity patterns of the brain, ignoring the dynamic information of the brain. In this regard, this study constructs a brain network based on electroencephalogram microstate analysis, and proposes a multi-scale graph convolutional model based on electroencephalogram microstate transition probability to reveal the dynamic changes of the brain during the microstate transition process, so as to achieve more efficient autism diagnosis. Summary of the Invention

[0003] The present invention proposes an autism assisted diagnosis method based on electroencephalogram microstate analysis and graph convolutional network to solve the problems mentioned in the above background art. The technical solution steps of the present invention are as follows:

[0004] Step 1: Preprocess the data in the self-collected electroencephalogram dataset (EEG Dataset for ASD Identification, EDAI) to obtain electroencephalogram signals with higher signal-to-noise ratio;

[0005] Step 1.1: Filter the original electroencephalogram signals using a band-pass filter of 0.5 - 50 Hz to remove meaningless frequency components and retain information related to brain activities;

[0006] Step 1.2: Adopt an interpolation method to fit the data around the electroencephalogram bad channels to estimate missing or damaged data, so as to minimize data distortion.

[0007] Step 1.3: Use independent component analysis technology to identify and extract source signals related to brain activities, and remove artifacts in the electroencephalogram signals.

[0008] Step 2: Conduct microstate analysis on the preprocessed electroencephalogram signals, and construct a state transition pattern graph by extracting the transition probability features of electroencephalogram microstates.

[0009] Step 2.1: Aggregate the electroencephalogram data of all subjects to generate electroencephalogram data for evaluating microstate prototypes, which is referred to as template data in this article;

[0010] Step 2.2: Extract the topographic maps of each sampling point in the template data, and use the clustering algorithm to perform statistical analysis on the topographic maps, and cluster them into 5 representative and interpretable topographic maps, called microstate prototypes;

[0011] Step 2.3: Perform inverse fitting on each EEG sample based on the microstate prototypes, and convert the original EEG signal into a microstate sequence in which 5 microstates alternate;

[0012] Step 2.4: Calculate the occurrence frequency, time coverage rate, average duration and microstate transition probability of each type of microstate;

[0013] Step 2.5: According to the three microstate time characteristics and microstate transition probability described in Step 2.4, construct the microstate transition mode into graph structure data to obtain the microstate transition probability map.

[0014] Step 2.5.1: Convert the original EEG signal into a microstate sequence according to the microstate selection. The original EEG signal is represented as a form in which five microstate prototypes appear alternately. By statistical analysis, the switching probability P(i,j) between different microstates is obtained. Then, the adjacency matrix A of the graph can be calculated as follows:

[0015]

[0016] where T(i,j) represents the number of times a microstate transitions to another microstate;

[0017] Step 2.5.2: Extract the activation intensity, spatial structure and time attribute features from the microstate sequence, including the GEV, GFP, occurrence frequency, time coverage rate and average duration of each type of microstate. Take each type of microstate as the node V of the graph, and the activation intensity, spatial structure and time attribute features of the microstate sequence as the node feature x i :

[0018] V = {v1, v2, …, v n} = {m1, m2, …, m5} (2)

[0019]

[0020] where m1, m2, …, m5 represent 5 types of microstates respectively, and GEV i , GFP i , F i , TC i , are the whole-brain domain power, global explained variance, occurrence frequency, time coverage rate and average duration of each type of microstate respectively;

[0021] Step 2.5.3: Each type of microstate serves as a node of the graph, and the activation intensity, spatial structure, and temporal attribute features of the microstate sequence are used as node features. The edges of the graph are constructed through the adjacency matrix obtained from the microstate transition probability matrix, thereby obtaining the microstate transition probability graph G:

[0022] G = (V, E, A) (4)

[0023] The node features in the microstate transition probability graph G can be represented as a matrix X:

[0024] X = [x1, x2, …, x n T (5)

[0025] where, x i is the feature representation of node v i .

[0026] Step 3: Design and build a deep learning model for autism recognition, input the microstate transition pattern graph into the initialized model, calculate the optimization loss, and update the model parameters for supervised training;

[0027] Step 3.1: Construct a multi-scale feature extraction model to achieve multi-scale representation of the nodes in the graph through multiple layers of graph convolution;

[0028] Step 3.1.1: The first convolution module consists of one layer of graph convolution, and this module is represented by a formula, as shown in formula (9):

[0029]

[0030] where, H1 is a layer of activation matrix, is the adjacency matrix of the graph, W0 is the initialized weight matrix, σ(·) is the activation function, such as the ReLu activation function, and the output f1(X, A) of the first layer can be expressed as:

[0031]

[0032] Step 3.1.2: The second convolution module consists of two layers of graph convolution, and the output f2(X, A) of the second convolution module can be expressed as:

[0033]

[0034] Step 3.1.3: The third convolution module consists of three layers of graph convolution, and the output f3(X, A) of the third convolution module can be expressed as:

[0035]

[0036] where, W l is the weight matrix of the l-th hidden layer;​

[0037] Step 3.1.4: The outputs of the three different-scale graph convolution modules are concatenated, and through the initial residual connection, the initial graph features and the multi-scale features are added together. To ensure that the dimensions of the initial graph features and the multi-scale features are consistent, before performing the residual connection, a fully connected layer is used to adjust the dimensions of the initial graph features to match the output of the multi-scale features. The final output f(X,A):

[0038]

[0039] where α is a hyperparameter representing the proportion retained from the input layer, which is set to 0.2, is the initial graph feature after passing through the fully connected layer.

[0040] Step 3.2: Introduce self-attention graph pooling to extract discriminative node features and spatially model the EEG microstate transition patterns;

[0041] Step 3.3: Aggregate the node embeddings through feed-forward attention to obtain the graph-level embedding, and use it as the discriminative feature for autism diagnosis to achieve effective identification of autism. The feed-forward attention mechanism inputs the node feature vector h n into the activation function a(·) to generate the weight vector. Calculate the adaptive weighted average of the node vector h n to produce the graph-level embedding vector g, and the process can be expressed as:

[0042] e n = a(h n ) (6)

[0043]

[0044] where the function a(·) is a learnable function that depends on the node feature vector h n , and g is the graph-level embedding that aggregates the node feature vectors. α n is the attention weight, reflecting the importance of each node n when constructing the graph-level representation of g. Finally, construct an MLP classifier to achieve autism recognition.

[0045] Step 4: Use the trained model as the autism recognition model, input the microstate transition pattern graph of the test data into the model for prediction, obtain the prediction results of autism recognition, and calculate the relevant evaluation metrics.

[0046] Step 4.1: Input the microstate transition probability graph of the EEG test data set into the trained model;

[0047] Step 4.2: Use the same graph convolutional neural network-based feature extraction module, self-attention graph pooling module, and graph-level embedding module as in the training phase to sequentially extract the dynamic information of the brain topology;

[0048] Step 4.3: Use the trained multi-layer perceptron network for autism recognition, compare it with the actual disease situation, and calculate performance metrics to evaluate the prediction performance of the model. Brief Description of the Drawings

[0049] Figure 1 It is a flowchart of the autism-assisted diagnosis method based on electroencephalogram microstate analysis and graph convolutional network of the present invention;

[0050] Figure 2 It is a model framework diagram of the autism-assisted diagnosis method based on electroencephalogram microstate analysis and graph convolutional network of the present invention;

[0051] Figure 3 It is an architecture diagram of the multi-scale feature extraction model of the present invention; Figure 4 It is a schematic diagram of the model training and evaluation process of the autism-assisted diagnosis method based on electroencephalogram microstate analysis and graph convolutional network of the present invention. Detailed Embodiments

[0052] The following further elaborates on the present invention in conjunction with preferred embodiments. More details are set forth in the following description for a thorough understanding of the present invention. However, the present invention is clearly capable of being implemented in many other ways different from this description. Those skilled in the art can make similar generalizations and deductions according to the actual application situation without departing from the connotation of the present invention. Therefore, the protection scope of the present invention should not be limited by the content of this specific embodiment.

[0053] A flowchart of an autism-assisted diagnosis method based on electroencephalogram microstate analysis and graph convolutional network, as Figure 1 shown, includes the following steps:

[0054] S1: Preprocess the data in the electroencephalogram dataset. After band-pass filtering, interpolation of bad leads, and independent component analysis, obtain clean electroencephalogram signals without artifacts;

[0055] S2: Perform microstate analysis on the preprocessed electroencephalogram signals and calculate microstate parameters;

[0056] S3: Construct a microstate transition pattern graph. Each type of microstate serves as a node of the graph, and the activation intensity, spatial structure, and temporal attribute features of the microstate sequence serve as node features. Construct the edges of the graph through the adjacency matrix obtained from the microstate transition probability matrix;

[0057] S4: Construct a multi-scale feature extraction module, which consists of three convolutional modules with different scales. Each module is stacked by different numbers of graph convolutional layers, and each layer of convolutional layer is connected by a ReLu activation layer. By stacking different numbers of graph convolutional layers, the first convolutional module focuses on local information, while the subsequent modules gradually expand the receptive field and extend the focus to global information;

[0058] S5: Calculate the attention weights between nodes, sort according to the attention scores of nodes, select the top k nodes with the highest scores to form a mask, and retain the features of the specified nodes and the topological information between these nodes according to the mask;

[0059] S6: Use graph-level embedding classification to discriminate the microstate transition probability graph features. The graph-level embedding captures the global graph topology and node attributes by mapping the nodes and edges of the graph to a low-dimensional vector space. This graph-level embedding is fed into a multi-layer perceptron as a feature vector for the final autism recognition;

[0060] S7: Perform microstate analysis on the EEG signals of the test dataset and construct a microstate transition pattern graph;

[0061] S8: Input the microstate transition pattern graph of the test data into the trained autism recognition model to obtain the prediction result;

[0062] S9: Calculate relevant metrics to evaluate the model performance.

[0063] The model framework diagram of an autism-assisted diagnosis method based on EEG microstate analysis and graph convolutional network is as Figure 2 shown:

[0064] This model framework can be divided into four parts, including EEG data preprocessing, microstate analysis, constructing a microstate transition pattern graph, and multi-scale feature extraction and autism recognition.

[0065] EEG data preprocessing: The processing techniques include band-pass filtering, interpolating bad channels, and independent component analysis;

[0066] Microstate analysis: Perform microstate evaluation on the preprocessed EEG data to obtain 5 types of microstate prototypes, and perform inverse fitting on the EEG samples to convert the original EEG signals into a microstate sequence that alternates between different microstates;

[0067] Construct a microstate transition probability graph: Regard the microstate transition probability as a graph with topological structure, where the microstates are the nodes of the graph, and the transition probability between microstates is the connection strength between the nodes in the graph. By extracting the transition probability features of EEG microstates, construct them in the form of a graph.

[0068] Scale Feature Extraction and Autism Recognition: To better learn the dynamic information of this topological structure, a feature extraction and autism diagnosis model based on a multi-scale graph convolutional neural network is constructed. It mainly includes the following steps: First, a multi-scale feature extraction model is constructed to achieve multi-scale representation of nodes in the graph through multiple layers of graph convolution; Second, self-attention graph pooling is introduced to extract discriminative node features and spatially model the electroencephalogram microstate transformation pattern; Finally, the extracted graph-level embedding is used as a discriminative feature for autism diagnosis to achieve effective autism recognition.

[0069] Among them, the architecture diagram of the multi-scale feature extraction model is as Figure 3 shown:

[0070] This model is composed of three convolutional modules with different scales. Each module is stacked by different numbers of graph convolutional layers, and each layer of convolutional layer is connected by a ReLu activation layer. The multi-level graph convolutional module aims to learn the feature representation of nodes at different abstraction levels. The role of each graph convolutional module is to capture and extract graph representations at different scales. By stacking different numbers of graph convolutional layers, the first convolutional module focuses on local information, while the subsequent modules gradually expand the receptive field and extend the focus to global information. Although increasing the number of hidden layers can increase the expressive power and complexity of the neural network, too many hidden layers will also lead to problems such as overfitting and increased training time. To solve the problem of possible network degradation caused by stacking multiple layers of graph convolution, an initial residual connection is added to the model, so that the network not only makes full use of multi-scale information, but also promotes information transmission and gradient flow through the residual connection, improving the multi-scale spatial modeling ability of the model.

[0071] The above is only the preferred solution of the present invention and is not a further limitation of the present invention. All equivalent changes made by using the content of the specification and drawings of the present invention are within the protection scope of the present invention.

Claims

1. An autism-assisted diagnosis method based on electroencephalogram (EEG) microstate analysis and graph convolutional network, the specific steps are as follows: Step 1: Preprocess the data in the self-collected EEG dataset (EEG Dataset for ASD Identification, EDAI) to obtain EEG signals with a higher signal-to-noise ratio; Step 2: Perform microstate analysis on the preprocessed EEG signals, and by extracting the transition probability features of EEG microstates, construct them into a state transition pattern graph. Step 3: Build and train a deep learning model for autism identification. Input the microstate transition pattern graph into the initialized model, calculate the optimization loss, and update the model parameters for supervised training; Step 4: Use the trained model as an autism identification model. Input the microstate transition pattern graph of the test data into this model for prediction, obtain the prediction results of autism identification, and calculate relevant evaluation metrics.

2. According to the autism-assisted diagnosis method based on EEG microstate analysis and graph convolutional network described in claim 1, the sub-steps of preprocessing the data in the EEG dataset in step 1 are: Step 1.1: Filter the original EEG signals using a band-pass filter of 0.5 - 50 Hz to remove meaningless frequency components and retain the information related to brain activities; Step 1.2: Adopt an interpolation method to fit the data around the EEG bad channels to estimate the missing or damaged data, thereby minimizing data distortion. Step 1.3: Use independent component analysis technology to identify and extract the source signals related to brain activities, and remove the artifacts in the EEG signals.

3. According to the autism-assisted diagnosis method based on EEG microstate analysis and graph convolutional network described in claim 1, the sub-steps of EEG microstate analysis and constructing the microstate transition pattern graph in step 2 are: Step 2.1: Aggregate the EEG data of all subjects to generate EEG data for evaluating microstate prototypes, which is referred to as template data in this article; Step 2.2: Extract the topographic maps of each sampling point in the template data, and use a clustering algorithm to perform statistical analysis on the topographic maps, and cluster them into 5 representative and interpretable topographic maps, called microstate prototypes; Step 2.3: Based on the microstate prototypes, perform backward fitting on each EEG sample to convert the original EEG signals into a microstate sequence of alternating 5 microstates; Step 2.4: Calculate the occurrence frequency, time coverage rate, average duration, and microstate transition probability of each type of microstate; Step 2.5: Convert the original EEG signals into a microstate sequence according to the microstate selection. The original EEG signals are represented in the form of alternating 5 microstate prototypes. By statistical analysis, the switching probability P(i,j) between different microstates is obtained, and the adjacency matrix A of the graph can be calculated as: Among them, T(i,j) represents the number of times a microstate transfers to another microstate; Step 2.6: Extract activation intensity, spatial structure, and temporal attribute features from the microstate sequence, including the GEV, GFP, occurrence frequency, temporal coverage, and average duration of each type of microstate. Consider each type of microstate as a node V of the graph, and the activation intensity, spatial structure, and temporal attribute features of the microstate sequence as node features x i : V = {v1, v2, …, v n} = {m1, m2, …, m5} (2) Among them, m1, m2, …, m5 respectively represent five types of microstates, GEV i , GFP i , F i , TC i , are respectively the whole-brain domain power, global explained variance, occurrence frequency, time coverage rate, and average duration of each type of microstate; Step 2.7: Each type of microstate serves as a node of the graph, and the activation intensity, spatial structure, and temporal attribute features of the microstate sequence serve as node features. The edges of the graph are constructed using the adjacency matrix obtained from the microstate transition probability matrix, thereby obtaining the microstate transition probability graph G: G = (V, E, A) (4) The node features in the microstate transition probability graph G can be represented as a matrix X: X = [x1, x2, …, x n T (5)​ Among them, x i is the feature representation of node v i .

4. For an autism-assisted diagnosis method based on electroencephalogram microstate analysis and graph convolutional network according to claim 1, the sub-steps of designing and building a deep learning model for autism recognition in the said step 3 are as follows: Step 3.1: Construct a multi-scale feature extraction model to achieve multi-scale representation of the nodes in the graph through multiple layers of graph convolution; Step 3.2: Introduce self-attention graph pooling to extract discriminative node features and perform spatial modeling on the electroencephalogram microstate transformation pattern; Step 3.3: Aggregate node embeddings through feed-forward attention to obtain graph-level embeddings, and use them as discriminative features for autism diagnosis to achieve effective identification of autism. The feed-forward attention mechanism inputs the node feature vector h n into the activation function a(·) to generate a weight vector. Calculate the adaptive weighted average of the node vector h n to produce the graph-level embedding vector g, and the process can be expressed as: e n = a(h n ) (6) Among them, The function a(·) is a learnable function that depends on the node feature vector h n , where g is the graph-level embedding that aggregates the node feature vectors. α n is the attention weight, which reflects the importance of each node n when constructing the graph-level representation of g. Finally, an MLP classifier is constructed to achieve autism recognition.

5. For the deep learning model built and used for autism recognition according to claim 5, the sub-steps of constructing a multi-scale feature extraction model in the said step 3.1 are as follows: Step 3.1.1: The first convolutional module consists of one layer of graph convolution, and this module is represented by a formula, as shown in formula (9): Among them, H1 is a layer of activation matrix, is the adjacency matrix of the graph, W0 is the initialized weight matrix, σ(·) is the activation function, such as the ReLu activation function, and the output f1(X,A) of the first layer can be expressed as: Step 3.1.2: The second convolutional module consists of two layers of graph convolution, and the output f2(X, A) of the second convolutional module can be expressed as: Step 3.1.3: The third convolutional module consists of three layers of graph convolution, and the output f3(X, A) of the third convolutional module can be expressed as: where, W l is the weight matrix of the l-th hidden layer; Step 3.1.4: The outputs of the three different-scale graph convolution modules are concatenated, and through an initial residual connection, the initial graph features and multi-scale features are added together. To ensure that the dimensions of the initial graph features and multi-scale features are consistent, before performing the residual connection, a fully connected layer is used to adjust the dimensions of the initial graph features to match the output of the multi-scale features. The final output f(X, A): Among them, α is a hyperparameter representing the proportion retained from the input layer, which is set to 0.

2. It is the initial graph feature after full connection.

6. For an autism-assisted diagnosis method based on electroencephalogram microstate analysis and graph convolutional network according to claim 1, the sub-steps of using the trained model as an autism recognition model for testing and performance evaluation in the said step 4 are as follows: Step 4.1: Input the microstate transition probability graph of the electroencephalogram test data set into the trained model; Step 4.2: Use the same feature extraction module based on graph convolutional neural network, self-attention graph pooling module, and graph-level embedding module as in the training stage to sequentially extract the dynamic information of the brain topology; Step 4.3: Use the trained multi-layer perceptron network for autism recognition, compare it with the actual disease situation, and calculate performance metrics to evaluate the prediction performance of the model.

Images