A system for classifying autism based on resting-state electroencephalogram signals

Abstract

The application relates to the technical field of biomedical information processing, in particular to an autism classification system based on resting-state electroencephalogram signals, which comprises an EEG signal acquisition and preprocessing unit and an EEG signal classification unit, the EEG signal classification unit is used for classification by using a trained Rest-HGCN network model; the Rest-HGCN network model comprises a resting-state mixed graph network module, an attention learning module and a classification module; the resting-state mixed graph network module comprises a cognitive graph branch and a data-driven graph branch and is used for extracting corresponding feature mappings; the attention learning module is used for fusing the feature mappings extracted by the resting-state mixed graph network module to obtain a final feature mapping; and the classification module is used for classifying the final feature mapping to obtain a classification result. Through the classification system, the problems of ASD patient EEG feature extraction difficulty and low recognition rate in the prior art can be effectively solved, and only a small amount of features are needed to achieve the purpose of more efficient ASD classification recognition.

Description

Technical Field

[0001] This invention relates to the field of biomedical information processing technology, and in particular to an autism classification system based on resting-state electroencephalogram (EEG) signals. Background Technology

[0002] Autistic Spectrum Disorder (ASD) is the most common and representative pervasive neurodevelopmental disorder, primarily beginning in infancy and early childhood. Its clinical manifestations include social impairment, limited interests, and repetitive, stereotyped behaviors. Currently, researchers are attempting to identify ASD early using objective indicators from magnetic resonance imaging (MRI) and computed tomography (CT). However, due to the lack of precise correlation between neurological examination results and the clinical or pathophysiological characteristics of ASD, the use of these imaging techniques for identifying children with ASD is not currently recommended. Previous studies have found that the functional complexity of electroencephalography (EEG) signals in children with ASD decreases with brain development, and that children with ASD exhibit significantly different EEG patterns in the right and middle regions of the brain compared to normal control children. These studies indicate that the EEG activation patterns of patients with ASD differ from those of non-ASD patients and can be reflected in resting or task-oriented EEG. Compared to MRI and CT, EEG offers unique advantages such as higher temporal resolution, ease of operation, convenient acquisition, and lower cost, and has been proven to provide more objective electrophysiological indicators for identifying ASD. Furthermore, EEG examinations have a wider range of applications compared to MRI and CT, fewer restrictions on the examinee, and are suitable for a wider age range.

[0003] In the prior art, a Chinese invention patent with publication number CN112190261A and publication date of January 8, 2021 was proposed. The technical solution is to use machine learning methods to input features into a classifier to achieve pattern classification. However, this technical solution will result in poor classification performance.

[0004] With the rapid development of deep learning (DL) technology, researchers have attempted to apply DL techniques to ASD recognition, with convolutional neural networks (CNNs) and recurrent neural networks (RNNs) being widely used. Currently, most studies employ DL technology to automatically extract features from functional magnetic resonance imaging (fMRI) images for ASD recognition. While achieving good recognition performance, this approach suffers from high implementation costs and limited scalability. Benefiting from lower acquisition costs and better implementation results, ASD recognition models based on EEG signals and DL technology have gradually been proposed and applied to ASD-assisted recognition systems. Currently, most studies use CNNs and RNNs. Although these models improve ASD recognition performance to some extent, they do not effectively mine the latent graph features between different EEG channels. Furthermore, while CNNs can be used for multi-channel feature extraction, they cannot describe the complex functional relationships between channels. These limitations lead to poor recognition performance of CNN- or RNN-based ASD recognition models. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes an autism classification system based on resting-state electroencephalogram (EEG) signals. This system effectively solves the problems of difficulty in extracting EEG features from ASD patients and low recognition rates in existing technologies, requiring only a small number of features to achieve more efficient ASD classification and recognition.

[0006] This invention is achieved by adopting the following technical solution:

[0007] An autism classification system based on resting-state electroencephalogram (EEG) signals, comprising:

[0008] The EEG signal acquisition and preprocessing unit is used to acquire EEG signals and preprocess the acquired EEG signals to calculate differential entropy features and PLV network features.

[0009] The EEG signal classification unit is used for classification using a trained Rest-HGCN network model; the Rest-HGCN network model includes a resting-state mixture graph network module, an attention learning module, and a classification module.

[0010] The resting-state hybrid graph network module includes a cognitive graph branch and a data-driven graph branch; the cognitive graph branch is used to extract the corresponding feature map Z based on the input PLV network features and differential entropy features. c The data-driven graph branch is used to extract the corresponding feature map Z based on the input differential entropy features. d;

[0011] The feature map Z c and feature map Z d They are respectively:

[0012]

[0013]

[0014] Where ReLU is the activation function, θ fk Let X be a learnable parameter and X be a differential entropy feature. These are the scaled Laplace matrices, respectively;

[0015] The attention learning module is used to map the feature Z... c and feature map Z d Perform fusion to obtain the final feature map Z:

[0016] Z = ω c .Z c +ω d .Z d

[0017] Where, ω c ,ω d ∈R n*n , indicating that it is assigned to the feature map Z c and feature map Z d The attention values ​​of N nodes;

[0018] The classification module is used to classify the final feature map Z and obtain the classification result.

[0019] The feature map Z c The extraction method includes the following steps:

[0020] Step 1: Define the initial PLV network characteristics as a graph adjacency matrix Ac, and calculate the regularized Laplacian matrix L of the graph adjacency matrix Ac. c :

[0021]

[0022] Among them, I N Let D be the identity matrix and L be the length of the matrix. c The degree matrix;

[0023] Step 2: Calculate the regularized Laplacian matrix L c proportionally scaled Laplace matrix

[0024]

[0025] Where, λ max L is the regularized Laplace matrix c The maximum value of the eigenvalue;

[0026] Step 3: Scale the Laplacian matrix Convolution operation with the differential entropy feature input map, and according to spectral graph filtering theory, the feature mapping Z of the cognitive graph branch is obtained. c .

[0027] The feature map Z d The extraction method includes the following steps:

[0028] Step 1: Predefine a learnable parametric graph matrix of size N*N to dynamically capture potential graph patterns in differential entropy features; use the backpropagation algorithm to dynamically learn the potential optimal graph adjacency matrix A. d Compute the graph adjacency matrix A d The regularized Laplace matrix L d :

[0029]

[0030] Among them, I N Let L be the identity matrix, and D be the regularized Laplacian matrix. d The degree matrix;

[0031] Step 2: Calculate the regularized Laplacian matrix L d proportionally scaled Laplace matrix

[0032]

[0033] Where, λ max L is a scaled Laplace matrix d The maximum value of the eigenvalue;

[0034] Step 3: Based on the spectral filtering theory, obtain the feature map Z of the data-driven graph branch. d .

[0035] The potential optimal graph adjacency matrix A d Update with a learning rate of ρ:

[0036]

[0037] Among them, the loss function is applied to A d The partial derivative is defined as

[0038] The classification module is used to classify the final feature map Z. Specifically, the final feature map Z is input into an average pooling layer, then into a fully connected layer, and finally through a softmax layer to obtain the final classification result.

[0039] The Rest-HGCN network model utilizes a definition defined as C e The network parameters are iteratively updated using a cross-entropy-based loss function and a backpropagation algorithm; wherein the loss function is:

[0040] Loss=C e (l,l p )+α||θ||

[0041] Where l is the true label of the training data, l p θ is the predicted label vector, θ is the model parameter, and α is the weight coefficient L2 used to prevent overfitting.

[0042] The method for calculating the differential entropy feature is as follows:

[0043]

[0044] Where DE(X) is the differential entropy feature of the EEG signal X, p(x) is the probability density function, and [a,b] represents the range of values.

[0045] The calculation method for the PLV network characteristics is as follows:

[0046]

[0047] Wherein, the value range of PLV(t) is [0,1]; This represents the phase difference between electrode signals x and y at time t; t is the length of the time series.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. This invention, based on the resting-state EEG signals of ASD patients, combines Phase Locking Value (PLV), a cognitive brain network feature, with spectrogram theory to mine stable graph patterns in EEG signals. It also integrates this with a data-driven graph strategy through a parameter-sharing mechanism, proposing a hybrid graph convolutional neural network model based on resting-state EEG signals, namely the Rest-HGCN network model. This model identifies effective shared EEG features between the two graph strategies, achieving highly robust ASD classification. The Rest-HGCN network model of this invention can stably extract graph patterns between multi-channel EEG signals, effectively solving problems that traditional deep learning methods cannot address, such as the lack of multi-channel information and difficulty in extracting stable features based on CNNs and RNNs. It also solves the problem that traditional graph network models using fixed graph patterns based on simple channel spatial relationships and channel feature similarities cannot dynamically extract interaction patterns and functional connections between channels.

[0050] 2. The Rest-HGCN network model employed in this invention effectively extracts stable differential patterns between normal individuals and autistic patients in resting-state EEG signals. It uses the brain functional network as a cognitive prior map to further extract complex functional connectivity patterns between different brain regions. This strategy helps extract differential features in activation patterns between normal children and children with ASD. Ultimately, it achieved an accuracy rate of 87% in single-subject experiments and 85% in cross-subject experiments. Compared with other systems, it achieved optimal recognition performance.

[0051] 3. Resting-state EEG signals offer advantages such as high temporal resolution, reduced analytical complexity, lower operational costs, and no side effects on subjects. Unlike the recognition performance reported by ASD studies based on resting-state EEG temporal features, this invention, through its classification system, can more effectively uncover the connectivity relationships between EEG channels. This not only preserves the inherent advantages of using resting-state EEG signals but also effectively improves the recognition performance of existing resting-state-based EEG signals. Attached Figure Description

[0052] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments, wherein:

[0053] Figure 1 This is a schematic diagram of the Rest-HGCN network model in this invention;

[0054] Figure 2 This is a confusion matrix diagram of two experiments in this invention;

[0055] Figure 3 This is a visualization of the features of the single-subject experiment of the present invention;

[0056] Figure 4 This is a visualization of the features of the cross-subject experiment of the present invention;

[0057] Figure 5 This is a statistical graph showing the differences in resting EEG networks between the ASD group and the healthy control group in this invention. Detailed Implementation

[0058] Example 1

[0059] As a basic embodiment of the present invention, the present invention includes an autism classification system based on resting-state electroencephalogram (EEG) signals, comprising an EEG signal acquisition and preprocessing unit and an EEG signal classification unit.

[0060] The EEG signal acquisition and preprocessing unit is used to acquire EEG signals and preprocess the acquired EEG signals to calculate differential entropy features and PLV network features.

[0061] Specifically, EEG signal preprocessing can employ conventional methods in the field, such as filtering; differential entropy features can also employ conventional methods in the field, such as obtaining the differential entropy features of a frequency band by dividing the EEG leads into frequency bands; PLV network features can also employ conventional methods in the field, including:

[0062] Given two nodes with time series x(t) and y(t) and instantaneous phases φx(t) and φy(t), the Hilbert transform is used to obtain the corresponding analytic signal H. x (t) and H y (t):

[0063]

[0064] In the formula, X(t) and Y(t) are the real parts of the EEG signals x(t) and y(t) after Hilbert transform, respectively. [ (t) and Y [ (t) represents the imaginary part of the EEG signals x(t) and y(t) after Hilbert transform, and i represents the imaginary number i. 2 =-1, defined as follows:

[0065]

[0066] In the formula, PV represents the Cauchy principal value; subsequently, the corresponding analytic signal phase φ x (t) and φ y (t) can be obtained through the following calculation.

[0067]

[0068] Finally, the connection weight w between electrodes i and j is calculated. plv ,

[0069]

[0070] In the formula, Δt represents the sampling period, and N represents the number of sampling points; w plv Between [0, 1], 1 represents complete phase synchronization and 0 represents no phase synchronization.

[0071] The EEG signal classification unit is used for classification using a trained Rest-HGCN network model. The Rest-HGCN network model includes a resting-state mixture graph network module, an attention learning module, and a classification module.

[0072] The resting-state hybrid graph network module includes a cognitive graph branch and a data-driven graph branch. The cognitive graph branch is used to extract the corresponding feature map Z based on the input PLV network features and differential entropy features. c The data-driven graph branch is used to extract the corresponding feature map Z based on the input differential entropy features. d .

[0073] The feature map Z c and feature map Z d They are respectively:

[0074]

[0075]

[0076] Where ReLU is the activation function, θ fk Let X be a learnable parameter and X be a differential entropy feature. These are the scaled Laplace matrices, respectively.

[0077] The attention learning module is used to map the feature Z... c and feature map Z d Perform fusion to obtain the final feature map Z:

[0078] Z = ω c .Z c +ω d .Z d

[0079] Where, ω c ,ω d ∈R n*n , indicating that it is assigned to the feature map Z c and feature map Z d The attention values ​​of N nodes.

[0080] The classification module is used to classify the final feature map Z and obtain the classification result.

[0081] Example 2

[0082] As the preferred embodiment of the present invention, please refer to the appendix to the specification. Figure 1 The present invention includes an autism classification system based on resting-state electroencephalogram (EEG) signals, comprising an EEG signal acquisition and preprocessing unit and an EEG signal acquisition and preprocessing unit.

[0083] The EEG signal acquisition and preprocessing unit is used to acquire EEG signals and preprocess the acquired EEG signals to calculate differential entropy features and PLV network features.

[0084] Specifically, EEG signal acquisition can be achieved using an EEG amplifier. The electrode placement standard of the EEG amplifier is the international standard 10-20 system, with a sampling rate of 500Hz and a bandpass filter range of 0.5-45Hz, for acquiring EEG signals in a resting state with eyes closed.

[0085] The preprocessing of the acquired EEG signals specifically refers to the following steps performed on the resting EEG in sequence: average reference processing, frequency band bandpass filtering (δ wave: [1Hz-4Hz), θ wave: [4Hz-8Hz), α wave: [8Hz-13Hz), and β wave: [13Hz-30Hz], 5-second data segmentation processing, baseline correction processing of segmented data, and removal of electrooculography artifacts with a threshold of 70μV.

[0086] Differential entropy features (DE features) are an extension of Shannon entropy and are used to measure the complexity of continuous random variables. Specifically, for a bandpass filtered EEG signal X, assuming that the distribution of EEG signal X follows a Gaussian distribution, the DE features of EEG signal X are calculated as follows:

[0087]

[0088] Where DE(X) is the differential entropy feature of the EEG signal X, p(x) is the probability density function, and [a,b] represents the range of values.

[0089] The method for calculating PLV network features is as follows:

[0090] To evaluate the functional correlations between EEG signals from different channels, the PLV method was employed in the construction of the resting-state EEG network. Specifically, for two EEG signal sequences x(t) and y(t), instantaneous phase synchronization was used to estimate the cognitive functional connectivity between different brain regions; therefore, the PLV network features were defined as follows:

[0091]

[0092] Wherein, the value range of PLV(t) is [0,1]; This represents the phase difference between electrode signals x and y at time t; t is the length of the time series.

[0093] The EEG signal classification unit is used for classification using a trained Rest-HGCN network model. (Refer to the attached instruction manual.) Figure 1 The Rest-HGCN network model includes a resting-state hybrid graph network module, an attention learning module, and a classification module.

[0094] The resting-state hybrid graph network module includes a cognitive graph branch and a data-driven graph branch. The PLV network features and differential entropy features serve as inputs to the cognitive graph branch, from which the corresponding feature map Z is extracted. c Simultaneously, the differential entropy feature serves as the data-driven graph branch, and by inputting the data-driven graph branch, the corresponding feature map Z is extracted. d We will use PLV network features as prior cognitive graph information in resting-state EEG network graph pattern learning, and further explore the potential graph patterns of resting-state EEG networks using a data-driven strategy on differential entropy features.

[0095] Specifically, based on the cognitive graph branch, the feature map Z c The specific extraction method includes the following steps:

[0096] Step 1: Define the initial PLV network characteristics as a graph adjacency matrix Ac, and calculate the regularized Laplacian matrix L of the graph adjacency matrix Ac. c :

[0097]

[0098] Among them, I N Let D be the identity matrix and L be the length of the matrix. c The degree matrix.

[0099] Step 2: Calculate the regularized Laplacian matrix L c proportionally scaled Laplace matrix

[0100]

[0101] Where, λ max L is the regularized Laplace matrix c The maximum value of the eigenvalue.

[0102] Step 3: Scale the Laplacian matrix By convolving the input graph with the differential entropy features, and using the graph convolution operation formula in spectral graph filtering theory, the feature mapping Z of the cognitive graph branch is calculated. c :

[0103]

[0104] Where ReLU is the activation function, θ fk Let X be a learnable parameter and X be a differential entropy feature. This is a scaled Laplace matrix.

[0105] Specifically, the spectral filtering theory is as follows:

[0106] A directed graph can be represented as G = {V, E, A}, where V = {v i Let ni = 1 be the node set, E be the edge set, and A ∈ Rn × n be the adjacency matrix describing the connections in the graph. For graph neural networks, nodes in the graph aggregate information from their neighboring nodes to obtain new embeddings. Ai,j represents the source node v. i to destination node v j The connection is as follows: If Ai,j = 0, vi and vj are not connected; if they are connected, then Ai,j ≠ 0.

[0107] Graph convolution is defined as the signal x∈R N Multiplied by filter g θ =diag(θ), the filter is defined by θ∈R in the Fourier domain. N Parameterization:

[0108] g θ *x=Ug θ U T

[0109] Where U is composed of the eigenvectors of the normalized Laplacian matrix, defined as:

[0110]

[0111] ∧ is a diagonal matrix composed of the eigenvalues ​​of L, U∧U T It is the Fourier transform of x. g θ It is a function of the L characteristic, i.e., g θ (∧).

[0112] To locate the filter in space and reduce computational complexity, a K-order Chebyshev polynomial is used, defined as: T K-1 (x)=2xT K-1 (x)-T k-2 (x), where T0(x)=1, T1(x)=x.

[0113] Then, the final graph convolution operation formula can be expressed as:

[0114]

[0115] in, Let λ be the Laplace matrix scaled proportionally to L. max This represents the maximum value of the eigenvalues ​​of L.

[0116] Specifically, based on data-driven graph branching, a learnable parametric graph matrix of size N*N is predefined to dynamically capture latent graph patterns in differential entropy features. The predefined graph is continuously updated during model learning, and the corresponding optimal data-driven graph filter is obtained when the model converges. The potential optimal graph adjacency matrix A is dynamically learned using the backpropagation (BP) algorithm. d ; and apply the loss function to A d The partial derivative is defined as:

[0117]

[0118] Using the chain rule, partial derivatives It can be redefined as:

[0119]

[0120] Potential optimal graph adjacency matrix A d It can be updated with a learning rate of ρ:

[0121]

[0122] in, This represents taking the partial derivative of the matrix; the loss function is the loss function in the Rest-HGCN network model.

[0123] The feature map Z d The specific extraction method includes the following steps:

[0124] Step 1: Calculate the graph adjacency matrix A d The regularized Laplace matrix L d :

[0125]

[0126] Among them, I N Let L be the identity matrix, and D be the regularized Laplacian matrix. d The degree matrix.

[0127] Step 2: Calculate the regularized Laplacian matrix L d proportionally scaled Laplace matrix

[0128]

[0129] Where, λ max L is a scaled Laplace matrix d The maximum value of the eigenvalue.

[0130] Step 3: Based on the spectral filtering theory, calculate the feature map Z of the data-driven graph branch. d :

[0131]

[0132] Where ReLU is the activation function, θ fk Let X be a learnable parameter and X be a differential entropy feature. This is a scaled Laplace matrix.

[0133] To further capture the discriminative patterns in the two branches and reduce the impact of redundant information, the attention learning module in the Rest-HGCN network model is used to effectively allocate attention resources on the two graph branches, as shown below:

[0134] (ω c ,ω d ) = ATT(Z c Z d )

[0135] Where, ω c ,ω d ∈R n*n Indicates assignment to Z c and Z d The attention values ​​of N nodes.

[0136] The specific calculation process, taking node i as an example:

[0137] Node i in the graph feature map Z c The embedding in is represented as Furthermore, the attention value of node i is first obtained by using the weight matrix W∈R h*h The nonlinear transformation is used to obtain the result, which is then multiplied by the shared attention vector q. c ∈R h*h Its formula is expressed as:

[0138]

[0139] From the above formula, we can obtain Attention value Then we finally get node i paired with Z. c The attention value is:

[0140]

[0141] Similarly, Z can be obtained. d Attention value of node i Finally, for all N nodes, we can obtain Z. c and Z a Attention value ωc and ω a Then, the final feature map Z is obtained as follows:

[0142] Z = ω c .Z c +ω d .Z d .

[0143] The classification module is used to classify the final feature map Z. Specifically, the final feature map Z is input into an average pooling layer of the Rest-HGCN network model, then into a fully connected layer, and finally through a softmax layer to obtain the final prediction classification result.

[0144] Furthermore, to achieve robust ASD classification, a method defined as C is used. e The network parameters are iteratively updated using a cross-entropy-based loss function and a back propagation (BP) algorithm, and the loss function can be defined as follows:

[0145] Loss=C e (l,l p )+α||θ||

[0146] Where l is the true label of the training data, l p θ is the predicted label vector, θ is the model parameter, and α is the weight coefficient L2 used to prevent overfitting.

[0147] Example 3

[0148] In another embodiment of the present invention, the present invention includes an autism classification system based on resting-state electroencephalogram (EEG) signals, comprising an EEG signal acquisition and preprocessing unit and an EEG signal classification unit. The specific method of using the classification system may include:

[0149] Step S1. Acquire EEG signals using the EEG signal acquisition and preprocessing unit, and preprocess the acquired EEG signals to calculate the differential entropy features and EEG network features. Specifically, this includes the following steps:

[0150] Step S 11 Data collection. The data collected in this study were longitudinally gathered from 280 ASD participants and 119 TD participants, aged 6-11 years. Data was acquired using a 128-channel EEG acquisition device at a sampling rate of 1000Hz.

[0151] Step S 12The acquired EEG signals were preprocessed, including averaging, 1-20Hz bandpass filtering, 5-second data segmentation, baseline correction of the segmented data, and removal of electrooculography artifacts with a threshold of 70μV.

[0152] Step S 13 Step S 12 The obtained data were used to calculate the differential entropy features and PLV network features of the subjects on the five frequency bands delta, theta, alpha, beta, and full of the training set (first batch of data).

[0153] Step S2. (This refers to step S...) 13 The obtained differential entropy features (DE features) and PLV network features are input into the resting-state mixture graph network module of the Rest-HGCN network model to obtain the corresponding feature mapping Z. c and feature map Z d .

[0154] Step S3. Utilize the attention learning module of the Rest-HGCN network model to effectively allocate attention resources on the two graph branches to obtain the final feature map Z.

[0155] Step S4. Finally, in order to achieve robust ASD classification and recognition, we utilize the method defined as C. e The network parameters are iteratively updated using a cross-entropy loss function and a backpropagation (BP) algorithm. The final feature map Z obtained by convolution of the mixture graph is input into the classification module of the Rest-HGCN network model. Specifically, it is input into an average pooling layer, then into a fully connected layer, and finally through a softmax layer to obtain the final prediction result.

[0156] The performance of the Rest-HGCN network model based on resting-state EEG signals in single-subject experiments on the ABC-CT dataset is shown in the table below:

[0157]

[0158] The performance of the Rest-HGCN network model based on resting-state EEG signals in cross-subject experiments on the ABC-CT dataset is shown in the table below:

[0159]

[0160] As shown in the two tables above, the classification accuracy based on time features reached 87% in the single-subject experiment and 85% in the multi-subject experiment, significantly outperforming the recognition performance of other models. Furthermore, the experimental results also reveal that fusing features from all frequency bands effectively improves classification results compared to single-band-based experiments. This implies that multi-band fusion feature extraction can capture more representative features and achieve better classification performance.

[0161] To further explore the effectiveness of the two branches of the Rest-HGCN network model in ASD classification, ablation experiments were conducted in both single-subject and cross-subject scenarios. Specifically, ablation experiments were performed on each module of the Rest-HGCN network model to explore the effectiveness of the proposed graph learning strategy, including the cognitive graph branch, the data-driven graph branch, and the attention learning module. The specific experimental results are shown in the table below:

[0162]

[0163] In the table above, "Only Bn" represents only the experimental results of the cognitive graph branch of the PLV network, "Only Da" represents only the results of the data-driven graph branch, and "Without Att" represents the results without an attentional learning mechanism. The table shows that compared to a single graph learning module, the hybrid graph learning strategy can more comprehensively mine the latent graph features of resting-state EEG signals. This may indicate that both branches contain useful feature information, and this information may be interdependent and constrained. Furthermore, the ablation results also demonstrate that the attentional learning mechanism plays a crucial role in eliminating information redundancy and achieving efficient representation.

[0164] To further evaluate the performance of the Rest-HGCN model, confusion matrices for normal children and children with ASD were also investigated. The corresponding results are shown in the appendix to the instruction manual. Figure 2 As shown, (a) represents the results of a single-subject experiment, and (b) represents the results of a multi-subject experiment. (See attached instruction manual.) Figure 2 The study found that in single-subject experiments, the difference in recognition rates between the two groups was not significant (87%). However, in cross-experimental studies, the recognition rate of children with ASD was higher than that of typically developing children, and the Rest-HGCN showed a sensitivity of 88% for autism, effectively improving the classification efficiency of autism.

[0165] In addition to the experiments mentioned above, visualization experiments using t-distributed random neighbor embedding (T-SNE) technology were conducted on single-subject and cross-subject experiments, respectively. In the T-SNE visualization analysis, the feature space of the initial, intermediate and final states during the model training process was visualized.

[0166] Refer to the instruction manual appendix Figure 3The figure shows the T-SNE feature visualization of the model training process of Rest-HGCN in a single-subject experiment on the ABC-CT dataset. (a) is the sample feature distribution map when the model starts training, (b) is the sample feature distribution map in the middle stage of training, and (c) is the feature distribution map when training is completed.

[0167] Refer to the instruction manual appendix Figure 4 The figure shows the T-SNE feature visualization of the model training process of Rest-HGCN across subjects on the ABC-CT dataset. (e) is the sample feature distribution map when the model starts training, (f) is the sample feature distribution map in the middle stage of training, and (g) is the feature distribution map when training is completed.

[0168] Instruction manual attached Figure 3 Included with instruction manual Figure 4 In the diagram, "0" represents a normal child and "1" represents a child with ASD. Initially, normal children and children with ASD are completely mixed in the original feature space and cannot be distinguished. As the model is further trained, the two types of feature samples gradually separate in intermediate states. The visualization results of the experiments in the final state show that the ability to distinguish between normal children and children with ASD is significantly improved. This indicates that the Rest-HGCN network model can extract high-resolution features well in both single-subject and multi-subject experiments.

[0169] The final performance of the Rest-HGCN network model shows that combining the two strategies with the attention mechanism can achieve the best performance. It further proves that the two graph branches can learn from and constrain each other through parameter sharing and the attention mechanism, thereby capturing more stable differential patterns and realizing robust ASD recognition based on resting-state EEG signals.

[0170] Most existing technologies use data extracted from brain network analysis as features in classifiers. This embodiment, however, uses this data as a graph structure representing multi-channel relationships, combined with a graph neural network, to extract the relationships between resting-state EEG signal channels. This approach better extracts the differential patterns between ASD patients and healthy individuals, thereby improving ASD classification performance. To further explore the neural interpretability of the Rest-HGCN network model, analysis of variance was used to statistically analyze the resting-state EEG networks of the ASD group and the healthy control group. The differences between the ASD group and the control group in different frequency bands were statistically significant (p<0.05), as shown in the appendix to the specification. Figure 5 As shown. To visually represent the differences, only the upper triangular portion of the undirected PLV network matrix is ​​retained. (See attached instruction manual.) Figure 5It can be observed that the resting-state EEG networks of the ASD group and the control group have different connection patterns, which may prove that the graph patterns extracted by the Rest-HGCN model proposed in this embodiment can serve as effective biomarkers for early classification and identification of ASD.

[0171] In summary, any other corresponding modifications made by those skilled in the art after reading this invention document, without creative mental effort, based on the technical solutions and concepts of this invention, are all within the scope of protection of this invention.

Claims

1. An autism classification system based on resting-state electroencephalogram (EEG) signals, characterized in that: include: The EEG signal acquisition and preprocessing unit is used to acquire EEG signals and preprocess the acquired EEG signals to calculate differential entropy features and PLV network features. The EEG signal classification unit is used for classification using a trained Rest-HGCN network model; the Rest-HGCN network model includes a resting-state mixture graph network module, an attention learning module, and a classification module. The resting-state hybrid graph network module includes a cognitive graph branch and a data-driven graph branch; the cognitive graph branch is used to extract corresponding feature maps based on the input PLV network features and differential entropy features. The data-driven graph branch is used to extract corresponding feature maps based on the input differential entropy features. ; The feature mapping and feature mapping They are respectively: in, For activation function, For learnable parameters, X It is a characteristic of differential entropy. , These are the scaled Laplace matrices, for k Chebyshev polynomial; The attention learning module is used to map features. and feature mapping Perform fusion to obtain the final feature map. Z : in, , indicating assignment to feature map and feature mapping of N Attention values ​​for each node; The classification module is used for the final feature mapping. Z The classification is performed to obtain the classification results.

2. The autism classification system based on resting-state electroencephalogram (EEG) signals according to claim 1, characterized in that: The feature mapping The extraction method includes the following steps: Step 1: Define the initial PLV network features as a graph adjacency matrix Ac, and calculate the regularized Laplacian matrix of the graph adjacency matrix Ac. : in, It is the identity matrix. D for The degree matrix; Step 2: Calculate the regularized Laplacian matrix proportionally scaled Laplace matrix : in, The regularized Laplace matrix The maximum value of the eigenvalue; Step 3: Scale the Laplacian matrix By convolving the input graph with the differential entropy feature graph, and based on spectral graph filtering theory, the feature mapping of the cognitive graph branch is obtained. .

3. The autism classification system based on resting-state electroencephalogram (EEG) signals according to claim 1 or 2 is characterized in that: The feature mapping The extraction method includes the following steps: Step 1: Predefine a size of N * N A learnable parametric graph matrix is ​​used to dynamically capture latent graph patterns in differential entropy features; the potential optimal graph adjacency matrix is ​​dynamically learned using the backpropagation algorithm. A d Compute the graph adjacency matrix A d The regularized Laplacian matrix : in, It is the identity matrix. D The regularized Laplace matrix The degree matrix; Step 2: Calculate the regularized Laplacian matrix proportionally scaled Laplace matrix : in, For scaled Laplace matrix The maximum value of the eigenvalue; Step 3: Based on the spectral filtering theory, obtain the feature mapping of the data-driven graph branch. .

4. The autism classification system based on resting-state electroencephalogram (EEG) signals according to claim 3, characterized in that: The potential optimal graph adjacency matrix A d by Learning rate update: Among them, the loss function is paired A d The partial derivative is defined as .

5. An autism classification system based on resting-state electroencephalogram (EEG) signals according to claim 1, characterized in that: The classification module is used for the final feature mapping. Z Classification specifically refers to mapping the final features. Z The input is an average pooling layer, then a fully connected layer, and finally a softmax layer to obtain the final classification result.

6. An autism classification system based on resting-state electroencephalogram (EEG) signals according to claim 4, characterized in that: The Rest-HGCN network model utilizes the definition... The network parameters are iteratively updated using a cross-entropy-based loss function and a backpropagation algorithm; wherein the loss function is: in, l These are the true labels of the training data. It is the predicted label vector. These are model parameters. These are weighting coefficients used to prevent overfitting. .

7. An autism classification system based on resting-state electroencephalogram (EEG) signals according to claim 1, characterized in that: The method for calculating the differential entropy feature is as follows: in, The differential entropy characteristic of the EEG signal X, Let be the probability density function, and [a,b] represent the range of values.

8. The autism classification system based on resting-state electroencephalogram (EEG) signals according to claim 1, characterized in that: The calculation method for the PLV network characteristics is as follows: in, The value range is [0,1]; , indicating time t Time electrode signal x and y The phase difference; t This represents the length of the time series.

Images