A cortical brain network construction method based on a causal convolutional graph neural network

Abstract

The application provides a cortical brain network construction method based on a causal convolutional graph neural network, and belongs to the field of cortical brain network construction. The method directly mines a brain cortex source interaction mode from collected head surface EEG signals, directly mines a cortical network from the head surface electroencephalogram signals in an end-to-end manner, and directly estimates a cortical directed network mode corresponding to the head surface recorded electroencephalogram signals based on a deep learning method to mine an end-to-end implicit learning relationship between the electroencephalogram signals and the cortical source space brain network, thereby avoiding problems such as ill-conditioning and instability of signal tracing in a traditional explicit solving framework, and problems such as model and hypothesis constraints in traditional causal network construction. The method can robustly estimate a physiological physiological significance cortical causal network mode from the head surface electroencephalogram signals, and has important significance for fine brain causal interaction exploration research in the time-space two-dimensional dimension.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of cerebral cortex network construction, and particularly relates to a cerebral cortex network construction method based on a causal convolutional graph neural network. BACKGROUND

[0002] Electroencephalogram (EEG) is a reflection of the electrical activity of neurons in the brain conducted to the surface of the scalp, which has the advantages of high temporal resolution, relatively low cost and non-invasive acquisition, and can be used as a powerful tool to study the neural activity of the brain. Brain network refers to the connection network between different regions of the brain, which is usually obtained by analyzing the time series of multi-channel electroencephalogram, and can be analyzed in the head surface space (head surface network) or the cortical space (cortical network). Due to the low spatial resolution, volume effect, and interference such as head movement and blinking during recording inherent in the head surface space EEG, the head surface network is unreliable and ambiguous, and therefore cannot accurately describe the cortical activity pattern of the brain. Therefore, the analysis of the cortical network is currently the mainstream method in neuroscience. At present, a feasible method is to invert (trace) the scalp signals to the cortical source space through the EEG inverse problem, and to construct the cortical brain network based on the modeling of the traced signals. The accuracy of such a strategy depends on the solution of the EEG inverse problem and the modeling of the interaction of the cortical source signals. However, the EEG inverse problem is severely underdetermined, and different source activities on the cortex can produce the same scalp activity. Traditional source analysis methods such as LORETA, SOMP, LASSO, MNE, etc. solve the problem by imposing different physical and physiological constraints. Due to the noise sensitivity, the physiological priori does not conform to the actual situation, and the regularization parameter is suboptimal, the traditional source analysis methods are prone to instability, and therefore produce neurophysiological results that are not reasonable. Accordingly, the brain network constructed based on the above traced signals is also unreliable. At the same time, the traditional method of constructing brain network is also plagued by problems such as prior constraints and explicit solution. In recent years, the powerful data mining ability of deep learning has attracted much attention, and the causal convolutional neural network for processing time series problems and the graph neural network for mining interaction relationships have been widely used in various fields. The present application proposes a causal convolutional graph neural network to establish an end-to-end direct mapping from the scalp EEG signal to the cortical source space brain network in a supervised learning manner, so as to directly solve the interaction relationship of the deep cortical source from the scalp acquisition signal, thereby avoiding the problems existing in the process of solving the inverse problem and explicitly constructing the network. SUMMARY

[0003] In order to avoid the problems of solving the source equation and the influence of prior constraints and noise on the traditional explicit network construction, the interaction mode of the cerebral cortex source is directly mined from the acquired scalp EEG signal. The present application proposes a cerebral cortex network construction method and system based on a causal convolutional graph neural network.

[0004] The application provides a cortical brain network construction method based on a causal convolutional graph neural network, and comprises the following steps.

[0005] Step S1: the electroencephalogram signals collected by the head table are subjected to the following preprocessing process:

[0006] The reference electrode standardization technique is used to re-reference the electroencephalogram signals, a band-pass filter is used to filter out signal noise components, the signals are down-sampled, the signals are segmented, and the signal segments containing artifacts are removed;

[0007] Step S2: a fixed number of preprocessed electroencephalogram signals are randomly selected as inputs of simulated cortical source neural activity signals, a multiple autoregressive equation is used to simulate neural activity signals of the remaining cortical sources, and a cortical brain network connection matrix is constructed;

[0008] The generation steps of the cortical source neural activity and the cortical brain network connection matrix are as follows:

[0009] Step S21: let T time steps simulate the cortical source signals There are m+n cortical sources; the head electroencephalogram signals of different channels are randomly selected as the activity signals corresponding to the m cortical sources The activity signals corresponding to the remaining n cortical sources The amplitude initial value is 0;

[0010] Step S22: the multiple linear regression fitting is performed on S1(t), and the least squares are used to solve the parameters of the linear autoregressive model The model parameters of each order are taken The sum of the absolute values of the element values in the corresponding positions is obtained, and the cortical brain network connection matrix corresponding to S1(t) is obtained The order P of the model is [1, 3];

[0011] Step S23: a matrix of Gaussian distribution with a mean of 0 is generated The element values in the top α% are reserved, B is set to [W O], and the state space system matrix K is set to [B A] T , wherein O is a zero matrix 0 (P×m×n) ;

[0012] Step S24: the eigenvalues of the state space system matrix are calculated, and if the eigenvalues are not less than 1, the step S23 is repeated;

[0013] Step S25: the cortical source signal S2(t) is generated according to the multiple autoregressive equation:

[0014]

[0015] Wherein, ε(t) represents the system noise at time t, represents the A matrix corresponding to the p order;

[0016] Step S26: S2(t) corresponding to the cortical brain network connection matrix Let Q = [D1O], then the cortical source signal S corresponding to the cortical brain network connection matrix D = [Q D2] T , wherein O is a zero matrix 0 (P×m×n) ;

[0017] Step S3: forward modeling of the cortical source activity signal of step S2 to obtain the simulated scalp EEG signal;

[0018] Step S4: dividing the scalp EEG signal obtained in step S3 into a training set and a validation set, adding random noise and processing based on channel normalization, inputting the simulated scalp EEG signal of the training set to the causal convolution-based graph neural network each time for training to obtain the cortical brain network connection matrix, using an optimization algorithm to calculate the loss function and its gradient and update the parameters of the graph neural network, and using the validation set to determine the optimal parameters of the graph neural network according to the size of the loss function value;

[0019] Step S5: importing the optimal graph neural network parameters determined after training, using step 1 to process the scalp EEG signal to be analyzed, and inputting the processed scalp EEG signal to the trained graph neural network to obtain the cortical brain network connection matrix;

[0020] Step S6: setting a threshold to obtain significant directed connections between cortical sources;

[0021] Further, the specific process of forward modeling of the scalp EEG signal in step S3 is: Y = HS + E is calculated to obtain the scalp signal , wherein is the current density on the cortical space, is the mapping transfer matrix from the cortical space to the scalp, is noise.

[0022] Further, the causal convolution-based graph neural network of step 4 is constructed as follows:

[0023] The graph neural network structure is sequentially composed of a cortical source signal feature extraction module and a cortical source interaction mining module; wherein the cortical source signal feature extraction module is a spatiotemporal coding feature extractor, which learns the feature mapping process from the scalp to the deep cortical space in time sequence to obtain the feature sequence of the cortical space source; the cortical source interaction mining module mines the interaction relationship between the cortical sources according to the feature sequence of the cortical space source to obtain a directed connection matrix; each element of the connection matrix represents the connection weight between two cortical sources, and the row index is the sending node and the column index is the receiving node.

[0024] Further, the training process of the step 4 causal convolution-based graph neural network is as follows:

[0025] Step S41: divide the data set into training set and validation set, set the number of iterations and initialize the model hyperparameters (including learning rate);

[0026] Step S42: add random noise to the training set and input it to the cortical source signal feature extraction module after channel normalization processing to obtain the cortical feature signal Then, the cortical brain network connection matrix is obtained through the cortical source interaction mining module The loss function is: The model parameters of the two modules are updated synchronously by using the joint training strategy, Bsize represents the number of samples in each batch of the training set, T represents the sequence length, represents the directed connection strength between the i-th source signal and the j-th source signal in the z-th sample corresponding to the nB-th batch, represents the directed connection strength between the i-th source signal and the j-th source signal in the z-th sample corresponding to the nB-th batch predicted by the model, represents the amplitude of the j-th time step in the i-th channel corresponding to the nB-th batch, represents the amplitude of the j-th time step in the i-th channel corresponding to the nB-th batch predicted by the model;

[0027] Step S43: after each iteration training, input the validation to the neural network model corresponding to the iteration round to obtain the cortical feature signal and the cortical brain network connection matrix Calculate the loss function, if the loss value is less than the recorded minimum validation loss value, save the current model parameters, and set the minimum validation error as the current loss value;

[0028] Step S44: repeat steps S42 and S43 for each iteration until the number of iterations reaches the initial set number of iterations, and the training is completed.

[0029] Further, the cortical source signal feature extraction module in the graph neural network structure is constructed as follows:

[0030] The cortical source signal feature extraction module includes a time series convolution network and a linear mapping layer, the time series convolution network system is composed of multiple residual connection modules, each residual connection module is composed of multiple residual connection modules, the output feature of the residual connection module is obtained by stacking the information input by the module to the output of the last feature submodule, and the feature submodule is sequentially composed of a one-dimensional dilated causal convolution layer, a weight normalization layer, an activation function layer, and a random zero layer.

[0031] The residual connection module: o=F(x+f(x)), wherein an activation function operation F, a combination of feature sub-modules is denoted as f, and an input is x;

[0032] The one-dimensional expansion causal convolution operation: let a one-dimensional input sequence And a filter: f: The expansion convolution operation F on the sequence elements: Wherein d is an expansion factor, r is a filter length, s-d·i is a number of historical features, an input is x, and a convolution operation is performed on the sequence after 0 padding to meet the causal characteristics based on time sequence and ensure that the input and output sequence lengths are the same, and the padding length is r-1.

[0033] Further, the cortical source interaction mining module in the graph neural network structure is constructed as follows:

[0034] The module adopts an encoder-decoder structure, wherein the encoder sequentially comprises: a first reshaping layer, a first time sequence convolution network, a first pooling layer, a second time sequence convolution network, a second pooling layer, a first edge interaction encoding layer, a second reshaping layer, a first convolution network, a first average feature layer, a third reshaping layer, a node encoding layer, a second edge interaction encoding layer, a fourth reshaping layer, a second convolution network, a second average feature layer, and a splicing layer, wherein the output of the first average feature layer and the output feature of the second average feature layer are input to the splicing layer for feature splicing.

[0035] The edge interaction encoding layer operation:

[0036] Each connection mode between the cortical sources is grouped into a sending source end and a receiving source end, and one-hot encoding is adopted to encode the serial numbers of the sending source end and the receiving source end into matrices, and the layer input signal is subjected to matrix multiplication operation with the encoding matrices of the sending source end and the receiving source end to extract the sending source end feature sequence and the receiving source end feature sequence, and the sending source end feature sequence and the receiving source end feature sequence corresponding to the connection mode are spliced, and the feature sequence of the sending-receiving source is the feature corresponding to the connection mode.

[0037] The node encoding layer operation:

[0038] The encoding matrix of the serial number of the receiving source end is transposed, and then subjected to matrix multiplication operation with the layer input to obtain the average feature of all cortical sources flowing into the cortical source by dividing the obtained feature sequence by the number of cortical sources excluding itself.

[0039] The convolution network structure is as follows:

[0040] The structure sequentially comprises a one-dimensional conventional convolution layer, a weight normalization layer, and a time sequence convolution network.

[0041] The average feature layer operation:

[0042] The average value of elements along the row dimension of each sample is calculated;

[0043] The concatenation layer operation:

[0044] The concatenation of sequences along the column dimension of each sample is performed;

[0045] The decoder structure is as follows:

[0046] The decoder is an artificial neural network structure, which is composed of multiple multilayer perceptron modules, linear mapping layers and parameterized rectified linear unit scoring functions, each multilayer perceptron module is sequentially composed of a fully connected layer activated by a nonlinear function, a random zero setting layer, a fully connected layer activated by a nonlinear function, and a layer normalization layer;

[0047] The parameterized rectified linear unit activation function:

[0048]

[0049] Where λ is a weight factor, which is a constant term;

[0050] The layer normalization layer operation:

[0051] The mean μ and the standard deviation σ of the input data along the row dimension are calculated, and the formula is calculated ε is e -6 , trainable parameters γ, β.

[0052] In a second aspect, the application provides a cortical brain network construction system based on a causal convolutional graph neural network, comprising:

[0053] A preprocessing module: preprocessing the head surface electroencephalogram signals collected by the sensor to be analyzed, and reducing noise interference;

[0054] A data generation module: simulating a cortical source neural activity signal, constructing a cortical brain network and a corresponding head surface electroencephalogram signal, and providing samples for training a model and verifying the performance of the model;

[0055] A training module: inputting the training samples into the graph neural network to obtain a cortical feature signal and a cortical brain network connection matrix D, calculating the sum of the average mean square error of the predefined cortical brain network connection matrix and the estimated cortical brain network connection matrix and the average mean square error of the cortical source neural activity signal S and the model learned cortical feature signal , updating the parameters of the model using a self-adaptive momentum random optimization algorithm, setting a verification sample to verify the effect of the model after each round of training to determine the model parameters of the graph neural network;

[0056] An analysis module: using the trained graph neural network to analyze the electroencephalogram signal dataset to obtain a directed brain network matrix, and obtaining significant directed connection results in the dataset according to a set threshold.

[0057] The present application has the advantages of:

[0058] The features of the future information based on the causal convolution cannot be leaked to the current features to simulate the space-time transmission characteristics of the head surface signal inversion to the cortical source space, compared with the recurrent neural network, a plurality of cortical source activity electric signals can be quickly and parallelly reconstructed; the proposed causal convolution graph neural network realizes the time-sequential feature coding based on the time sequence of the interaction between the cortical sources, so that the cortical network has higher accuracy. After supervised learning, it can be used to mine the interaction network mode of the deep cortical source from the head surface collected electric signal. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A step schematic diagram of a cortical brain network construction method based on a causal convolution graph neural network is provided.

[0060] Figure 2 A cortical source signal feature extraction module in a graph neural network structure is provided.

[0061] Figure 3 A cortical source interaction mining module in a graph neural network structure is provided.

[0062] Figure 4 The cortical brain network connection topological structure under the resting state. DETAILED DESCRIPTION

[0063] The embodiments of the present application will be further described below with reference to the accompanying drawings.

[0064] In a first aspect, a cortical brain network construction method based on a causal convolution graph neural network is provided according to the embodiments of the present application, please refer to Figure 1 , comprising the following steps:

[0065] Step S1: the electroencephalogram signals collected from the head surface are subjected to the following preprocessing procedures: using reference electrode standardization technology to process the signals, using band-pass filtering to filter out other frequency band information of the signals, reducing the signal sampling rate, signal segmentation slicing, and removing the signal segments containing artifacts;

[0066] Step S2: randomly selecting a fixed number of channels of the preprocessed electroencephalogram signals as simulated partial cortical source neural activity signals, using a multivariate autoregressive equation to simulate the neural activity signals of the remaining cortical sources and constructing a cortical brain network connection matrix;

[0067] Step S3: forward modeling the cortical source activity signal of step S2 to obtain a simulated scalp EEG signal using a linear model;

[0068] Step S4: dividing the scalp EEG signal obtained in step S3 into a training set and a validation set, adding random noise and processing based on channel normalization, inputting the simulated scalp EEG signal of the training set to the causal convolution-based graph neural network each time for training to obtain a cortical brain network connection matrix, using an optimization algorithm to calculate the loss function and its gradient and update the parameters of the graph neural network, and using the validation set to determine the optimal parameters of the graph neural network according to the size of the loss function value;

[0069] Step S4: importing the optimal graph neural network parameters determined after training, processing the scalp EEG signal to be analyzed using step 1, and inputting the processed scalp EEG signal to the trained graph neural network to obtain a cortical brain network connection matrix;

[0070] Step S5: setting a threshold to obtain significant directed connections between cortical sources; in this example, the threshold is set to the sum of the mean and 0.5 times the variance

[0071] Further, the step S1 specifically comprises:

[0072] Step S11: re-referencing the EEG signal using the Reference Electrode Standardization Technique (REST);

[0073] Step S12: filtering out signals of a specific frequency band and power frequency noise interference using a band-pass filter; in this embodiment, the upper limit of the filter for the EEG signal used to generate the data is 30Hz, and the lower limit is 1Hz;

[0074] Step S13: down-sampling the EEG signal; in this embodiment, the sampling rate of the EEG signal used to generate the data is 1000Hz, and the sampling rate after processing is 256Hz;

[0075] Step S14: segmenting the potential signal continuously collected for a period of time in a non-overlapping manner according to a certain same interval L; in this embodiment, the window length is 2s;

[0076] Step S15: removing the artifact segment of the signal segment obtained in step S14. That is, comparing each time point in the signal segment with the set artifact threshold one by one, and if a time point in the segment exceeds the threshold, the signal segment is removed; in this embodiment, the artifact threshold is set to 75μV;

[0077] Further, the step S2 specifically comprises:

[0078] Step S21: set T time step simulation cortex source signal There are m+n cortex sources; randomly select the head surface electroencephalogram signals of different channels as the activity signals corresponding to the m cortex sources The activity signals corresponding to the remaining n cortex sources The amplitude initial value is 0;

[0079] Step S22: perform multiple linear regression analysis on S1, and solve the parameters of the linear autoregressive model by least squares (Determine the order p of the model before modeling by AIC criterion) Take the model parameters of each order The sum of the absolute values of the element values at the corresponding positions is obtained, and the cortex brain network connection matrix corresponding to S1 is obtained

[0080] Step S23: generate a matrix of Gaussian distribution with mean 0 The element values of the top α% are retained, let B=[W O], and the state space system matrix K=[B A] T , where O is a zero matrix 0 (P×m×n) ;

[0081] Step S24: calculate the eigenvalues of the state space system matrix, and if they are not less than 1, repeat step S23;

[0082] Step S25: generate the cortex source signal S2(t) according to the multiple autoregressive equation:

[0083]

[0084] Where ε(t) represents the system noise at time t, represents the A matrix corresponding to the p-th order;

[0085] Step S26: the cortex brain network connection matrix corresponding to S2(t) Let Q=[D1O], then the cortex brain network connection matrix D corresponding to the cortex source signal S is [Q D2] T , where O is a zero matrix 0 (P×m×n) ;

[0086] Further, the specific process of forward modeling of the head surface electroencephalogram signal in step S3 is: calculate the head surface signal by Y=HS+E Where is the current density in the cortex space, is the conduction space matrix from the cortex source space to the head surface, Noise introduced in electroencephalogram signal recording. In this example, the cortical sources are selected to be the dorsal medial prefrontal cortex (dMPFC, MNI coordinates [x y z]: [0 52 26]), the anterior medial prefrontal cortex (aMPFC, MNI coordinates [x y z]: [-6 52 -2]), the ventral medial prefrontal cortex (vMPFC, MNI coordinates [x y z]: [0 26 -18]), the left lateral temporal cortex (LLTC, MNI coordinates [x y z]: [-60 -24 -18]), the right lateral temporal cortex (RLTC, MNI coordinates [x y z]: [60 -24 -18]), and the posterior cingulate cortex (PCC, MNI coordinates [x y z]: [0 -58 27]);

[0087] Further, the step 4 causal convolution-based graph neural network is constructed as follows:

[0088] The graph neural network structure is sequentially composed of a cortical source signal feature extraction module and a cortical source interaction mining module, wherein the cortical source signal feature extraction module is designed as a space-time coding feature extractor, learning the feature mapping process from the head table (sensor acquisition space) to the deep cortical space in time sequence, obtaining the feature sequence of the cortical space source, and the cortical source interaction mining module mines the interaction relationship between the cortical sources according to the feature sequence of the cortical space source, obtaining a directed connection matrix, the element value of which represents a connection weight, the matrix row index is a sending node, and the matrix column index is a receiving node.

[0089] Further, please refer to Figure 2 , the cortical source signal feature extraction module in the graph neural network structure is constructed as follows:

[0090] The cortex source signal feature extraction module comprises a time sequence convolution network architecture and a linear mapping layer, the convolution operation in the time sequence convolution network architecture makes the information transmission comply with the causal logic (that is, the current information is only determined by the features of historical information, and the features of future information cannot be leaked), the time sequence convolution network architecture is composed of a plurality of residual connection modules, the output features of each residual connection module are obtained by superimposing the information of the module input to the output of the last feature submodule, and the feature submodule is sequentially composed of a one-dimensional dilated causal convolution layer, a weight normalization layer, an activation function layer and a random zero setting layer;

[0091] Further, the residual connection module: o=F(x+f(x)), wherein the activation function operation F, the combination of the feature submodules is denoted as f, and the input is x;

[0092] Further, the one-dimensional dilated causal convolution operation is as follows: let a one-dimensional input sequence be x, and a filter be f: The expansion convolution operation F on the sequence elements is as follows: Wherein d is a dilated factor, r is a filter length, s-d·i is a historical feature quantity, the input is x, and the 0 padding is performed on the sequence before the convolution operation to ensure that the causal characteristics based on the time sequence are met and the length of the input and output sequences is the same, and the padding length is (r-1), and then the conventional one-dimensional sequence convolution operation is performed.

[0093] Further, please refer to Figure 3 The cortex source interaction mining module in the graph neural network structure is constructed as follows:

[0094] The module adopts an encoder-decoder structure, wherein the encoder is sequentially composed of a reshaping layer, a time sequence convolution network, a pooling layer, a time sequence convolution network, an edge interaction encoding layer, a reshaping layer, a convolution network, an average feature layer, a node encoding layer, a reshaping layer, a node encoding layer, an edge interaction encoding layer, a reshaping layer, a convolution network, an average feature layer and a splicing layer, wherein the output of the first average feature layer (obtained shallow layer encoding features of node interaction) and the output features of the second average feature layer (obtained deep layer encoding features of node interaction) are input to the splicing layer for feature splicing.

[0095] Further, the edge interaction encoding layer operation is as follows:

[0096] ​For each connection mode between the cortex sources (i.e. information flows from cortex source 1 to cortex source 2 or from cortex source 2 to cortex source 1), a one-hot encoding is used to encode the sequence number of the sending source end and the receiving source end into a matrix, respectively, and the layer input signal is subjected to a matrix multiplication operation with the encoding matrix of the sending source end and the receiving source end, respectively, to extract the sending source end feature sequence and the receiving source end feature sequence, and the sending-receiving source feature sequence is spliced with the sending source end feature sequence and the receiving source end feature sequence corresponding to the connection mode.

[0097] Further, the node encoding layer operation is as follows:

[0098] The encoding matrix of the sequence number of the receiving source end is transposed, and then subjected to a matrix multiplication operation with the layer input to obtain a feature sequence, which is divided by the number of cortex sources excluding itself to obtain the average feature of all cortex sources flowing into the cortex source;

[0099] Further, the convolutional network structure is as follows:

[0100] The structure is sequentially composed of a one-dimensional regular convolutional layer, a weight normalization layer and a time series convolutional network;

[0101] Further, the average feature layer operation is as follows:

[0102] The average value of elements is calculated along the row dimension of each sample;

[0103] Further, the splicing layer operation is as follows:

[0104] The sequence is spliced along the column dimension of each sample;

[0105] Further, the decoder structure is as follows:

[0106] The decoder is an artificial neural network structure, which is composed of multiple multilayer perceptron (MLP) modules, a linear mapping layer and a parametric rectified linear unit (PReLu) scoring function with parameters, each MLP module is sequentially composed of a fully connected layer activated by a nonlinear function, a random zero setting layer, a fully connected layer activated by a nonlinear function, and a layer normalization layer; in this example, the fully connected layer activated by a nonlinear function adopts an ELU activation function:

[0107]

[0108] wherein ξ is a weight factor, which is a constant term;

[0109] Further, the parametric rectified linear unit activation function with parameters is as follows:

[0110]

[0111] wherein, λ is a weight factor, and is a constant term;

[0112] Further, the layer normalization layer operation is:

[0113] The mean μ and the standard deviation σ are calculated along the row dimension of the input data, and are calculated according to the formula (ε is e -6 , and the trainable parameters γ, β);

[0114] Further, the training process of step 3 is as follows:

[0115] Step S31: divide the data set into a training set and a validation set, set the number of iterations and initialize the model hyperparameters (including the learning rate);

[0116] Step S32: add random noise to the training set and input it to the cortical source signal feature extraction module after channel normalization processing to obtain the cortical feature signal Then, the cortical brain network connection matrix D is obtained through the cortical source interaction mining module, and the loss function is: The model parameters of the two modules are updated synchronously by adopting a joint training strategy, Bsize represents the number of samples in each batch of the training set, T represents the sequence length, represents the directed connection strength between the i-th source signal and the j-th source signal in the z-th sample corresponding to the nB-th batch, represents the directed connection strength between the i-th source signal and the j-th source signal in the z-th sample corresponding to the nB-th batch predicted by the model, represents the amplitude of the j-th time step in the i-th channel corresponding to the nB-th batch, represents the amplitude of the j-th time step in the i-th channel corresponding to the nB-th batch predicted by the model;

[0117] Step S33: after each iteration training, input the validation to the neural network model corresponding to the iteration round to obtain the cortical feature signal S and the cortical brain network connection matrix D, calculate the loss function, if the loss value is less than the recorded minimum validation loss value, save the current model parameters, and set the minimum validation error as the current loss value;

[0118] Step S34: repeat steps S32 and S33 every iteration until the number of iterations reaches the initial set number of iterations, and the training is completed.

[0119] This embodiment constructs the average cortical brain network connection matrix of 7 normal people in the resting state, which illustrates the feasibility of the application:

[0120] The EEG signals of 7 normal persons in a resting state (about 5 minutes, sampling rate of 1000 Hz) are collected, preprocessed according to the specific parameters described in steps S1 and specific examples, and the obtained signals are input into the causal convolution-based graph neural network after channel normalization processing, to obtain the cortical brain network connection matrix of each subject. The element values of the corresponding positions of the cortical brain network connection matrices of all subjects are added and averaged to obtain the average brain network connection matrix of the 7 normal persons. The mean and variance of the average brain network connection matrix are calculated to obtain the set connection screening threshold. The significant connections of the average brain network connection matrix are obtained and marked in red, and the non-significant connections are marked in gray. The greater the weight value, the darker the color. Please refer to Figure 4 : aM. represents the aMPFC region, and dM. represents the dMPFC region. In the figure, the aMPFC, vMPFC, and dMPFC frontal lobe regions interact with each other and interact with the bilateral temporal lobe regions and occipital lobe regions in the resting state.

[0121] Therefore, the present application proposes a causal convolution-based graph neural network to directly obtain the interaction relationship of deep cortical sources in an end-to-end manner through signals collected by sensors arranged on the head, so as to avoid the EEG inverse problem solving problem, and to obtain a direct mapping function from the EEG signal to the cortical brain network in a supervised learning manner.

[0122] The above only describes the preferred embodiments of the present application, and any equivalent changes and modifications made within the scope of the patent application of the present application shall be included in the scope of the present application.

Claims

1. A method for constructing a cortical brain network based on a causal convolutional graph neural network, comprising the following steps: Step S1: The electroencephalogram signals collected on the scalp are subjected to the following preprocessing steps: The electroencephalogram signals are re-referenced using a reference electrode standardization technique, filtered to remove signal noise components using a band-pass filter, down-sampled, segmented, and the signal segments containing artifacts are removed; Step S2: A fixed number of channels of the preprocessed electroencephalogram signals are randomly selected as inputs for simulating the neural activity signals of the cortical sources, and a multivariate autoregressive equation is used to simulate the neural activity signals of the remaining cortical sources and construct a cortical brain network connection matrix; The steps for generating the cortical source neural activity and the cortical brain network connection matrix are as follows: Step S21: set T time step simulation cortex source signal There are m+n cortex sources; randomly select the head surface electroencephalogram signals of different channels as the activity signals corresponding to the m cortex sources The activity signals corresponding to the remaining n cortex sources The initial value of the amplitude is 0; Step S22: performing multiple linear regression fitting on S1(t), and solving the parameters of the linear autoregressive model by least squares Taking the model parameters of each order Summing the absolute values of the element values of the corresponding positions to obtain the cortical brain network connection matrix corresponding to S1(t) The order P of the model is in [1, 3]. Step S23: generating a matrix of Gaussian distribution with mean 0 The values of the top a% elements are reserved before sorting, let B = [W O], then the state space system matrix K = [B A] T where O is a zero matrix 0 (P×m×n) ; Step S24: Calculate the eigenvalues of the state space system matrix, and if they are not less than 1, repeat step S23; Step S25: Generate the cortical source signal S2(t) according to the multivariate autoregressive equation: wherein ε(t) represents the system noise at time t, represents the A matrix corresponding to the pth order; Step S26: The cortical brain network connection matrix corresponding to S2(t) Let Q = [D1 O], then the cortical brain network connection matrix D = [Q D2] corresponding to the cortical source signal S T where O is a zero matrix 0 (P×m×n) ; Step S3: The cortical source activity signals of step S2 are forward modeled to obtain simulated electroencephalogram signals on the scalp; Step S4: The electroencephalogram signals on the scalp obtained in step S3 are divided into a training set and a validation set, random noise is added, and channel normalization is performed, the simulated electroencephalogram signals on the scalp of the training set are input into the causal convolution-based graph neural network each time, the cortical brain network connection matrix is obtained, an optimization algorithm is used to calculate the loss function and its gradient and update the parameters of the graph neural network, and the optimal parameters of the graph neural network are determined according to the size of the loss function value using the validation set; Step S5: Import the optimal graph neural network parameters determined after training, process the electroencephalogram signals on the scalp to be analyzed using step 1, input the processed electroencephalogram signals on the scalp into the trained graph neural network, and obtain the cortical brain network connection matrix; Step S6: Set a threshold to obtain significant directed connections between cortical sources.

2. The method of claim 1, wherein the causal convolutional graph neural network is based on a graph neural network (GNN) and a causal convolutional neural network (CNN). The forward modeling of the step S3 obtains the electroencephalogram signal of the head surface in detail: the head surface signal is calculated by Y=HS+E wherein is the current density on the cortical space, is the mapping transfer matrix from the cortical space to the head surface, is the noise.

3. The method of claim 1, wherein the method comprises: The graph neural network of step 4 is constructed as follows: The graph neural network structure is sequentially composed of a cortical source signal feature extraction module and a cortical source interaction mining module; the cortical source signal feature extraction module is a spatiotemporal coding feature extractor that learns the feature mapping process from the scalp to the deep cortical space in time sequence, obtaining a feature sequence of the cortical space source; the cortical source interaction mining module mines the interaction between the cortical sources according to the feature sequence of the cortical space source, obtaining a directed connection matrix; each element of the connection matrix represents the connection weight between two cortical sources, with the row index as the sending node and the column index as the receiving node.

4. The method of claim 3, wherein the causal convolutional graph neural network is constructed based on a plurality of brain regions, and the plurality of brain regions are divided into a plurality of brain regions of different brain networks according to a brain network parcellation method. The training process of the causal convolution-based graph neural network in step 4 is as follows: Step S41: Divide the data set into a training set and a validation set, set the number of iterations and initialize the model hyperparameters (including the learning rate); Step S42: add random noise to the training set and input it into the cortical source signal feature extraction module after channel normalization processing to obtain the cortical feature signal Then, the cortical brain network connection matrix is obtained through the cortical source interaction mining module The loss function is: The model parameters of the two modules are updated synchronously using a joint training strategy. Bsize represents the number of samples in each batch of the training set, T represents the sequence length, represents the directed connection strength between the ith source signal and the jth source signal in the zth sample corresponding to the nth B batch, represents the directed connection strength between the ith source signal and the jth source signal in the zth sample corresponding to the nth B batch predicted by the model, represents the amplitude of the jth time step in the ith channel corresponding to the nth B batch, represents the amplitude of the jth time step in the ith channel corresponding to the nth B batch predicted by the model; Step S43: after each iteration training, the validation is input to the neural network model corresponding to the iteration round to obtain the cortical feature signal and the cortical brain network connection matrix The loss function is calculated, and if the loss value is less than the recorded minimum validation loss value, the current model parameter is saved, and the minimum validation error is set as the current loss value; Step S44: Repeat steps S42 and S43 for each iteration until the number of iterations reaches the initial number of iterations, and the training is complete.

5. The method of claim 4, wherein the causal convolutional graph neural network is based on a graph neural network (GNN) and a causal convolutional neural network (CNN). The cortical source signal feature extraction module in the graph neural network structure is constructed as follows: The cortex source signal feature extraction module comprises a time sequence convolution network and a linear mapping layer, the time sequence convolution network system is composed of a plurality of residual connection modules, each residual connection module is composed of a plurality of residual connection modules, the output feature of the residual connection module is obtained by superimposing the information of the module input to the output of the last feature submodule, and the feature submodule is sequentially composed of a one-dimensional dilated causal convolution layer, a weight normalization layer, an activation function layer and a random zero setting layer; The residual connection module is o=F(x+f(x)), wherein the activation function operation F, and the combination of the feature submodule is denoted as f, and the input is x; The one-dimensional dilated causal convolution operation is: a one-dimensional input sequence And a filter: f: A convolution operation F on the expansion of sequence elements: Wherein d is a dilated factor, r is a filter length, s-d·i is a number of historical features, the input is x, in order to meet the causal characteristics based on time sequence and ensure that the input and output sequence lengths are the same, the sequence is filled with 0 before the convolution operation, the filling length is r-1, and then the conventional one-dimensional sequence convolution operation is carried out.

6. The method of claim 4, wherein the causal convolutional graph neural network is trained using a loss function that is based on a reconstruction error of the causal convolutional graph neural network. The cortex source interaction mining module in the graph neural network structure is constructed as follows: The module adopts an encoder and a decoder structure, wherein the encoder sequentially comprises a first reshaping layer, a first time sequence convolution network, a first pooling layer, a second time sequence convolution network, a second pooling layer, a first edge interaction encoding layer, a second reshaping layer, a first convolution network, a first average feature layer, a third reshaping layer, a node encoding layer, a second edge interaction encoding layer, a fourth reshaping layer, a second convolution network, a second average feature layer and a splicing layer in sequence, wherein the output of the first average feature layer and the output feature of the second average feature layer are input to the splicing layer for feature splicing; The edge interaction encoding layer operates as follows: Each connection mode between the cortex sources is grouped into a sending source end and a receiving source end, and is encoded into a matrix by using one-hot encoding, and the layer input signal is subjected to matrix multiplication operation with the encoding matrix of the sending source end and the receiving source end to extract the sending source end feature sequence and the receiving source end feature sequence, and the sending source end feature sequence and the receiving source end feature sequence corresponding to the connection mode are spliced, and the feature sequence of the sending-receiving source is the feature corresponding to the connection mode; The node encoding layer operates as follows: The encoding matrix of the serial number of the receiving source end is transposed, and then subjected to matrix multiplication operation with the layer input to obtain the average feature of all cortex sources flowing into the cortex source by dividing the obtained feature sequence by the number of cortex sources excluding itself; The convolution network structure is as follows: The structure is sequentially composed of a one-dimensional conventional convolution layer, a weight normalization layer and a time sequence convolution network; The average feature layer operates as follows: The average value of elements is calculated along the row dimension of each sample; The splicing layer operates as follows: The sequence is spliced along the column dimension of each sample; The decoder structure is as follows: The decoder is an artificial neural network structure, which is composed of a plurality of multilayer perceptron modules, a linear mapping layer and a parameterized rectified linear unit scoring function, each multilayer perceptron module is sequentially composed of a fully connected layer activated by a nonlinear function, a random zero setting layer, a fully connected layer activated by a nonlinear function and a layer normalization layer; The parameterized rectified linear unit activation function is as follows: Wherein, λ is a weight factor, which is a constant term; The layer normalization layer operates as follows: The input data is calculated along the row dimension for the mean μ and the standard deviation σ, according to the formula ε is e -6 The trainable parameters γ, β.

7. A cortex brain network construction method system based on the causal convolution graph neural network of claim 1, the system comprising: A preprocessing module: preprocessing the head surface electroencephalogram signal collected by the sensor to be analyzed to reduce noise interference; Data generation module: simulate the cortical source neural activity signal, construct the cortical brain network and the corresponding head surface electroencephalogram signal, and provide samples for training model and verifying model performance; a training module: inputting training samples into the graph neural network to obtain a cortical feature signal and a cortical brain network connection matrix D, calculating the sum of the average mean square error of the pre-defined cortical brain network connection matrix and the estimated cortical brain network connection matrix and the average mean square error of the cortical source neural activity signal S and the model learned cortical feature signal , gradient, and updating the parameters of the model using a random optimization algorithm with adaptive momentum, setting a validation sample to verify the effect of the model after each round of training to determine the model parameters of the graph neural network. Analysis module: use the trained graph neural network to analyze the electroencephalogram signal dataset to obtain the directed brain network matrix, and obtain the significant directed connection results in the dataset according to the set threshold.

Images