A method and system for recognizing EEG signals for functional remodeling of movement disorders

Abstract

The present invention relates to an EEG signal recognition method and system for functional remodeling of movement disorders. The recognition method comprises the following steps: S1: setting a sampling period to obtain multiple EEG signals; S2: preprocessing the EEG signals and extracting a training set; S3: dividing the EEG signals in the training set into non-overlapping time windows, calculating the Pearson coefficient of each time window, and obtaining the time-domain brain functional connection; S4: establishing a brain biotopological model based on the time-domain brain functional connection and spatial domain information; S5: extracting the initial features within a specific frequency band, and using the features on the calibrated training set as input features; S6: constructing a learnable graph convolutional neural network model based on the input features and the brain biotopological model, and outputting a predicted label. The present invention fully integrates the EEG information in the time-frequency-space domain to learn more discriminative motor imagery features, so that the model reasonably optimizes the electrode relationship during the training process, thereby improving the accuracy of EEG signal recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal recognition, and in particular to an electroencephalogram (EEG) signal recognition method for movement disorder function remodeling and an electroencephalogram (EEG) signal recognition system for movement disorder function remodeling. Background Art

[0002] Movement is one of the most basic functions in people's daily life. Patients with movement disorders whose motor functions are impaired to varying degrees due to accidents or diseases have their ability to live and their quality of life seriously affected. At present, rehabilitation training for patients with movement disorders often involves passive movement of the affected part, and the assistive devices are only associated with the muscles at the affected part and are not connected to the nervous system, which makes it difficult to mobilize the enthusiasm and autonomy of people with movement disorders. Therefore, patients with movement disorders are in urgent need of more effective auxiliary rehabilitation tools and more proactive training methods. Electroencephalogram (EEG) has the advantages of spontaneity and high precision, and is widely used in medical health, smart driving, life and entertainment. Reshaping the brain based on EEG signals and converting them into instructions to control movement can help patients with movement disorders repair and supplement the output of brain information, and enhance the information interaction between trained muscles and the central nervous system.

[0003] At present, the difficulties in accurately identifying EEG signals are mainly manifested in the following three aspects: 1. EEG signals have non-stationary characteristics and individual differences. That is, facing the same stimulus, individuals with different cultural backgrounds and growth environments have different EEG representations. The model trained by the data set may only be applicable to the current individual and lack generalization ability. 2. Interference of redundant electrode information. The network nodes of the current deep network model are generally full-channel EEG signals, which causes some signals generated by electrodes with low correlation with movement to enter the network training, generating a large amount of redundant information, introducing interference information for the model to learn autonomously, resulting in negative learning. 3. Brain functional networks are difficult to construct. The brain is different from other structures of the human body. The connections between brain regions do not have a clear physical topological structure. The spatial position connection between EEG channels is also different from the functional connection between them. Previous studies often ignore the information connection between brain regions. Summary of the invention

[0004] Based on this, it is necessary to provide an EEG signal recognition method and system for movement disorder functional remodeling to address the problem of low EEG signal recognition accuracy. The present invention is implemented through the following technical solutions: An EEG signal recognition method for movement disorder functional remodeling includes the following steps:

[0005] S1: Set the sampling period and collect multiple EEG signals from different brain regions of multiple subjects in real time through downsampling.

[0006] S2: Preprocess the EEG signals to remove artifacts and then map them to a preset frequency band. Allocate multiple EEG signals in each frequency band according to a preset ratio and divide them into a training set and a test set.

[0007] S3: Divide the EEG signals in the training set into multiple non-overlapping time windows according to a preset time period; extract the training labels of the EEG signals in each time window; perform threshold screening on the training labels and normalize them to obtain time-domain brain functional connectivity; brain functional connectivity is used to characterize the degree of connection between different brain regions in the same time window.

[0008] S4: Establish a brain biotopological model based on the time-domain brain functional connectivity and the physical distance of the electrodes in the space domain. Obtain the connections between EEG channels based on the brain biotopological model and construct an initial graph adjacency matrix. The method for constructing the initial graph adjacency matrix includes the following steps:

[0009] S41: Initialize the graph adjacency matrix based on the theory that the connection strength between brain regions decays inversely with the physical distance.

[0010]

[0011] d i,j represents the physical distance between electrode i and electrode j, η is a sparse parameter, and A is constructed as an undirected symmetric matrix to initialize the local channel relationship.

[0012] S42: Set the global channel relationship according to the initialized graph adjacency matrix. The global channel relationship is expressed as:

[0013] A i,j =A i,j -1;

[0014] In order to maintain the asymmetry between brains, some global connections are added to better simulate the information transmission between brains using asymmetric information. i,j That is, the adjacency matrix value between the i-th and j-th electrodes selected as the global channel.

[0015] S5: intercepting EEG signals within a preset frequency band, dividing the EEG data of each frequency band into samples containing t seconds of EEG signals using non-overlapping time windows of t seconds, and dividing the EEG signals into multiple EEG signal sets according to the preset frequency band range; extracting the initial features of the EEG signals in each EEG signal set, enhancing the initial feature weights of the key frequency bands on the training set, and selecting multiple initial features with the highest weight values ​​as input features of the network nodes.

[0016] S6: Construct a learnable graph convolutional neural network model based on the input features of the network nodes and the brain biotopology model. Input the fused adjacency matrix and frequency domain features into the learnable graph convolutional neural network to reasonably optimize the electrode relationship. The learnable graph convolutional neural network autonomously learns the identification features and obtains the predicted label.

[0017] The above recognition method fully integrates the EEG information in the time-frequency-space domain to learn more discriminative motor imagery features, effectively initializes the graph adjacency matrix through the spatial domain electrode topological relationship, and at the same time, constructs a time domain functional network to supplement the time domain information, so that the model can reasonably optimize the electrode relationship during the training process, autonomously learn better features, and improve the accuracy of EEG signal recognition.

[0018] In one embodiment, the method for acquiring time-domain brain functional connectivity includes the following steps:

[0019] S31: dividing the EEG signals in the corresponding frequency band into multiple time groups according to the data set acquisition paradigm, each time group contains multiple EEG data with a time window of t seconds.

[0020] S32: Calculate the Pearson coefficient of each window based on the EEG data in each time group. The Pearson coefficient is used to characterize the degree of connection between electrodes in the corresponding time period. The Pearson coefficient is expressed as:

[0021] A Per ∈R S*F*N*N ,

[0022] Among them, S is the number of samples, F is the number of frequency bands, and N is the number of electrodes.

[0023] S33: Divide the EEG signal into different categories; take the average of the Pearson coefficients of the corresponding windows in each category to obtain the training label.

[0024] S34: Threshold selection is performed on the training labels of each category to obtain multiple matrices; the multiple matrices are summed and normalized to obtain the time domain brain functional connectivity A Te ∈R N*N .

[0025] In one embodiment, the EEG signal X Te According to the data set acquisition paradigm, the test segments are divided into k time groups, each group contains multiple EEG data with a time window of t seconds. The number of samples in each segment r varies according to the length of the time group. The preprocessed EEG signal X Te It is expressed as:

[0026] X Te ∈R N×L

[0027] Where N is the number of electrodes and L is the number of samples after time domain sampling. The number of samples L can be expressed as:

[0028]

[0029] Among them, S P represents the sampling rate during preprocessing, Indicates the length of the nth window in the mth segment.

[0030] In one embodiment, the Pearson coefficient is expressed as:

[0031]

[0032] Among them, σ represents the standard deviation, E represents the expected function, μ represents the expected value, and v and θ represent the EEG data on different electrodes in the same time range. The EEG data in the time domain are mapped to 5 functional frequency bands, and then the Pearson coefficient between electrodes is calculated for each non-overlapping t-second time window in each frequency band of the training data set to obtain A Per ∈R S*F*N*N , where S is the number of samples, F is the number of frequency bands, and N is the number of electrodes as above. Next, according to the known training data labels, the mean is taken for each category, and then the threshold is selected to retain the connection of reasonable strength, and the threshold is applied to each category. Finally, the matrices of all categories and all frequency bands obtained by screening are summed and normalized to finally obtain the time domain brain functional connectivity A Te ∈R N*N .

[0033] In one embodiment, each frequency band is used as a feature detector, and the frequency band attention module is used to enhance the initial feature weights of the key frequency bands. Then the input feature is represented as:

[0034]

[0035] Among them, δ1 represents the activation function, W1 and W0 represent the weight coefficients of the multi-layer perceptron, and The representation features are compressed by average pooling and maximum pooling respectively, and the output feature vectors are added to ensure that the output vector dimension matches the number of feature frequency bands.

[0036] In one embodiment, the learnable graph convolutional neural network model includes two graph convolution layers, two fully connected layers and a loss function. The method of outputting a predicted label by taking the fused adjacency matrix and the initial features of the frequency band as input includes the following steps: S61: Calculating the normalized Laplacian matrix of the graph according to the initialized adjacency matrix. S62: Update the Laplacian matrix of the graph according to the time domain brain network matrix and the adjacency matrix. S63: Obtaining the characteristics of network self-learning through two layers of graph convolution and the Laplacian matrix of fused time domain information.

[0037] In one embodiment, the normalized Laplacian matrix of the graph is It is expressed as:

[0038]

[0039] Where I∈R N×N is the identity matrix, D∈R N×N Represents the diagonal matrix of A, which is used to describe the connection relationship between each node, that is, the electrode and other electrodes. and are the normalized D and A respectively.

[0040] In one embodiment, the Laplacian matrix of the update graph is represented as:

[0041]

[0042] In one embodiment, the characteristic output of network self-learning is:

[0043]

[0044] Among them, δ2 represents the activation function, which is used to ensure that the value of the output of the lth layer is non-negative, H (l+1) is the input of the l+1th layer and the output of the lth layer. represents the parameter matrix, u represents the feature dimension of the current graph convolutional layer input, and u l Indicates the dimension of the expected output of this layer.

[0045] The present invention also provides an EEG signal recognition system for functional remodeling of movement disorders, the recognition system comprising: a sampling module, a signal processing module, a brain biotopology model and a learnable graph convolutional neural network model.

[0046] The sampling module is used to collect EEG signals of multiple subjects.

[0047] The signal processing module includes a preprocessing unit, a feature extraction unit and a classification unit. The preprocessing unit is used to remove artifact interference of the EEG signal and map the EEG signal to the corresponding frequency band. The feature extraction unit is used to extract the features of the EEG signal. The classification unit is used to classify the EEG signal according to a preset time period.

[0048] The brain biotopology model is used to initialize the graph adjacency matrix according to the time-domain brain functional connectivity of EEG signals, add global channel relationships, and use asymmetric information to simulate information transmission between brains.

[0049] The learnable graph convolutional neural network model includes an input feature acquisition module, a convolution module, a fully connected module and a loss function; the input feature acquisition module includes an updated adjacency matrix acquisition module and a frequency band attention acquisition module; wherein the updated adjacency matrix is ​​obtained by fusing and normalizing the brain function connection matrix and the initialization graph adjacency matrix; the frequency band attention is obtained by enhancing and calibrating the initial frequency domain features; the convolution module includes two convolution layers, and the convolution module is used to train the Laplacian matrix of the fused time domain information; the fully connected layer is used to screen the adjacency matrix; the loss function is used to update the adjacency matrix through back propagation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of the method for recognizing EEG signals for functional remodeling of movement disorders according to Example 1 of the present invention;

[0051] Figure 2 for Figure 1 Framework diagram of the learnable graph convolutional neural network model. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] It should be noted that when a component is referred to as being "mounted on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component at the same time. When a component is considered to be "fixed to" another component, it may be directly fixed on the other component or there may be a central component at the same time.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "or / and" used herein includes any and all combinations of one or more of the related listed items.

[0055] See also Figure 1 and Figure 2 , Figure 1 This is a flow chart of the method for recognizing EEG signals for functional remodeling of movement disorders according to this embodiment; Figure 2 for Figure 1 The framework diagram of the graph convolutional neural network model that can be learned in the present embodiment provides an EEG signal recognition method for functional remodeling of movement disorders. The recognition method comprises the following steps:

[0056] S1: Set the sampling period and obtain the EEG signals of multiple subjects by downsampling. EEG signals are the overall reflection of the electrophysiological activity of brain nerve tissue on the surface of the cerebral cortex. There are two main forms of electrical activity in the cerebral cortex, namely spontaneous EEG activity and induced electrical activity. Spontaneous EEG activity refers to the corresponding brain wave frequency fluctuations that people show autonomously under different emotional states. There are two ways to collect EEG signals, namely embedded collection and non-embedded collection. The embedded collection method refers to opening the brain and implanting sensors inside. However, this method is not suitable for daily use. The non-embedded EEG signal collection method is to collect EEG signals by placing electrode sensors outside the cerebral cortex. In this embodiment, an EEG cap set up by the international standard 10-20 system is used to collect the original EEG signals of S subjects, and the downsampling frequency is 200Hz. There are multiple electrodes in the EEG cap, and the EEG signal is obtained by recording the potential difference of each electrode.

[0057] S2: Preprocess the EEG signal to remove artifact interference. Due to incorrect recording settings during signal acquisition and the good conductivity of the scalp, technical or biological artifacts such as active power line interference, blinking and muscle activity are often mixed with EEG signals, making EEG signal processing difficult. Artifact interference can be removed by a combination of one or more methods such as regression, artifact subtraction, PCA, ICA and wavelet transform. Map multiple EEG signals to a preset frequency band. Randomly distribute multiple EEG signals according to a preset ratio and divide them into a training set and a test set. In this embodiment, the time domain EEG signal is mapped to a frequency band of 1-50Hz.

[0058] The preprocessed EEG signal is expressed as: X Te ∈R N×L , where N represents the number of electrodes and L is the number of samples after time domain sampling.

[0059] S3: Divide the EEG signals in the training set into non-overlapping time windows according to a preset time period, calculate the Pearson coefficient of each window, extract the training data label, and obtain the time domain brain functional connection according to a preset threshold. The calculation method of the time domain brain functional connection includes the following steps:

[0060] S31: Transform the EEG signal into Te According to the data set acquisition paradigm, the test segments are divided into k time groups, each group contains multiple EEG data with a time window of t seconds. The number of samples r in each segment varies according to the length of the time group, so the number of samples L can be expressed as:

[0061]

[0062] Among them, S P represents the sampling rate during preprocessing, Indicates the length of the nth window in the mth segment.

[0063] S32: Calculate the Pearson coefficient of each window based on the EEG data in each time group. The Pearson coefficient is used to characterize the degree of connection between electrodes in the corresponding time period. The expression of the Pearson coefficient is:

[0064]

[0065] Among them, σ represents the standard deviation, E represents the expected function, μ represents the expected value, and v and θ represent the EEG data on different electrodes within the same time range.

[0066] Then the Pearson coefficient is:

[0067] A Per ∈R S*F*N*N ,

[0068] Among them, S is the number of samples, F is the number of frequency bands, and N is the number of electrodes.

[0069] S33: Classify the EEG signal into different categories. Take the average of the Pearson coefficients of the corresponding windows in each category to obtain the training label A MI , A MI ∈R Ca*N*N , where Ca represents the category.

[0070] S34: To A MI Select the threshold, apply the threshold to intercept the TOP-K connections in each category, sum up the filtered matrix, and normalize it to finally obtain the time domain brain functional connection A Te ∈R N*N .

[0071] In this embodiment, the Ca representing the category is 2, that is, the categories are respectively imagining left hand movement and imagining right hand movement. The Pearson coefficient matrix for the EEG signal also retains the connections of the TOP20 strengths to ensure that the connections related to motor control can be selected from them, and a threshold is applied to each category. Finally, the matrices of all categories and all frequency bands obtained by the screen are summed up, and Z-Score normalization is used to obtain the time domain brain functional connectivity.

[0072] S4: Establish a brain biotopological model based on time-domain brain functional connectivity. The robustness of an object in deformation is called topology. The concept of topology can exist in biochemical systems and is used to characterize the robustness of corresponding biochemical processes. According to the brain biotopological model, the connection between EEG channels is obtained and the initial graph adjacency matrix is ​​constructed. The method of constructing the initial graph adjacency matrix includes the following steps:

[0073] S41: Initialize the adjacency matrix. In the spatial domain, the weighted adjacency matrix is ​​initialized with the help of the physical distance of the brain electrodes that collect brain neural activity, which reduces redundant electrode connections and effectively initializes the adjacency matrix. According to the theory that the connection strength between brain regions decays inversely with the physical distance, the graph adjacency matrix A is initialized. i,j Initialization gets:

[0074]

[0075] Among them, the graph adjacency matrix A∈R N×N represents the edge weight relationship between all nodes, i.e., the topological structure of the EEG channel, d i,j Represents the physical distance between electrode i and electrode j. The physical distance between electrodes is calculated from the three-dimensional coordinates of the data marked in the standard lead placement method. η is a sparse parameter. In this embodiment, η is set to 0.2. The connection of TOP20 strength is retained to ensure that the brain connection is reliable. A is constructed as an undirected symmetric matrix and then the local channel relationship is added for initialization.

[0076] S42: Set the global channel relationship to:

[0077] A i,j =A i,j -1

[0078] In this embodiment, 9 pairs of electrodes are selected empirically to construct global channels, such as (FP1, FP2), (AF3, AF4), (F5, F6), (FC5, FC6), (C5, C6), (CP5, CP6), (P5, P6), (PO5, PO6) and (O1, O2), where F represents the frontal lobe, P represents the parietal lobe, C represents the center, O represents the occipital lobe, odd numbers represent the left brain, and even numbers represent the right brain, using asymmetric information to better simulate information transmission between brains.

[0079] S5: intercept the EEG signal within a preset frequency band range, divide the EEG data of each frequency band into samples containing t seconds of EEG signals using a non-overlapping time window of t seconds, and divide the EEG signal into multiple EEG signal sets according to the preset frequency band range; extract the initial features of the EEG signals in each EEG signal set, use the frequency band attention module to enhance the initial feature weights of the key frequency bands on the training set, and select multiple initial features with the highest weight values ​​as the input features of the network nodes.

[0080] In this embodiment, a bandpass filter is used to intercept EEG signals within the frequency range of 1-50Hz, and the EEG data of each channel is divided into samples containing 1s EEG signals using a 1s non-overlapping time window, and the signal is divided into five frequency bands δ (1-3Hz), θ (4-7Hz), α (8-13Hz), β (14-30Hz) and γ (31-50Hz). The initial features of the EEG signals in each frequency band are extracted, and the frequency band attention module is used to improve the sensitivity to the feature information of the key frequency band. The features on the calibrated training set are used as the input features of the network nodes. The input features are expressed as:

[0081]

[0082] Among them, δ1 represents the activation function, which performs nonlinear mapping of the network, W1 and W0 are the weights of the multi-layer perceptron, and In this embodiment, the sigmoid function is selected as the activation function, the input features are compressed by average pooling and maximum pooling respectively, and the adjusted vector is output, and each vector element is added, and the output vector dimension matches the number of feature frequency bands.

[0083] S6: Construct a learnable graph convolutional neural network model based on the input features of the network nodes and the brain biotopology model. The fused adjacency matrix and frequency domain features are sent to the network model to reasonably optimize the electrode relationship, and the EEG data in the test set are used to autonomously learn the identification features of the learnable graph convolutional neural network to obtain the predicted label. Among them, the method for obtaining the predicted label includes the following steps:

[0084] S61: Calculate the normalized Laplacian matrix of the graph based on the initialized adjacency matrix A for:

[0085]

[0086] The first step in graph convolution research is to use the graph Laplacian matrix L = DA to represent the relationship between nodes. N×N is the identity matrix, D∈R N×NRepresents the diagonal matrix of A, that is, D ii =∑ j A ij , used to describe the connection relationship between each node, that is, the electrode and other electrodes, and are the normalized D and A respectively.

[0087] S62: Based on time domain brain functional connectivity A Te And the adjacency matrix A updates the Laplacian matrix of the graph, then the Laplacian matrix of the updated graph can be expressed as:

[0088]

[0089] In this embodiment, The setting range is 0.1 to 0.5. For a single subject, the physical shape of the brain remains stable, that is, the spatial domain information provides a good reference for brain connections. At the same time, this embodiment takes into account the non-stationary characteristics of EEG signals, constructs a time domain functional network to supplement the dynamic information in time sequence, and integrates the time domain information so that the model can reasonably optimize the electrode relationship during the training process.

[0090] S63: After two layers of graph convolution, the characteristics of network self-learning can be obtained by fusing the Laplacian matrix of time domain information. The single-layer output is:

[0091]

[0092] Among them, δ2 represents the activation function, which performs nonlinear mapping of the network and ensures that the value output by the lth layer is non-negative. (l+1) It is the input of the l+1th layer and the output of the lth layer. In this embodiment, the number of graph convolution layers is 2, the output dimension of a single node is 32, and the LeakyRelu function with a parameter of 0.15 is used for activation. The number of fully connected layers is 2, with 128 nodes and 2 task categories, respectively, and the loss is sparsely distributed by a regularization coefficient of 0.01. During training, the model updates the adjacency matrix through backpropagation to optimize the relationship between electrodes and learn the optimal model.

[0093] In step S5, the initial features of each frequency band are extracted, and the frequency band attention module is used to improve the sensitivity of the key frequency band feature information. The features on the calibrated training set are used as the input of the network nodes on the deep neural network classifier. Then, a certain number of training rounds are set for the learnable graph convolutional neural network model constructed in S6, and finally the best model in the training is selected for testing. The EEG signal in the test set is input into the final learnable graph convolutional neural network model, and the obtained test results are used as the corresponding prediction results.

[0094] In order to facilitate the recognition of EEG signals, this embodiment also provides an EEG signal recognition system for movement disorder functional remodeling, and the recognition system includes: a sampling module, a signal processing module, a model generation module and a calculation module.

[0095] The sampling module is used to collect EEG signals of multiple subjects. The collection module may include multiple EEG signal collection devices, such as EEG caps, magnetic resonance equipment, etc. Each subject can collect EEG signals by wearing an EEG cap.

[0096] The signal processing module includes a preprocessing unit, a feature extraction unit and a classification unit. The preprocessing unit is used to remove artifact interference of EEG signals and map EEG signals to corresponding frequency bands. EEG signals are very weak and have small amplitudes, usually at the microvolt level, and are easily interfered by other signals unrelated to EEG activity. These unrelated signals are artifacts. Common artifacts include electrooculographic artifacts (blink artifacts and eye movement artifacts), myoelectric artifacts and electrocardiographic artifacts. Common interferences include DC offset and power frequency interference. During the test phase, artifact interference can be reduced by guiding subjects to practice, cultivating habits, optimizing the environment, and reducing the electrode impedance of EEG signal acquisition equipment. In data processing, artifact interference can be removed by one or more combinations of methods such as regression methods, artifact subtraction, PCA, ICA and wavelet transform. The feature extraction unit is used to extract the features of EEG signals. In this embodiment, the feature extraction unit is mainly used to extract the training data labels of EEG signals and the initial features of each frequency band. The classification unit is used to classify EEG signals according to a preset time period. In this embodiment, the classification of EEG signals includes two parts. One part is to divide the EEG signals mapped to the corresponding frequency band into non-overlapping time windows according to a preset time period, and the other part is to divide the EEG signals into imagined left hand movement and imagined right hand movement according to the training data labels of the EEG signals.

[0097] The brain biotopology model is used to initialize the graph adjacency matrix according to the time-domain brain functional connectivity of the EEG signal to obtain the global channel relationship. The brain biotopology model can be established by the following method: calculate the Pearson coefficient of each window according to the divided non-overlapping time windows, combine the training labels obtained by taking the mean of the training data labels, perform threshold interception on the training labels and summarize and normalize them to obtain brain functional connectivity. Initialize the graph adjacency matrix according to the brain functional connectivity, obtain the global channel relationship, and then establish the brain biotopology model.

[0098] The learnable graph convolutional neural network model is used to obtain the predicted label based on the initial features of each frequency band of the EEG signal and the normalized graph adjacency matrix. The learnable graph convolutional neural network model can be trained by the following methods: the sensitivity to the initial feature information of the frequency band is improved through the frequency band attention module, and the initial features on the calibrated training set are used as the input features of the network nodes. The Laplacian matrix of the graph is calculated according to the initialized adjacency matrix, and the Laplacian matrix of the graph is updated through the brain biotopology model. The input features are passed through two layers of graph convolution. The learnable graph convolutional neural network model can be trained by integrating the Laplacian matrix of the time domain information to obtain the final learnable graph convolutional neural network model.

[0099] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for those of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for recognizing EEG signals for functional remodeling of movement disorders, which is used to obtain EEG signals of multiple brain regions of each subject according to the motor imagination of each subject, and then establish a learnable graph convolutional neural network model based on the characteristics of the EEG signals to recognize the EEG signals of the subject and predict the subject's motor intention; characterized in that: It includes the following steps: S1: Set the sampling period to collect multiple EEG signals from different brain regions of multiple subjects in real time; S2: After pre-processing the plurality of EEG signals to remove artifact interference, the plurality of EEG signals are respectively mapped to preset frequency bands; Extract multiple EEG signals in each frequency band as a training set according to a preset ratio; S3: Divide the EEG signals in the training set into multiple non-overlapping time windows according to a preset time period; extract the training labels of the EEG signals in each time window; perform threshold screening on the training labels and normalize them to obtain time-domain brain functional connectivity; the brain functional connectivity is used to characterize the degree of connection between different brain regions in the same time window; S4: establishing a brain biotopological model according to the time-domain brain functional connectivity and the physical distance of the electrodes in the space domain; obtaining the connection between EEG channels according to the brain biotopological model, and constructing an initial graph adjacency matrix; wherein the method for constructing the initial graph adjacency matrix comprises the following steps: S41: The graph adjacency matrix is ​​initialized based on the theory that the connection strength between brain regions decays inversely with the physical distance; the initialized graph adjacency matrix is ​​expressed as: Among them, d i,j represents the physical distance between electrode i and electrode j, A i,j is the value of the i-th row and j-th column in the graph adjacency matrix A; η is a sparse parameter; S42: Setting a global channel relationship according to the initialized graph adjacency matrix; the global channel relationship is used to simulate information transmission between brains according to asymmetric information; the global channel relationship is expressed as: A i,j =A i,j -1; S5: intercepting multiple EEG signals within a preset frequency band, dividing the EEG data of each frequency band into samples containing t seconds of EEG signals using a non-overlapping time window of t seconds, and dividing the EEG signals into multiple EEG signal sets according to the preset frequency band range; extracting the initial features of the EEG signals in each EEG signal set, enhancing the weights of the initial features of the key frequency bands on the training set, and selecting multiple initial features with the highest weight values ​​as input features of the network nodes; S6: Construct a learnable graph convolutional neural network model based on the input features of the network nodes and the brain biotopology model; input the fused adjacency matrix and frequency domain features into the learnable graph convolutional neural network to reasonably optimize the electrode relationship, and the learnable graph convolutional neural network autonomously learns the identification features to obtain the prediction label; In S6, the learnable graph convolutional neural network model includes two graph convolutional layers, two fully connected layers and a loss function; the method of taking the fused adjacency matrix and the calibrated frequency band features as input and outputting the predicted label includes the following steps: S61: Calculate the normalized Laplacian matrix of the graph according to the initialized adjacency matrix; S62: Update the Laplacian matrix of the graph according to the time domain brain network matrix and the adjacency matrix; S63: Obtain the characteristics of network self-learning through two layers of graph convolution and Laplacian matrix integrating time domain information; In S61, the normalized Laplacian matrix of the graph is It is expressed as: Where I∈R n×n is the identity matrix, D∈R n×n Represents the diagonal matrix of A, which is used to describe the connection relationship between each node, that is, the electrode and other electrodes. and are the normalized D and A respectively; In S62, the Laplacian matrix of the update graph is expressed as: in, Indicates the proportion of the Laplacian matrix of the graph before each iteration in updating the Laplacian matrix of the graph; In S63, the characteristic output of the network self-learning is: Among them, δ2 represents the activation function, which is used to ensure that the value of the output of the lth layer is non-negative, H (l+1) It is the input of the l+1th layer and the output of the lth layer; represents the parameter matrix, u represents the feature dimension of the current graph convolutional layer input, and u l Represents the dimension of the expected output of layer l.

2. The method for recognizing EEG signals for functional remodeling of movement disorders according to claim 1, characterized in that: In S3, the method for obtaining time-domain brain functional connectivity includes the following steps: S31: dividing the EEG signals in the corresponding frequency band into multiple time groups according to the data set acquisition paradigm, each time group contains multiple EEG data with a time window of t seconds; S32: Calculate the Pearson coefficient of each window according to the EEG data in each time group. The Pearson coefficient is expressed as: A Per ∈R S*F*N*N , Among them, S is the number of samples, F is the number of frequency bands, and N is the number of electrodes; S33: Divide the EEG signal into different categories; take the average of the Pearson coefficients of the corresponding windows in each category to obtain the training label; S34: Threshold selection is performed on the training labels of each category to obtain multiple matrices; the multiple matrices are summed and normalized to obtain the time domain brain functional connectivity A Te ∈R N*N .

3. The method for recognizing EEG signals for functional remodeling of movement disorders according to claim 2, characterized in that: In S31, the EEG signal X Te According to the data set acquisition paradigm, the test segments are divided into k time groups, each group contains multiple EEG data with a time window of t seconds. The number of samples in each segment r varies according to the length of the time group. The preprocessed EEG signal X Te It is expressed as: X Te ∈R N×L Where N represents the number of electrodes, L represents the number of samples after time domain sampling, and the number of samples L is expressed as: Among them, S P represents the sampling rate during preprocessing, r m is the number of samples in the mth segment, Indicates the length of the nth window in the mth segment.

4. The method for recognizing EEG signals for functional remodeling of movement disorders according to claim 2, characterized in that: In S32, the Pearson coefficient is calculated as follows: Among them, σ represents the standard deviation, E represents the expected function, μ represents the expected value, and v and θ represent the EEG data on different electrodes within the same time range.

5. The method for recognizing EEG signals for functional remodeling of movement disorders according to claim 1, characterized in that: In S5, each frequency band is used as a feature detector, and the frequency band attention module is used to enhance the initial feature weights of the key frequency bands. The input feature is expressed as: Among them, δ1 represents the activation function, W1 and W0 represent the two-layer weight coefficients of the multi-layer perceptron, and The representation features are compressed by average pooling and maximum pooling respectively, and the output feature vectors are added to ensure that the output vector dimension matches the number of feature frequency bands.

6. An EEG signal recognition system for functional remodeling of movement disorders, which adopts the motor imagery EEG signal recognition method as claimed in any one of claims 1 to 5, characterized in that: It includes: A sampling module, which is used to collect EEG signals of multiple subjects; A signal processing module, which includes a preprocessing unit, a feature extraction unit and a classification unit; The preprocessing unit is used to remove artifact interference of the EEG signal and map the EEG signal to a corresponding frequency band; The feature extraction unit is used to extract the features of the EEG signal; the classification unit is used to classify the EEG signal according to a preset time period; Brain biotopology model, which is used to initialize the graph adjacency matrix based on the time-domain brain functional connectivity of EEG signals, add global channel relationships, and use asymmetric information to simulate information transmission between brains; A learnable graph convolutional neural network model includes an input feature acquisition module, a convolution module, a fully connected module and a loss function; the input feature acquisition module includes an updated adjacency matrix acquisition module and a frequency band attention acquisition module; wherein the updated adjacency matrix is ​​obtained by fusing and normalizing the brain function connection matrix and the initialization graph adjacency matrix; the frequency band attention is obtained by enhancing and calibrating the initial frequency domain features; the convolution module includes two convolution layers, and the convolution module is used to train the Laplacian matrix of fused time domain information; the fully connected layer is used to screen the adjacency matrix; the loss function is used to update the adjacency matrix through back propagation.

Images