An epilepsy electroencephalogram classification method based on a multi-feature multi-relation graph convolutional neural network

Abstract

The application belongs to the technical field of communication and relates to a kind of epilepsy electroencephalogram classification method based on multi-feature multi-relation graph convolutional neural network, comprising: reading electroencephalogram data, using graph to describe 10-second electroencephalogram slice data characteristics as input data;Graph convolution is carried out on the input data to generate node-level embedding;The node embedding of the whole graph is subjected to a pooling operation, and a graph embedding is returned;The graph-level embedding is input into a fully connected layer, the output value of the model is compared to determine the result, and then the category, i.e., the result, is output;The spatial similarity and spectral similarity between the electroencephalogram channels are extracted, and the two channel similarities are fused as the edge relationship matrix between the nodes of the whole graph, which can mine and utilize the correlation information between different channels of electroencephalogram, extract multi-feature combination, and has high classification accuracy and good accuracy.

Description

Technical fields:

[0001] This invention belongs to the field of communication technology and relates to an epilepsy EEG classification method based on a multi-feature multi-relation graph convolutional neural network. It can mine and utilize the correlation information between different channels of EEG to classify and judge EEG data, with high accuracy and precision. Background technology:

[0002] The classification of epilepsy EEG data generally relies on interpretation by professionals, such as doctors. However, this manual interpretation is time-consuming and inefficient. Furthermore, due to varying technical quality and doctor qualifications, different people may obtain different interpretations of the same epilepsy EEG data, often resulting in misinterpretations and low accuracy. With the application of communication technology in medicine, traditional machine learning methods have become widely used. These methods combine the neurophysiological features of five language regions (three frontal lobes and two temporal lobes) from two hemispheres of multiple subjects into multiple feature subsets, and use the extreme gradient boosting algorithm to classify these feature subsets, achieving good classification results. In recent years, with the successful application of deep learning in image classification and speech recognition, scholars have proposed a deep learning-based system—a pyramid-shaped one-dimensional convolutional neural network (P-1D-CNN) model—to extend deep learning to graph structures for the classification and prediction of brain diseases. However, current techniques lack sufficient depth in mining the spatial information across multiple EEG channels and extract too few feature combinations, leading to poor classification accuracy.

[0003] Therefore, it is necessary to design an epilepsy EEG classification method based on a multi-feature, multi-relation graph convolutional neural network. This method extracts four time-frequency domain features and two nonlinear dynamic features from epilepsy EEG signals. By concatenating the features at the subject slice dimension as input features for the model, spatial and spectral similarities between EEG channels are extracted. The two channel similarities are then fused into an edge relationship matrix between the nodes of the overall graph. This method can mine and utilize the correlation information between different channels of the EEG, achieving a high average classification accuracy for EEG data from different epilepsy patients, which is superior to traditional machine learning algorithms and single-feature, single-relation graph convolutional neural networks. Summary of the Invention:

[0004] The purpose of this invention is to overcome the shortcomings and deficiencies of existing technologies and to design a method for classifying epilepsy EEG based on multi-feature multi-relation graph convolutional neural networks. This method aims to solve the problems of insufficient spatial information mining between multiple channels of EEG, too few extracted feature combinations, and low classification accuracy and poor judgment in existing technologies.

[0005] To achieve the above objectives, the present invention provides an epilepsy EEG classification method based on a multi-feature, multi-relationship graph convolutional neural network, comprising: reading EEG data; using a graph to describe the features of a 10-second EEG slice as input data; performing graph convolution on the input data to generate node-level embeddings; performing pooling operations on the node embeddings of the entire graph to return a graph embedding; passing the graph-level embedding as input through a fully connected layer; comparing the results with the model's output values ​​to determine the outcome; and then outputting the category, i.e., the determination result.

[0006] The read EEG data includes two parts. The first part is multiple feature data extracted from 16 bipolar epileptic EEG channel pairs composed of partial EEG channel sets. The extracted feature matrices are selected and fused at the feature layer to form the final graph node data. The second part is the correlation coefficient between epileptic EEG channels, that is, through the mapping relationship, the hidden spatial relationship between channels is shown in the form of a graph, and its data dimension is 16×16.

[0007] Furthermore, after processing the EEG data recorded through L brain electrode channels, a total of M brain electrode pairs were selected as the raw data for the test. This raw data was then divided into N slices, each 10 seconds long. Thus, the raw data was divided into N M-channel EEG data sets, each 10 seconds long, with the slice becoming the smallest unit of data. A graph G is defined. i =(V,E) i Let represent the interaction relationships between channels at slice i. This dynamic graph is a fully connected graph. Here, V represents the set of M (M = |V|) channels, which can be represented by the characteristic matrix X. i ∈R m×d Let represent , where d represents the input feature dimension of each channel and m represents the number of channels. The set of undirected edges between node sets V can be represented by a weighted adjacency matrix A∈R. m×m Let A represent this. Each node has a self-connecting edge A. ii =1; i = 1, 2, ..., m.

[0008] Select several original brain electrode channels as channel set c, and then select channels c from the channel set. i and channel c j The potential difference of the electrodes serves as a new bipolar channel C k This forms a new bipolar channel set C. Multiple time-frequency features and nonlinear dynamic features are extracted from the channel set and concatenated on a channel-by-channel basis to form the channel's feature matrix F. Node k in the network uses channel C... k The characteristic matrix F k To express.

[0009] In the research of graph neural networks, adjacency matrices are used to represent the relationships between vertices, and the relationship A between channel i and channel j is designed. ij It consists of two parts, namely the spatial distance relationship between channel i and channel j, denoted as And the spectral correlation between channel i and channel j, denoted as

[0010] For spatial correlation, based on the positional information of each channel in the 10-10 international standard lead system, the specific coordinates of the channel in the Cartesian coordinate system are obtained. In Cartesian coordinate space, a point (x...) on a sphere of radius r... i ,y i ,z i ) and point (x) j ,y j ,z j Geodesic distance between ) Defined as:

[0011]

[0012] Regarding spectral correlation, according to Guido Nolte et al. [6] Let x i (f) and x j (f) are the time series of channels i and j, respectively. and The Fourier transform of the cross spectrum. Define the cross spectrum as...

[0013] The spectral correlation between channel i and channel j is defined as the absolute value of the normalized cross spectrum.

[0014]

[0015] The fundamental purpose of convolution in spatial graph convolutional neural networks is to extract spatial features from topological graphs. In each convolutional layer, the feature update of each node is based on its neighboring nodes. The features of all its neighbors are weighted and summed to obtain the node features of the new layer. [7] The data update rules for convolutional layers are as follows:

[0016]

[0017] in:

[0018] H (l) H represents the eigenvector matrix of all stages at layer l. (l+1) This represents the feature vector matrix of the next layer after one convolution operation; Where I is the identity matrix, and A represents graph G. iThe adjacency matrix of A, where D is the degree matrix of A, is calculated using the following method. σ is a non-linear activation function, such as the ReLU function; W (l) This represents the trainable parameter matrix of the current layer's convolutional transformation.

[0019] Furthermore, after multiple rounds of training, the program automatically obtains the optimal classification threshold. This threshold serves as the classification standard, and by comparing the output value of the model, it determines whether the subject's EEG data belongs to the category of epilepsy or non-epilepsy.

[0020] A multi-feature, multi-relation graph convolutional neural network (GCNN) is proposed for classifying epilepsy EEG signals. The time-frequency domain features and nonlinear dynamic features of the EEG signals are extracted separately. These features are then concatenated at the subject slice dimension as input features to the model. Spatial and spectral similarities between EEG channels are extracted, and these two channel similarities are fused into an edge relation matrix between overall graph nodes. Experiments were conducted on a dataset, and the results show that this model can effectively mine and utilize the correlation information between different EEG channels, achieving a high average classification accuracy for EEG data from different epilepsy patients, outperforming traditional machine learning algorithms and single-feature, single-relation graph convolutional neural networks.

[0021] An epilepsy EEG classification device based on a multi-feature, multi-relationship graph convolutional neural network comprises: an EEG data acquisition unit, a data processor, and a display. The EEG data acquisition unit and the data processor are electrically connected, as are the data processor and the display. The EEG data acquisition unit has a cap-like structure with electrodes evenly distributed on its top. These electrodes are used to acquire the subject's EEG signals. This acquisition process is conducted naturally, without the need for any harmful stimulation. The electrodes of the EEG data acquisition unit are electrically connected to the data processor. The data processor contains a program model that reads data from the brain electrode channels and the data between channels as input data. It then processes, slices, and extracts feature data from the EEG data. The program automatically derives the optimal classification threshold as a comparison classification standard. By comparing the model output values ​​of the dataset, it determines whether the subject in the dataset belongs to the epilepsy or non-epilepsy category, and displays the classification results on the display.

[0022] Compared with existing technologies, this invention extracts spatial and spectral similarities between EEG channels and integrates the two types of channel similarities as an edge relationship matrix between nodes in the overall graph. It can mine and utilize the correlation information between different channels of EEG, extract multi-feature combinations, and achieve high accuracy and precision in classification. Attached image description:

[0023] Figure 1 This is a schematic diagram of the working structure principle of the present invention.

[0024] Figure 2 This is a schematic diagram illustrating the principle of EEG data reading in this invention.

[0025] Figure 3 This is a schematic diagram illustrating the principle of extracting and splicing features from sliced ​​data according to the present invention.

[0026] Figure 4 The ROC curves of the E-GCNN and the single relation graph convolutional neural network of the present invention are shown.

[0027] Figure 5 The ROC curves for E-GCNN and single feature map convolutional neural network of the present invention are shown. Detailed implementation method:

[0028] The present invention will be further described below with reference to embodiments and accompanying drawings:

[0029] Example 1:

[0030] This embodiment describes an epilepsy EEG classification method based on a multi-feature, multi-relationship graph convolutional neural network, comprising: reading EEG data; using a graph to describe the features of a 10-second EEG slice as input data; performing graph convolution on the input data to generate node-level embeddings; performing pooling operations on the node embeddings of the entire graph to return a graph embedding; passing the graph-level embedding as input through a fully connected layer; comparing the results with the model's output values ​​to determine the outcome; and then outputting the category, i.e., the determination result.

[0031] The read EEG data includes two parts. The first part is multiple feature data extracted from 16 bipolar epileptic EEG channel pairs composed of partial EEG channel sets. The extracted feature matrices are selected and fused at the feature layer to form the final graph node data. The second part is the correlation coefficient between epileptic EEG channels, that is, through the mapping relationship, the hidden spatial relationship between channels is shown in the form of a graph, and its data dimension is 16×16.

[0032] Furthermore, after processing the EEG data recorded through L brain electrode channels, a total of M brain electrode pairs were selected as the raw data for the test. This raw data was then divided into N slices, each 10 seconds long. Thus, the raw data was divided into N M-channel EEG data sets, each 10 seconds long, with the slice becoming the smallest unit of data. A graph G is defined. i =(V,E) i Let represent the interaction relationships between channels at slice i. This dynamic graph is a fully connected graph. Here, V represents the set of M (M = |V|) channels, which can be represented by the characteristic matrix X. i ∈R m×d Let represent , where d represents the input feature dimension of each channel and m represents the number of channels. The set of undirected edges between node sets V can be represented by a weighted adjacency matrix A∈R. m×m Let A represent this. Each node has a self-connecting edge A. ii =1; i = 1, 2, ..., m.

[0033] Select several original brain electrode channels as channel set c, and then select channels c from the channel set. i and channel c j The potential difference of the electrodes serves as a new bipolar channel C k This forms a new bipolar channel set C. Multiple time-frequency features and nonlinear dynamic features are extracted from the channel set and concatenated on a channel-by-channel basis to form the channel's feature matrix F. Node k in the network uses channel C... k The characteristic matrix F k To express.

[0034] In the research of graph neural networks, adjacency matrices are used to represent the relationships between vertices, and the relationship A between channel i and channel j is designed. ij It consists of two parts, namely the spatial distance relationship between channel i and channel j, denoted as And the spectral correlation between channel i and channel j, denoted as

[0035] For spatial correlation, based on the positional information of each channel in the 10-10 international standard lead system, the specific coordinates of the channel in the Cartesian coordinate system are obtained. In Cartesian coordinate space, a point (x...) on a sphere of radius r... i ,y i ,z i ) and point (x) j ,y j ,z j Geodesic distance between ) Defined as:

[0036]

[0037] Regarding spectral correlation, according to Guido Nolte et al. [6] Let x i (f) and x j (f) are the time series of channels i and j, respectively. and The Fourier transform of the cross spectrum. Define the cross spectrum as...

[0038] The spectral correlation between channel i and channel j is defined as the absolute value of the normalized cross spectrum.

[0039]

[0040] The fundamental purpose of convolution in spatial graph convolutional neural networks is to extract spatial features from topological graphs. In each convolutional layer, the feature update of each node is based on its neighboring nodes. The features of all its neighbors are weighted and summed to obtain the node features of the new layer. [7] The data update rules for convolutional layers are as follows:

[0041]

[0042] in:

[0043] H (l) H represents the eigenvector matrix of all stages at layer l. (l+1) This represents the feature vector matrix of the next layer after one convolution operation; Where I is the identity matrix, and A represents graph G. i The adjacency matrix of A, where D is the degree matrix of A, is calculated using the following method. σ is a non-linear activation function, such as the ReLU function; W (l) This represents the trainable parameter matrix of the current layer's convolutional transformation.

[0044] Furthermore, after multiple rounds of training, the program automatically obtains the optimal classification threshold. This threshold serves as the classification standard, and by comparing the output value of the model, it determines whether the subject's EEG data belongs to the category of epilepsy or non-epilepsy.

[0045] All technical contents not disclosed in this embodiment adopt existing technologies. The data read can be information collected and stored in advance or information obtained by directly connecting to the brain data acquisition device. No stimulation or other operations harmful to the human body are performed.

[0046] Example 2:

[0047] This embodiment verifies the epilepsy EEG classification method based on multi-feature multi-relation graph convolutional neural network:

[0048] Dataset Selection: The dataset used is a subset of the EEG database, containing EEG information from 237 subjects. This includes 1360 EEG signal recordings from 133 epilepsy patients during seizures and 288 EEG signal recordings from 104 subjects at rest, totaling 1648 EEG signal files. Three unipolar montage recording references were used during data acquisition: 01_tcp_AR, 02_tcp_LE, and 03_tcp_AR_a. All samples in this dataset are stored in EDF format, with 31 EEG channels being the most common number in the EDF files. Most EEG data were sampled at 250Hz, with a few at 256Hz, 400Hz, and 512Hz.

[0049] Preprocessing of the dataset: To balance positive and negative samples, 1602 EEG recordings from 195 subjects were selected and retained. For each recording, 16 original channels were selected from all channels according to the three reference rules used when the data was collected (01_tcp_AR, 02_tcp_LE, 03_tcp_AR_a): ["F7","F3","F8","F4","T3","C3","T4","C4","T5","P3","T6","P4","O1","O2","FP1","FP2"]. The potential difference between two channel electrodes in the channel queue was used as a new channel, called a bipolar channel, to form a new channel set: ["O1-T5"]. [,"T5-T3","T3-F7","F7-FP1","O1-P3","P3-C3","C3-F3","F3-FP1","O2-P4","P4-C4","C4-F3","F4-FP2","O2-T6","T6-T4","T4-F8","F8-Fp2"]; All record files are divided into several independent slices in 10-second increments. That is, each record file is divided into several small record files of 10 seconds each that do not overlap. The remaining portions of less than 10 seconds are cleared. A total of 23,331 windows are generated. The data in each slice consists of the newly defined set of bipolar channel pairs. The detailed data distribution is shown in Table 1.

[0050] Table 1. Data distribution of subjects and subject slides

[0051]

[0052] Finally, the MNE-Python toolkit was used to resample each slice to 250Hz, and noise was removed by using a 1Hz high-pass filter and a 50Hz notch filter.

[0053] Feature engineering extraction:

[0054] For the preprocessed slice data, power spectral density, wavelet energy sum, wavelet energy ratio, wavelet energy entropy, approximate entropy, and Shannon entropy features are extracted from each bipolar channel to obtain a feature matrix of 6 single features. These 6 feature matrices are concatenated into a window feature matrix f, with each slice channel as the unit. The slice matrix f is compressed into a one-dimensional matrix, and all the compressed one-dimensional slice feature matrices are concatenated to form the total feature matrix F. The total feature matrix F is standardized to eliminate the differences caused by different features and reduce the impact on the program model.

[0055] Comparison and classification: The program automatically derives the optimal classification threshold, which serves as the comparison and classification standard. By comparing the model output values ​​of the dataset, it determines whether the subjects in the dataset belong to the category of epilepsy or non-epilepsy.

[0056] Classification Results Comparison: The classification results of the epilepsy EEG classification method based on multi-feature, multi-relationship graph convolutional neural network were compared with those of graph neural networks considering only a single feature and a single edge relationship. Accuracy, Precision, Recall, F1 score, and AUC (Area Under Curve) were used as evaluation metrics for comprehensive evaluation. Each model evaluation metric was at the subject level, meaning that for each subject's EEG data, the predicted values ​​of different slices were averaged as the output value of the program model after processing the subject's EEG data. Table 2 shows the performance comparison data between E-GCNN and the single-relationship graph model. The E-GCNN model based on graph convolutional neural network achieved an average accuracy of 87% and an AUC of 90% after 10-fold cross-validation. It also generally outperformed traditional machine learning algorithms in other metrics such as F1 score. As shown in Table 2, focusing on the node correlation aspect of graph neural networks, the E-GCNN model, which combines spatial and spectral correlation between channels, outperforms the E-GCNN-Dist model (using only spatial correlation) and the E-GCNN-Spec model (using spectral correlation), respectively. Average F1 score and accuracy are improved by 1%-5%. Table 3 shows a comparison of E-GCNN performance with single-feature models. E-GCNN, which combines multiple EEG features as input features to the final model, also outperforms single-feature graph neural networks, with improvements generally around 10%-20%, and even reaching 30% for some features. The receiver operating characteristic (ROC) curves of E-GCNN and single-relationship graph convolutional neural networks are shown below. Figure 4 As shown, the ROC curves of E-GCNN and single feature map convolutional neural networks are as follows: Figure 5 As shown, the AUC of E-GCNN is significantly higher than the other two, thus demonstrating that the graph structure with multiple relationships is better able to characterize the potential relationships between epileptic EEG channels.

[0057] Table 2. Performance of E-GCNN and Single Relationship Graph Model

[0058]

[0059] Table 3. Performance of E-GCNN and Single Feature Model

[0060]

[0061]

[0062] Example 3:

[0063] This embodiment relates to an epilepsy EEG classification device based on a multi-feature, multi-relationship graph convolutional neural network. Its main structure includes an EEG data acquisition unit, a data processor, and a display. The EEG data acquisition unit and the data processor are electrically connected, as are the data processor and the display. The EEG data acquisition unit has a cap-like structure with electrodes evenly distributed on its top. These electrodes are used to acquire the subject's EEG signals. This acquisition process is conducted in a natural state, without any stimulation or harmful operations. The electrodes of the EEG data acquisition unit are electrically connected to the data processor. The data processor contains a program model that reads brain electrode channel data and data between channels as input data. It then processes, slices, and extracts feature data from the EEG data. The program automatically derives the optimal classification threshold as a comparison classification standard. By comparing the model output values ​​of the dataset, it determines whether the subject in the dataset belongs to the epilepsy or non-epilepsy category, and displays the classification results on the display.

[0064] The data processor can be a computer, smartphone, or PLC controller, or other device capable of storing and processing data.

[0065] The data processor program model uses the epilepsy EEG classification method based on multi-feature multi-relation graph convolutional neural network described in Example 1 to process and classify EEG data.

Claims

1. A brainwave classification method based on a multi-feature, multi-relationship graph convolutional neural network, characterized in that... The method includes: reading EEG data, using a graph to describe the features of a 10-second EEG slice as input data; performing graph convolution on the input data to generate node-level embeddings; performing pooling on the node embeddings of the entire graph to return a graph embedding; passing the graph-level embedding as input through a fully connected layer, judging the result by comparing it with the model's output value, and then outputting the category; the input data includes two parts, the first part being 16 channels composed of a partial set of EEG channels (ci). The first part extracts multiple time-frequency features and nonlinear dynamic features from the channel cj pair. The extracted feature matrices are then fused and stitched together at the feature layer to form the final graph node data. The second part is the correlation coefficient between epileptic EEG channels, which represents the hidden spatial relationships between channels in the form of a graph through mapping relationships. Its data dimension is 16×16. After processing the EEG data recorded through L brain electrode channels, a total of M brain electrode pairs are selected as the raw data for testing. This raw data is divided into N slices, each 10 seconds long. Thus, the raw data is divided into N M-channel EEG data with a duration of 10 seconds, and the slice becomes the smallest data unit. A graph is defined. , indicating slice The graph shows the interactions between channels; it is a fully connected graph. Here, V represents a set of M channels, represented by the characteristic matrix. The expression is given by , where d represents the dimension of the input features for each channel and m represents the number of channels; The set of undirected edges between node sets V is represented by a weighted adjacency matrix. Let represent the self-connecting edges of each node. ; Select several primitive brain electrode channels as a channel set. channels in the channel set and channels The potential difference of the electrodes serves as a new bipolar channel. To form a new set of bipolar channels Extract multiple time-frequency features and nonlinear dynamic features from the channel set, and concatenate them channel by channel to form the channel's feature matrix. Node k in the network uses channels Feature matrix To indicate; In the research of graph neural networks, adjacency matrices are used to represent the relationships between vertices, and channels are designed. With channel Relationship It consists of two parts, namely the channel With channel The spatial distance relationship between them is denoted as ; and channels With channel The spectral correlation between them is denoted as ; For spatial correlation, based on the positional information of each channel in the 10-10 international standard lead system, the specific coordinate values ​​of the channels in the Cartesian coordinate system are obtained; in the Cartesian coordinate space, a point on a sphere with radius r... With point Geodesic distance between Defined as: For spectral correlation, let and Channels and channels time series and Fourier transform; define the cross spectrum as aisle and channels The spectral correlation is defined as the absolute value of the normalized cross spectrum. The fundamental purpose of convolution in spatial graph convolutional neural networks is to extract spatial features from topological graphs. In each convolutional layer, the feature update of each node is based on its neighboring nodes. The features of all its neighbors are weighted and summed to obtain the node features of the new layer. The data update rules of the convolutional layer are as follows: in: Indicates when all stages are The eigenvector matrix of the layer, This represents the feature vector matrix of the next layer after one convolution operation; ,in It is the identity matrix. Representation diagram The adjacency matrix, yes Degree matrix, calculation method It is a non-linear activation function, such as function; This represents the trainable parameter matrix of the current layer's convolutional transformation. Time-frequency domain features and nonlinear dynamic features of epileptic EEG signals are extracted separately. These features are then concatenated using subject slice dimensions as input features for the EEG classification method. Experiments on the dataset demonstrate that this model can mine and utilize the correlation information between different channels of EEG, improving the average classification accuracy for different EEG data.

2. An EEG classification device based on a multi-feature, multi-relationship graph convolutional neural network, the main structure of which includes: The system comprises an EEG data acquisition device, a data processor, and a display. Electrical connections exist between the EEG data acquisition device and the data processor, as well as between the data processor and the display. The EEG data acquisition device has a cap-like structure with electrodes evenly distributed on its top. These electrodes are used to acquire the subject's brainwave signals. The acquisition process is conducted in a natural state, without any harmful stimulation. Electrical connections exist between the electrodes of the EEG data acquisition device and the data processor. The data processor contains a program model that reads data from the brain electrode channels and the data between channels as input data. The EEG data is processed, sliced, and feature data extracted. The program automatically determines the optimal classification threshold as the comparison classification standard. The model output value of the comparison dataset is used to classify the results, and the classification results are displayed on the monitor. The program model in the data processor processes the EEG data to be tested recorded through L brain electrode channels, and selects M brain electrode pairs as the raw data for the test. These raw data are divided into N slices with a duration of 10 seconds each. In this way, the raw data is divided into N M-channel EEG data with a duration of 10 seconds, and the slice becomes the smallest data unit. A graph is defined. , indicating slice The graph shows the interactions between channels; it is a fully connected graph. Here, V represents a set of M channels, represented by the characteristic matrix. The expression is given by , where d represents the dimension of the input features for each channel and m represents the number of channels; The set of undirected edges between node sets V is represented by a weighted adjacency matrix. Let represent the self-connecting edges of each node. ; Select several primitive brain electrode channels as a channel set. channels in the channel set and channels The potential difference of the electrodes serves as a new bipolar channel. To form a new set of bipolar channels Extract multiple time-frequency features and nonlinear dynamic features from the channel set, and concatenate them channel by channel to form the channel's feature matrix. Node k in the network uses channels Feature matrix To indicate; In the research of graph neural networks, adjacency matrices are used to represent the relationships between vertices, and channels are designed. With channel Relationship It consists of two parts, namely the channel With channel The spatial distance relationship between them is denoted as ; and channels With channel The spectral correlation between them is denoted as ; For spatial correlation, based on the positional information of each channel in the 10-10 international standard lead system, the specific coordinate values ​​of the channels in the Cartesian coordinate system are obtained; in the Cartesian coordinate space, a point on a sphere with radius r... With point Geodesic distance between Defined as: For spectral correlation, let and Channels and channels time series and Fourier transform; define the cross spectrum as aisle and channels The spectral correlation is defined as the absolute value of the normalized cross spectrum. The fundamental purpose of convolution in spatial graph convolutional neural networks is to extract spatial features from topological graphs. In each convolutional layer, the feature update of each node is based on its neighboring nodes. The features of all its neighbors are weighted and summed to obtain the node features of the new layer. The data update rules of the convolutional layer are as follows: in: Indicates when all stages are The eigenvector matrix of the layer, This represents the feature vector matrix of the next layer after one convolution operation; ,in It is the identity matrix. Representation diagram The adjacency matrix, yes Degree matrix, calculation method , It is a non-linear activation function, such as function; This represents the trainable parameter matrix of the current layer's convolutional transformation; We extracted the time-frequency domain features and nonlinear dynamic features of epileptic EEG signals, and used the sliced ​​features as input features for the EEG classification method. Experiments on the dataset showed that this model can mine and utilize the correlation information between different channels of EEG, and improve the average classification accuracy of different EEG data.

3. The EEG classification device based on a multi-feature, multi-relationship graph convolutional neural network according to claim 2, characterized in that... The data processor is a device that can store and process data, such as a computer, smartphone, or PLC controller.

Images