A method for extracting EEG features based on temporal graph structure learning

Abstract

This invention discloses a method for extracting EEG features based on time-series graph structure learning, comprising: acquiring EEG signal data; preprocessing the EEG signals and dividing them into signal slices of consistent duration; dividing the data into training and testing sets; for each sample, extracting frequency domain features in three-dimensional matrix form and power spectrum features in one-dimensional vector form, converting to logarithmic space and standardizing; inputting the frequency domain features in three-dimensional matrix form into a grouped time-series convolutional neural network to extract node features; using the node features for graph structure learning; using the power spectrum features in one-dimensional vector form as node attributes to calculate the graph representation of the obtained graph structure; inputting the graph representation into a classifier, calculating the loss function, training and evaluating the model performance. This invention can simultaneously handle the time-varying nature of single-channel EEG signals and the time-varying nature of the connectivity between channels, achieving good performance in classifying interictal and preictal EEG signals in epilepsy.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal processing, and in particular to an EEG feature extraction method based on time-series graph structure learning. Background Technology

[0002] Electroencephalography (EEG) signals are the overall reflection of the electrophysiological activity of brain nerve tissue on the surface of the cerebral cortex. EEG feature extraction technology has important research and clinical significance in the diagnosis and treatment of neurological diseases such as epilepsy and Parkinson's disease.

[0003] Currently, EEG feature extraction methods are mainly divided into two categories: feature extraction methods designed manually by experts and feature extraction methods using deep learning models. Deep learning models that have been successfully applied to EEG feature extraction include fully connected neural networks and convolutional neural networks.

[0004] For example, Chinese patent document CN110263606A discloses a method for extracting and classifying scalp EEG features based on an end-to-end convolutional neural network; Chinese patent document CN117435052A discloses a data processing method and device for brain-computer interfaces.

[0005] However, because EEG signals are multi-channel, time-varying signals, there are correlations between the signals in different channels. These correlations are not only specific to different individuals, but also change over time even within the same individual. Current feature extraction methods struggle to effectively model these correlations, and the accuracy of EEG signal classification tasks based on existing feature extraction methods, such as classifying interepidemic and preepidemic EEG signals, still has room for improvement. Summary of the Invention

[0006] This invention provides a method for extracting EEG features based on time-series graph structure learning, which can simultaneously handle the time-varying nature of single-channel EEG signals and the time-varying nature of the connectivity between channels, and has good performance in classifying EEG signals during the interictal and preictal phases of epilepsy.

[0007] A method for extracting EEG features based on temporal graph structure learning includes the following steps:

[0008] (1) Extract EEG signal data for preprocessing. Use the sliding window method to divide the EEG signal into signal slices of consistent duration, and divide the signal slice data into training set and test set;

[0009] (2) For each signal slice data in the training set and the test set, extract the frequency domain features in the form of a three-dimensional matrix and the power spectrum features in the form of a one-dimensional vector; wherein, the power spectrum features in the form of a one-dimensional vector are the power spectrum features in the form of a one-dimensional vector formed by splicing the power spectrum features of each channel and each frequency.

[0010] (3) Transform all extracted frequency domain features and power spectrum features to logarithmic space and standardize them;

[0011] (4) Construct a grouped temporal convolutional neural network formed by stacking several temporal residual modules. Input the frequency domain features processed in step (3) of the training set into the grouped temporal convolutional neural network to obtain the output of the temporal residual modules from low layer to high layer.

[0012] (5) For the output of each layer of time residual module, each channel of the output is used as a graph node, and the data of each channel is used as the feature of the corresponding node. The connection relationship between the graph nodes is obtained by using the graph structure learning method.

[0013] (6) Using the power spectrum features of each channel and frequency processed in step (3) as the attributes of the graph nodes and the connection relationship between the graph nodes output in step (5) as the graph adjacency matrix, the graph representation is calculated using a graph convolutional neural network.

[0014] (7) Input the graph representation output in step (6) into the classifier to obtain the classification score, calculate the loss function, train and update the model parameters in steps (4) to (7) according to the loss function using gradient descent; after training, use the test set to test and evaluate the model.

[0015] (8) In the application process, the EEG signal data to be extracted is preprocessed in steps (1) to (3) and then passed through the model in steps (4) to (6) to obtain the graphical representation, which is the EEG feature to be extracted.

[0016] Further, in step (2), the frequency domain features in the form of a three-dimensional matrix are extracted, specifically by applying the short-time Fourier transform method to the signal of each channel to obtain the frequency domain features in the form of a two-dimensional matrix corresponding to that channel; and stacking the frequency domain features in the form of a two-dimensional matrix corresponding to all channels to form the frequency domain features in the form of a three-dimensional matrix.

[0017] The power spectrum features are extracted in one-dimensional vector form. Specifically, the Welch method is applied to the signal of each channel to obtain the power spectrum features of that channel; the power spectrum features corresponding to all channels are concatenated to form a one-dimensional vector power spectrum feature.

[0018] In step (3), the training set and test set are standardized using the sample mean and sample variance of the training set. Specifically, the frequency domain features of the training set and test set are standardized using the mean and variance of the frequency domain features in the training set; the power spectrum features of the training set and test set are standardized using the mean and variance of the power spectrum features in the training set.

[0019] In step (4), the temporal residual module is composed of multiple temporal convolutional modules and there are jump connections between the input and the output; each temporal convolutional module includes a temporal convolutional layer, a regularization layer, an activation function layer and a random deactivation layer.

[0020] In step (4), in the grouped temporal convolutional neural network, each channel of the input data uses an independent convolutional kernel, and different channels do not share parameters.

[0021] In step (5), the connection relationships between graph nodes are obtained using a graph structure learning method, specifically as follows:

[0022] (5-1) Similarity Calculation: The L2 similarity between nodes is calculated using the following formula:

[0023]

[0024] in, This represents the L2 similarity between nodes i and j. and Let be the t-th feature of nodes i and j, respectively, and T be the feature dimension of nodes i and j;

[0025] (5-2) Nonlinear compression: For each node i, its connection weight with other nodes is calculated using the following formula:

[0026]

[0027] Among them, A i The vector S represents the connection weights from node i to all nodes. i Let L be the norm similarity between node i and all nodes. This represents the average L2 similarity between node i and all nodes;

[0028] (5-3) Graph structure pruning: Prune the adjacency matrix composed of all node connection weights obtained in step (5-2), retain the connections whose connection weights rank in the top 10% of all connection weights in the entire graph, and delete the rest of the connections.

[0029] The specific process of step (6) is as follows:

[0030] (6-1) Using the power spectrum characteristics of each channel and frequency processed in step (3) as the graph node attributes and the graph node relationship output in step (5) as the graph adjacency matrix, add self-connection of nodes in the graph and transform it into an undirected graph to ensure that the adjacency matrix of the graph is a real symmetric matrix.

[0031] (6-2) The Laplacian matrix of the graph is calculated using the symmetric normalized Laplacian method. The specific calculation method is as follows:

[0032]

[0033] Where I is the identity matrix, D is the degree matrix of the input graph, and A s Let L be the adjacency matrix of the input graph, and L be the Laplacian matrix of the graph.

[0034] (6-3) A graph convolutional neural network with one layer is used for graph representation extraction. The convolutional layers of the graph convolutional neural network use the first-order Chebyshev expansion graph convolution method, and the specific calculation method is as follows:

[0035] y = [x, (LI)x]θ

[0036] Where y represents the graph representation, x is the vector obtained by concatenating the attributes of each node in the graph, L is the Laplacian matrix of the graph, I is the identity matrix, and θ is the parameter vector to be learned in the graph convolution method, i.e., the graph convolution kernel.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] To address the time-varying nature of inter-channel connectivity in electroencephalogram (EEG) signals, this invention first employs a grouped temporal convolutional neural network to learn graph node features, handling the time-varying characteristics of EEG signals. Then, a graph structure learning method is used to learn the connectivity relationships between graph nodes, handling the time-varying characteristics of inter-channel relationships in EEG signals. Finally, a graph representation extraction method is used to extract graph representations as the classification basis, completing the classification of EEG signals. This method can simultaneously handle the time-varying nature of single-channel EEG signals and the time-varying nature of inter-channel connectivity, demonstrating excellent performance in classifying interictal and preictal EEG signals in epileptic seizures. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of temporal convolution;

[0040] Figure 2 This is a schematic diagram of the timing residual module;

[0041] Figure 3 This is a schematic diagram of grouped temporal convolution. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0043] This embodiment uses the EEG feature extraction method based on time-series graph structure learning of the present invention to classify EEG signals during the interictal and preictal periods of epilepsy. This embodiment uses the Freiburg intracranial EEG dataset for epilepsy. The Freiburg intracranial EEG dataset contains intracranial EEG (iEEG) data from 21 patients with refractory epilepsy during the interictal and preictal periods, with a signal sampling rate of 256 Hz and 6 channels of data. Three channels are from the epileptic focus region, and the other three channels are from regions far from the epileptic focus. Since deep learning-based epilepsy prediction methods require a large amount of data for training, this invention selects patient data containing five epileptic seizure records from the dataset. The total duration and statistical information of each segment of the preictal data in the used data are shown in Table 1.

[0044] Table 1

[0045]

[0046]

[0047] The EEG feature extraction method based on temporal graph structure learning of this invention includes the following steps:

[0048] (1) Preprocessing of EEG signals: For EEG signal data, signal artifacts are removed, power frequency interference bands are eliminated, and the EEG signal is divided into signal slices of consistent duration using the sliding window method. The data is divided into training set and test set.

[0049] In this example, to eliminate power frequency interference, frequency components near 0Hz, 50Hz, and 100Hz were removed, resulting in each channel of the signal containing a total of 114 frequency components within the ranges of 1Hz-46Hz, 54Hz-96Hz, and 104Hz-128Hz. That is, the total number of frequency components across all channels is 6 × 114.

[0050] =684. The sliding window method has a window length of 10s and a sliding length of 1s. The interepileptic data of each patient are randomly divided into 5 parts, and each part is combined with 1 part of the preepileptic data to form a 1-fold cross-validation dataset; the data of all folds are merged to form a 5-fold cross-validation dataset.

[0051] (2) Extracting primary features: Extract the following two types of primary features for each sample in the training and test sets:

[0052] (2-1) Extracting frequency domain features in three-dimensional matrix form: Apply the short-time Fourier transform method to the signal of each channel to obtain the frequency domain features in two-dimensional matrix form corresponding to that channel; stack the frequency domain features in two-dimensional matrix form corresponding to all channels to form frequency domain features in three-dimensional matrix form.

[0053] In this invention, the overlap rate of the short-time Fourier transform method is set to 0.5.

[0054] (2-2) Extracting power spectral density features in vector form: Apply the Welch method to the signal of each channel to obtain the power spectral density features of that channel; concatenate the power spectral density features corresponding to all channels to form a power spectral density feature in vector form.

[0055] (3) Feature preprocessing: The two primary features extracted in step (2) are transformed to logarithmic space using the natural logarithm function. Then, the training set and test set are standardized using the sample mean and sample variance of the training set. The specific calculation method is as follows:

[0056]

[0057] in, Let represent the t-th dimension feature of the standardized i-th sample. Let represent the t-th dimension feature of the i-th sample in the logarithmic space. σ represents the mean of the t-th dimension of the feature of all samples in the logarithmic space. t Let t represent the standard deviation of the t-th dimension feature of all samples in the logarithmic space.

[0058] (4) Constructing a grouped temporal convolutional neural network: Construct a grouped temporal convolutional neural network formed by stacking several temporal residual modules.

[0059] The following is combined Figure 1 Explain the temporal convolution operation. Figure 1 It contains an input layer and three consecutive temporal convolutional layers. Temporal convolution has two characteristics: first, temporal causality, meaning that the feature learning process at any given time step can only utilize historical information and cannot utilize future information. Figure 1 In the diagram, it can be seen that in each temporal convolutional layer, the neuron corresponding to time 18, which is highlighted, cannot utilize information after time 18; secondly, it maintains the consistency of the input and output sequence lengths. Figure 1 In this model, the input layer and the three temporal convolutional layers all have the same number of neurons. Temporal convolution uses dilated convolution to expand the receptive fields of higher-level neurons. Assuming the kernel size of the temporal convolution is k and the dilation coefficient is d, the temporal convolutional layer will dilate the receptive fields of lower-level neurons every 2... d – One neuron samples once, up to k samples, in Figure 1In the process, the kernel size of the temporal convolutional layer is k=3, the d=0 of the temporal convolutional layer 1, the d=1 of the temporal convolutional layer 2, and the d=2 of the temporal convolutional layer 3.

[0060] The following is combined Figure 2 Explain the temporal residual module. A temporal residual module is composed of multiple temporal convolutional modules connected in series, with jump connections between the input and output. Each temporal convolutional module consists of a temporal convolutional layer, a regularization layer, an activation function layer, and a random deactivation layer connected in series.

[0061] The following is combined Figure 3 Explain how grouped temporal convolution works. Figure 3 The diagram illustrates the operation of the lower-level timing residual module 1 and the higher-level timing residual module 2: rectangular strips of different colors represent different channel signals. In the timing residual module, different channel signals do not share parameters, ensuring that the output channel of the timing residual module is only related to the input channel, and the information interaction between channels is only completed through the subsequent graph construction process.

[0062] In this embodiment, the temporal convolutional neural network includes three temporal residual modules, each containing two temporal convolutional modules. The regularization layer of the temporal convolutional modules uses batch normalization, the activation function layer uses ReLU as the activation function, and the deactivation probability of the random deactivation layer is 0.5. The kernel size of the grouped temporal convolutional network used is 2, and the number of kernels is 684.

[0063] (5) Constructing node relationships between channels: For the output corresponding to each layer of temporal residual module in step (4), each channel of the output is taken as a graph node, and each channel data is taken as the corresponding node feature. The graph structure learning method is used to obtain the connection relationship between each graph node. The graph structure learning method used is further divided into the following steps:

[0064] (5-1) Similarity Calculation: The L2 similarity between each node is calculated using the following formula:

[0065]

[0066] in, This represents the L2 similarity between nodes i and j. and Let be the t-th dimension features of nodes i and j, respectively, and T be the feature dimension of nodes i and j.

[0067] (5-2) Nonlinear compression: For each node i, its connection weight with other nodes is calculated using the following formula:

[0068]

[0069] Among them, A i The vector S represents the connection weights from node i to all nodes. i Let L be the norm similarity between node i and all nodes. This represents the average L2 similarity between node i and all nodes.

[0070] (5-3) Graph structure pruning: Prune the graph structure output in step (5-2), retain the connections whose connection weights rank in the top 10% of all connection weights in the graph, and delete the rest of the connections.

[0071] (6) Graph Representation Learning: Based on the graph structure output in step (5), a channel node relationship graph is generated and input into a graph convolutional neural network to obtain the graph representation. The graph representation method used can be further divided into the following steps:

[0072] (6-1) Preprocessing of the graph: The power spectrum characteristics of each channel and frequency processed in step (3) are used as the graph node attributes, and the connection relationship between the graph nodes output in step (5) is used as the graph adjacency matrix. Node self-connection is added to the graph and it is transformed into an undirected graph to ensure that the adjacency matrix of the graph is a real symmetric matrix.

[0073] (6-2) Graph Laplace: The Laplace matrix of a graph is calculated using the symmetric normalized Laplace method. The specific calculation method is as follows:

[0074]

[0075] Where I is the identity matrix, D is the degree matrix of the input graph, and A s Let L be the adjacency matrix of the input graph, and L be the Laplacian matrix of the graph.

[0076] (6-3) Graph Convolution: Graph representation extraction is performed using a graph convolutional neural network with one layer. The convolutional layers of the graph convolutional neural network use a first-order Chebyshev unfolded graph convolution method, and its specific calculation method is as follows:

[0077] y = [x, (LI)x]θ

[0078] Where x is the vector obtained by concatenating the attributes of the graph nodes, L is the Laplacian matrix of the graph, I is the identity matrix, and θ is the parameter vector to be learned in the graph convolution method, i.e., the graph convolution kernel.

[0079] In this example, the number of convolutional kernels in the graph convolutional neural network is set to 32, and the activation function used is the tanh function.

[0080] (7) Classification and Loss Function Calculation: Input the graph representation output from step (6) into the fully connected neural network classifier, calculate the loss function, and update the model parameters using gradient descent. This can be broken down into the following steps:

[0081] (7-1) Fully connected neural network classifier: The graph representation output from step (6) is input into the neural network classifier. The neural network classifier consists of two fully connected layers and uses skip connection technology and batch normalization technology.

[0082] In this example, the number of neurons in the two fully connected layers of the neural network classifier are 256 and 128, respectively, and the ReLU function is used as the activation function.

[0083] (7-2) Calculation of loss function: Calculate the cross-entropy loss of the classification score output by the classifier in step (7-1).

[0084] (7-3) Update model parameters using gradient descent: Update model parameters using gradient descent with the Adam optimizer.

[0085] In this example, the model weights are initialized using the He initialization method; a decay coefficient of 0.0001 is applied during the learning process; the batch size is set to 64; and the learning rate of the Adam optimizer is 0.00001. The model is trained for 10 epochs. At the end of each training epoch, the model's performance on the training set is evaluated, and the model with the highest F1 score is selected as the final model. The F1 score is defined as follows:

[0086] F1 score=TP / (TP+0.5*(FP+FN))

[0087] Where TP represents the number of positive samples correctly classified as positive samples, FP represents the number of negative samples misclassified as positive samples, and FN represents the number of positive samples misclassified as negative samples.

[0088] (8) Model testing and evaluation: The model is tested using the test set data, and the F1 score is used to evaluate the model performance.

[0089] In this embodiment, the performance of the EEG feature extraction method based on instance-level graph learning network (IGLN) of this invention is compared with that of other deep learning methods on the Freiburg epilepsy intracranial EEG dataset. The results are shown in Table 2. It can be seen that the instance-level graph learning method of this invention achieves better performance than other deep learning methods in the epilepsy prediction problem.

[0090] Table 2

[0091] method accuracy Recall rate Accuracy F1 value CNN-1D-STFT 0.941±0.001 0.758±0.0022 0.912±0.010 0.787±0.002 CNN-2D-STFT 0.947±0.002 0.776±0.006 0.935±0.005 0.809±0.006 ChebNet 0.950±0.004 0.776±0.008 0.933±0.012 0.816±0.007 GCN 0.949±0.001 0.781±0.005 0.933±0.004 0.817±0.004 ResNet 0.940±0.001 0.760±0.007 0.888±0.007 0.789±0.005 Transformer 0.932±0.002 0.838±0.001 0.843±0.008 0.811±0.001 IGLN 0.959±0.004 0.857±0.008 0.917±0.013 0.870±0.007

[0092] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting EEG features based on temporal graph structure learning, characterized in that, Includes the following steps: (1) Extract EEG signal data for preprocessing. Use the sliding window method to divide the EEG signal into signal slices of consistent duration, and divide the signal slice data into training set and test set; (2) For each signal slice data in the training set and the test set, extract the frequency domain features in the form of a three-dimensional matrix and the power spectrum features in the form of a one-dimensional vector. (3) Transform all extracted frequency domain features and power spectrum features to logarithmic space and standardize them; (4) Construct a grouped temporal convolutional neural network formed by stacking several temporal residual modules. Input the frequency domain features processed in step (3) of the training set into the grouped temporal convolutional neural network to obtain the output of the temporal residual modules from low layer to high layer. (5) For the output of each layer of time residual module, each channel of the output is used as a graph node, and the data of each channel is used as the feature of the corresponding node. The connection relationship between the graph nodes is obtained by using the graph structure learning method. (6) Using the power spectrum features of each channel and frequency processed in step (3) as the attributes of the graph nodes and the connection relationship between the graph nodes output in step (5) as the graph adjacency matrix, the graph representation is calculated using a graph convolutional neural network. (7) Input the graph representation output in step (6) into the classifier to obtain the classification score, calculate the loss function, and use gradient descent to train and update the model parameters in steps (4) to (7) based on the loss function. After training, the model is tested and evaluated using a test set. (8) In the application process, the EEG signal data to be extracted is preprocessed in steps (1) to (3) and then passed through the model in steps (4) to (6) to obtain the graphical representation, which is the EEG feature to be extracted.

2. The EEG feature extraction method based on temporal graph structure learning according to claim 1, characterized in that, In step (2), the frequency domain features in the form of a three-dimensional matrix are extracted, specifically as follows: Short-time Fourier transform is applied to the signal of each channel to obtain the frequency domain features in the form of a two-dimensional matrix corresponding to that channel; the frequency domain features in the form of two-dimensional matrices corresponding to all channels are stacked to form the frequency domain features in the form of a three-dimensional matrix.

3. The EEG feature extraction method based on temporal graph structure learning according to claim 1, characterized in that, In step (2), the power spectrum features in one-dimensional vector form are extracted, specifically as follows: The Welch method is applied to the signal of each channel to obtain the power spectrum characteristics of that channel; the power spectrum characteristics of all channels are concatenated to form a one-dimensional vector power spectrum characteristic.

4. The EEG feature extraction method based on temporal graph structure learning according to claim 1, characterized in that, In step (3), the standardization specifically refers to: The frequency domain features of the log space in the training and test sets are standardized using the mean and variance of the frequency domain features in the log space of the training set; the power spectral features of the log space in the training and test sets are standardized using the mean and variance of the power spectral features in the log space of the training set.

5. The EEG feature extraction method based on temporal graph structure learning according to claim 1, characterized in that, In step (4), the temporal residual module is composed of multiple temporal convolutional modules and there are jump connections between the input and the output; each temporal convolutional module includes a temporal convolutional layer, a regularization layer, an activation function layer and a random deactivation layer.

6. The EEG feature extraction method based on temporal graph structure learning according to claim 1, characterized in that, In step (4), in the grouped temporal convolutional neural network, each channel of the input data uses an independent convolutional kernel, and different channels do not share parameters.

7. The EEG feature extraction method based on temporal graph structure learning according to claim 1, characterized in that, In step (5), the connection relationships between graph nodes are obtained using a graph structure learning method, specifically as follows: (5-1) Similarity Calculation: The L2 similarity between nodes is calculated using the following formula: in, This represents the L2 similarity between nodes i and j. and Let be the t-th feature of nodes i and j, respectively, and T be the feature dimension of nodes i and j; (5-2) Nonlinear compression: For each node i, its connection weight with other nodes is calculated using the following formula: Among them, A i The vector S represents the connection weights from node i to all nodes. i Let L be the norm similarity between node i and all nodes. This represents the average L2 similarity between node i and all nodes; (5-3) Graph structure pruning: Prune the adjacency matrix composed of all node connection weights obtained in step (5-2), retain the connections whose connection weights rank in the top 10% of all connection weights in the entire graph, and delete the rest of the connections.

8. The EEG feature extraction method based on temporal graph structure learning according to claim 1, characterized in that, The specific process of step (6) is as follows: (6-1) Using the power spectrum characteristics of each channel and frequency processed in step (3) as the graph node attributes and the graph node relationship output in step (5) as the graph adjacency matrix, add self-connection of nodes in the graph and transform it into an undirected graph to ensure that the adjacency matrix of the graph is a real symmetric matrix. (6-2) The Laplacian matrix of the graph is calculated using the symmetric normalized Laplacian method. The specific calculation method is as follows: Where I is the identity matrix, D is the degree matrix of the input graph, and A s Let L be the adjacency matrix of the input graph, and L be the Laplacian matrix of the graph. (6-3) A graph convolutional neural network with one layer is used for graph representation extraction. The convolutional layers of the graph convolutional neural network use the first-order Chebyshev expansion graph convolution method, and the specific calculation method is as follows: y = [x, (LI)x]θ Where y represents the graph representation, x is the vector obtained by concatenating the attributes of each node in the graph, L is the Laplacian matrix of the graph, I is the identity matrix, and θ is the parameter vector to be learned in the graph convolution method, i.e., the graph convolution kernel.

Images