A method and system for identifying focal epileptic seizures based on scalp EEG signals

Abstract

The present invention provides a method and system for identifying focal epileptic seizures based on scalp EEG signals, which belongs to the field of epilepsy identification in EEG signal processing. It includes two stages: training and testing. The training stage is to train a classification model that can automatically identify focal epileptic EEG signals. The main task of the testing stage is to use the model trained in the training stage to identify and classify the newly arrived scalp EEG signals and generate recognition results. The present invention utilizes the easy collection and safe and non-invasive characteristics of scalp EEG signals, combines the graph convolution deep learning method, and innovatively establishes a deep learning neural network structure to address practical problems in clinical practice, thereby realizing an automatic recognition method suitable for distinguishing focal epileptic seizures from non-focal epileptic seizures. It can effectively provide doctors with auxiliary diagnostic opinions without using intracranial EEG.

Description

Technical Field

[0001] The present invention relates to the field of epilepsy identification in electroencephalogram (EEG) signal processing, and in particular to a method and system for identifying focal epileptic seizures based on scalp EEG signals. Background Art

[0002] Epilepsy is one of the most common neurological diseases, causing recurring and irregular abnormalities in brain neurons. Sudden seizures often lead to convulsions and loss of consciousness, causing significant inconvenience to patients. Repeated seizures can further impair normal brain function, increase the risk of developing other chronic diseases, and even lead to loss of independence. Therefore, early treatment for epilepsy is crucial.

[0003] Today, epilepsy is primarily treated with surgery and medication. Generally speaking, medication can only alleviate epileptic seizures, while surgery is the only treatment that can cure them. However, not all epilepsy patients are candidates for surgical treatment. Generally, only those with focal seizures have the opportunity to undergo surgery. Patients with non-focal seizures can only be treated with medication due to the widespread location of abnormal neuronal discharges. By quickly identifying patients with focal seizures, doctors can help them prepare for surgery as early as possible, significantly reducing the pain epilepsy causes.

[0004] EEG brainwave signals are characteristic of neuronal activity. Using electrodes placed on the scalp, EEG records neural signals in the brain and is a widely used diagnostic method. By monitoring a patient's EEG signals and observing changes in EEG signals across different channels during an epileptic seizure, EEG has become a crucial basis for doctors to determine whether a patient's epilepsy is focal, and a crucial reference for designing subsequent diagnostic and treatment plans.

[0005] While EEG-based epilepsy detection and prediction have been extensively studied, epilepsy classification currently focuses on categorizing epilepsy based on different clinical symptoms. Few researchers have used neural networks to distinguish focal epileptic seizures. Distinguishing focal epileptic seizures hinges on determining whether the features in the patient's pre-ictal EEG data are spatially concentrated or diffusely distributed. Currently, deep learning methods for learning spatial information primarily rely on graph convolution. Constructing graph adjacency relationships between channels and building appropriate graph convolutional neural networks present significant challenges. Summary of the Invention

[0006] The main purpose of this invention is to provide a method and system for identifying focal epileptic seizures based on hypergraph convolution. By constructing graph adjacency relationships between electrode channels for focal and non-focal EEG data, a neural network model composed of graph convolution and time-domain convolution is designed to help doctors diagnose whether an epileptic seizure is focal. In actual clinical diagnosis, focal seizures can be quickly identified based on the patient's EEG signals. The identification results of this method can assist doctors in diagnosis, thereby providing better treatment for patients.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] The first object of the present invention is to provide a method for identifying focal epileptic seizures based on scalp EEG signals, comprising the following steps:

[0009] Step 1: Collect scalp EEG data from patients with focal seizures and patients with non-focal seizures, and segment the scalp EEG data;

[0010] Step 2: Preprocess the scalp EEG data segments obtained in step 1;

[0011] Step 3: Construct a hypergraph for the focal epileptic seizure recognition task based on the preprocessed scalp EEG data segments;

[0012] Step 4: Extract features from the preprocessed scalp EEG data segments;

[0013] Step 5: The features of the patient's scalp EEG data segments containing both focal and non-focal seizures extracted in step 4 and the hypergraph constructed in step 3 are input into the deep learning recognition model for training to obtain a trained focal epileptic seizure classification model;

[0014] Step 6: Obtain the scalp EEG data segment to be detected and extract features, and identify the probability of focal epileptic seizures through the focal epileptic seizure classification model obtained in step 5.

[0015] Furthermore, the specific steps of step 1 are as follows:

[0016] Step 1-1: Collect scalp EEG data from patients with focal seizures and those with non-focal seizures according to the 10-20 international electrode standard;

[0017] Step 1-2: Use the EdfReader tool to read the collected raw scalp EEG signals and intercept the scalp EEG data during the preictal period, which is defined as 30 minutes before the onset of the attack.

[0018] Step 1-3: Divide all collected preictal scalp EEG data into 12-second scalp EEG data segments.

[0019] Furthermore, the specific steps of step 2 are as follows:

[0020] Step 2-1: The obtained scalp EEG data segments are sent to a 1-100 Hz bandpass filter to filter out high-frequency noise;

[0021] Step 2-2: Feed the scalp EEG data segment output by the 1-100 Hz bandpass filter into a 50 Hz notch filter to filter out the AC noise from the power grid;

[0022] Step 2-3: Normalize the scalp EEG data segments output by the 50 Hz notch filter.

[0023] Furthermore, the hypergraph described in step 3 is a combination of a basic hypergraph and a class hypergraph.

[0024] Furthermore, the hypergraph construction method includes:

[0025] Step 3-1: In the construction of the basic hypergraph, for all pre-processed scalp EEG data segments of patients with both focal and non-focal seizures, the mutual information is used to calculate the pairwise correlation between all electrode channels. The five electrode pairs with the largest correlation are selected to form the hyperedges of the basic hypergraph.

[0026] Step 3-2: In the construction of the category hypergraph, for all preprocessed patient scalp EEG data segments containing only focal seizures, the mutual information is used to calculate the pairwise correlation between all electrode channels. The calculated correlation matrix is used as the input of the DBSCAN clustering algorithm to cluster all electrodes. The electrodes in each cluster are a set of hyperedges in the category hypergraph;

[0027] Similarly, for all preprocessed scalp EEG data segments containing only non-focal seizures, we used mutual information to calculate the pairwise correlations between all electrode channels. The calculated correlation matrix was used as the input of the DBSCAN clustering algorithm to cluster all electrodes. The electrodes in each cluster were a set of hyperedges in the category hypergraph.

[0028] Step 3-3: Combine the basic hypergraph and the category hypergraph to form a hypergraph used for the focal epileptic seizure recognition task.

[0029] Furthermore, the feature extraction process in step 4 is as follows: the preprocessed scalp EEG data segment is divided into several 1s non-overlapping data windows, the data in each data window is subjected to Fourier transform to extract spectral features, and the time-frequency spectrum features of the scalp EEG data segment are obtained after recombination.

[0030] Furthermore, the deep learning recognition model described in step 5 is a hypergraph convolutional neural network, including an input layer, an output layer, a first convolutional layer located between the input layer and the output layer, a plurality of hypergraph convolutional blocks consisting of one-dimensional convolutional layers and hypergraph convolutional layers, a readout layer connected to each hypergraph convolutional block, and a second convolutional layer connected to all readout layers;

[0031] The first convolution layer and several hypergraph convolution blocks are connected in sequence, and the output of the first convolution layer is connected to the output of each hypergraph convolution block through a skip layer, and the output of the previous hypergraph convolution block is connected to the output of each subsequent hypergraph convolution block through a skip layer; the outputs of all hypergraph convolution blocks are read by the Readout layer, adjusted to the same dimension and spliced, and then passed through the second convolution layer and the output layer to obtain the final output.

[0032] Furthermore, the features of the patient's scalp EEG data segment are used as input data of the input layer of the hypergraph convolutional neural network. At the same time, the hypergraph constructed in step 3 is used as an additional input of the hypergraph convolutional layer in each hypergraph convolutional block.

[0033] Furthermore, the specific steps of step 6 are as follows:

[0034] Step 6-1: Divide the scalp EEG data to be tested into scalp EEG data segments and perform preprocessing to extract data segment features;

[0035] Step 6-2: Combine the data segment features obtained in step 6-1 and the hyperparameters obtained in step 3 Figure 1 The samples are sent to the focal epileptic seizure classification model trained in step 5 for classification and recognition.

[0036] The second object of the present invention is to provide a focal epileptic seizure identification system based on scalp EEG signals, comprising:

[0037] An EEG data acquisition module, which is used to acquire scalp EEG data of patients with focal seizures and / or non-focal seizures, and to segment the scalp EEG data;

[0038] An EEG data preprocessing module, which is used to preprocess the scalp EEG data segments obtained by the EEG data acquisition module;

[0039] A hypergraph construction module, which is used to construct a hypergraph for focal epileptic seizure recognition task based on preprocessed scalp EEG data segments;

[0040] A feature extraction module is used to extract features from the preprocessed scalp EEG data segments;

[0041] A deep learning recognition model training module is used to input the features of the patient scalp EEG data segments containing both focal seizures and non-focal seizures extracted by the feature extraction module and the hypergraph obtained by the hypergraph construction module into the deep learning recognition model for training, thereby obtaining a trained focal epileptic seizure classification model;

[0042] The focal epileptic seizure identification module is used to identify the probability that the scalp EEG data segment to be detected is a focal epileptic seizure based on the trained focal epileptic seizure classification model.

[0043] Based on this concept, the technical solution of the present invention uses collected data on focal and nonfocal epileptic seizures. Through correlation and clustering calculations, a graph adjacency relationship between channels in the 10-20 international electrode standard is constructed, extracting spatial information features important for focal seizure identification. This spatial information feature is then combined with a hypergraph convolutional neural network to improve model recognition performance. Furthermore, by adding skip-layer connections to the network, the spatial information in the neural network's information flow is effectively transferred backward, avoiding the feature smoothing problem that can occur in graph convolution and increasing the probability of correct recognition.

[0044] The beneficial effects of the present invention are mainly reflected in:

[0045] The present invention obtains feature input with richer information by extracting time-frequency domain features and spatial features, and simultaneously establishes a hypergraph convolutional neural network with added skip layers to model the extracted spatial features, thereby improving the recognition ability of the model.

[0046] This invention utilizes the easy acquisition, safety and non-invasiveness of scalp EEG, combined with deep learning methods, to build and debug a neural network for actual clinical problems to be suitable for the automatic identification of focal epileptic seizures, which can quickly provide auxiliary diagnostic advice to clinicians. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0048] Figure 1 This is the overall architecture diagram of the present invention;

[0049] Figure 2 Detailed diagram of the process proposed by the present invention;

[0050] Figure 3 This is a diagram of the model architecture proposed in the present invention. DETAILED DESCRIPTION

[0051] The present invention is further described below with reference to specific embodiments, but the protection scope of the present invention is not limited thereto:

[0052] A method for identifying focal epileptic seizures based on scalp EEG signals, such as Figure 1 As shown in the figure, it mainly includes three processes: EEG signal acquisition, preprocessing and feature extraction, and focal seizure identification. Specifically, it includes the following steps:

[0053] Step 1: First, prepare the scalp EEG data of a real patient and mark the data segments before the onset of the attack;

[0054] Step 2: Preprocess the patient's scalp EEG data fragments, including filtering noise above 100 Hz and below 1 Hz, filtering 50 Hz AC noise, and normalizing it;

[0055] Step 3: constructing a hypergraph for the focal epileptic seizure identification task based on EEG data segments of patients with focal seizures and patients with non-focal seizures, wherein the hypergraph represents the adjacency relationship of electrode channels;

[0056] Step 4: Extract features from the scalp EEG data segments preprocessed in step 2;

[0057] Step 5: Input the features of the EEG data segments of focal seizures and non-focal seizures obtained in step 4, and the hypergraph obtained in step 3 into the deep learning recognition model for network training to obtain a trained focal epileptic seizure classification model;

[0058] Step 6: In the clinical focal seizure identification scenario, obtain the scalp EEG data segment to be detected, identify it using the focal epileptic seizure classification model obtained in step 5, and output the predicted probability that the scalp EEG data segment belongs to the focal seizure type as the identification result. The identification result is used to assist diagnosis.

[0059] In a specific implementation of the present invention, step 1 is a training set acquisition process for a deep learning recognition model, which is implemented by the following sub-steps:

[0060] Step 1-1: First, collect data from real patients. The data in this example uses scalp EEG data from patients with focal seizures and non-focal seizures collected by the Second Affiliated Hospital of Zhejiang University. The scalp EEG signals are collected according to the 10-20 international electrode standard. The statistical information of the patients is shown in Table 1.

[0061] Table 1: Statistics of epilepsy patients in the Second Affiliated Hospital of Zhejiang University

[0062] Focal seizures (n=53) Non-focal seizures (n=14) Total (n=67) Gender, male, proportion (%) 31(58.5％) 6(42.9％) 37(55.2％) Age at the time of EEG recording, years 24.2±10.3 27.6±13.1 24.9±10.9 Age at onset, years 13.7±10.5 12.5±7.4 13.5±9.9 Duration of epilepsy, years 10.4±7.1 15.0±10.3 11.4±8.0 Age at surgery, years 24.4±10.3 25.6±8.4 24.3±10.4

[0063] Step 1-2: Use the EdfReader tool to read the collected raw scalp EEG signals and intercept the preictal data. In this embodiment, the preictal period is defined as the scalp EEG data 30 minutes before the epileptic seizure;

[0064] Step 1-3: Divide all collected preictal data into 12s data segments;

[0065] Step 1-4: Screen data segments of appropriate proportions as training sets and test sets. In this embodiment, data segments of 80% of patients with focal seizures and 80% of patients with non-focal seizures are used as training sets, and the rest are used as test sets.

[0066] In a specific implementation of the present invention, the purpose of step 2 is to pre-process the pre-epilepsy seizure data segment, and the specific steps are as follows:

[0067] Step 2-1: The obtained data segments are sent to a 1-100 Hz bandpass filter to filter out high-frequency noise;

[0068] Step 2-2: Send the signal output by the bandpass filter to a 50Hz notch filter to filter out the AC noise of the power grid;

[0069] Step 2-3: Normalize all EEG data segments and record the set of original EEG data segments as x = [x1, x2, x3…, x n ], n represents the number of EEG data segments, x n Represents the nth EEG data segment, and normalizes each dimension of the original data. The operation is:

[0070]

[0071] Among them, x out，i is the output of the normalization operation of the ith EEG data segment, mean(x i ) represents the mean of the ith EEG data segment, std(x i ) represents the standard deviation of the i-th EEG data segment.

[0072] In a specific implementation of the present invention, the hypergraph in step 3 includes: a basic hypergraph and a class hypergraph, and the hypergraph construction method includes the following steps:

[0073] Step 3-1: In the basic hypergraph construction, for all training set data, use the mutual information to calculate the correlation between all electrode channels A = (a 1，1 ,...a p，q ,...,a N，N )∈R N×N, where N is the number of electrodes for signal acquisition. In this embodiment, N is 19; a p，q Represents the correlation between the signals collected from the p-th electrode and the q-th electrode in all training set data;

[0074] From {a u，v(u＞v)}Select the five electrode pairs with the largest correlation, and use the adjacency relationship of each electrode pair as a hyperedge of the basic hypergraph, indicating that there is a connection relationship between the electrode pairs;

[0075] Step 3-2: In the construction of the class hypergraph, for the data of focal seizures, the mutual information is used to calculate the correlation between all electrode channels, which is recorded as A f =(a f，1，1 ,...a f，p，q ,...,a f，N，N )∈R N×N , where a f，p，q represents the correlation between the signals collected from the pth electrode and the qth electrode in the focal seizure data;

[0076] The pairwise correlation matrix A f As the input of the DBSCAN clustering algorithm, the electrode channels are clustered to obtain three main clusters. The electrodes in each main cluster are a set of hyperedges in the category hypergraph, indicating that there is a connection between the electrodes in the cluster.

[0077] The data of non-focal seizures are processed similarly, and the pairwise correlation between channels is calculated, which is recorded as A g =(a g，1，1 ,...a g，p，q ,...,a g，N，N )∈R N×N , where a g，p，q represents the correlation between the signals collected from the pth electrode and the qth electrode in the non-focal seizure data;

[0078] The pairwise correlation matrix A g As the input of the DBSCAN clustering algorithm, the electrode channels were clustered to obtain three main clusters, and all hyperedges in the non-focal seizure class data were obtained;

[0079] The hypergraph obtained by combining the focal seizure data and the hypergraph obtained by non-focal seizures constitutes a category hypergraph;

[0080] Step 3-3: Combine the basic hypergraph and the category hypergraph to form a hypergraph for focal epileptic seizure identification task. In this embodiment, the dimension of the association matrix of the hypergraph is R 19×11 , 19 is the number of channels, and 11 is the sum of 5 hyperedges in the base hypergraph and 6 hyperedges in the category hypergraph.

[0081] In a specific implementation of the present invention, the feature extraction in step 4 is to extract the time-frequency features by Fourier transform, that is, The specific operation is to divide the 12s data segment into 12 1s non-overlapping data windows, perform Fourier transform on the data in each data window to extract the spectrum features, and then recombine to obtain the time-frequency spectrum features of the 12s data segment. Each sample becomes a dimension of R 19×12×100 The feature matrix of , 19 is the number of channels, 12 is the number of data windows, and 100 is the frequency feature dimension.

[0082] In a specific implementation of the present invention, step 5 is the training process of the deep learning recognition model, and the specific steps are as follows:

[0083] Step 5-1: Construct a hypergraph convolutional neural network. In this embodiment, Figure 3 As shown, the hypergraph convolutional neural network includes an input layer, an output layer, a first convolutional layer located between the input layer and the output layer, a plurality of hypergraph convolutional blocks consisting of one-dimensional convolutional layers and hypergraph convolutional layers, a readout layer connected to each hypergraph convolutional block, and a second convolutional layer connected to all readout layers;

[0084] The first convolution layer and several hypergraph convolution blocks are connected in sequence, and the output of the first convolution layer is connected to the output of each hypergraph convolution block through a skip layer, and the output of the previous hypergraph convolution block is connected to the output of each subsequent hypergraph convolution block through a skip layer; the outputs of all hypergraph convolution blocks are read by the Readout layer, adjusted to the same dimension and spliced, and then passed through the second convolution layer and the output layer to obtain the final output.

[0085] The specific parameters of this neural network architecture are shown in Table 2 below. Skip connections are added to the intermediate layers of the network (implemented via a skip connection aggregator). This means that the output of layer 2 is added to the outputs of layers 4, 6, and 8; the output of layer 4 is added to the outputs of layers 6 and 8; and the output of layer 6 is added to the output of layer 8. For example, when the output of layer 2 is added to the output of layer 4, the output of layer 2 (19×12×32) is transformed into a 19×11×32 feature matrix through an additional 1D convolutional layer in the skip connection aggregator. This feature matrix has the same dimensions as the output of layer 4, and the final output of layer 4 is obtained by direct matrix addition. The outputs of layers 4, 6, and 8 are then added together through a readout layer to form the input to layer 9. The readout layer adjusts the feature dimensions of the outputs of layers 4, 6, and 8 to the same level by cropping the output of layer 4 (19×11×32) along the second channel, extracting the final 19×5×32 matrix to achieve the same feature dimensions as the output of layer 8. The same operation is performed on the output of layer 6. The feature matrix obtained in step 4 is fed into the hypergraph convolutional neural network as input, and the hypergraph obtained in step 3 is used as the additional parameters required by the hypergraph convolutional layer.

[0086] Table 2: Neural network structure details

[0087]

[0088]

[0089] refer to Figure 2 Each convolutional layer is followed by a BatchNorm layer, and then the nonlinear activation function is applied. This ensures that the distribution of features in each layer of the neural network is as close to the normal distribution as possible, which can accelerate network training and improve classification accuracy.

[0090] In this embodiment, ReLU is selected as the activation function, Cross Entropy Loss is used as the loss function, the Adam gradient descent method is used for optimization, the learning rate is set to 1e-5, back propagation is used for network training, and the trained network is saved for use in testing.

[0091] In a specific implementation of the present invention, the specific steps of step 6 are as follows:

[0092] Step 6-1: During the test, the collected scalp EEG data to be tested is sliced, noise is filtered, normalized, and features are extracted, similar to steps 1, 2, and 4 in the training phase, to obtain the feature matrix to be tested;

[0093] Step 6-2: Combine the feature matrix and the hyperparameter obtained in step 3 Figure 1The results are sent to the model trained in step 5 for classification and identification to obtain the probability that the test patient has a focal epileptic seizure, and the results are generated into a report to assist doctors in diagnosis.

[0094] This embodiment also provides a focal epileptic seizure identification system based on scalp EEG signals, which is used to implement the above embodiment. The terms "module", "unit", etc. used below can refer to a combination of software and / or hardware that can implement a predetermined function.

[0095] This embodiment provides a focal epileptic seizure identification system based on scalp EEG signals, comprising:

[0096] An EEG data acquisition module, which is used to acquire scalp EEG data of patients with focal seizures and / or non-focal seizures, and to segment the scalp EEG data;

[0097] An EEG data preprocessing module, which is used to preprocess the scalp EEG data segments obtained by the EEG data acquisition module;

[0098] A hypergraph construction module, which is used to construct a hypergraph for focal epileptic seizure recognition task based on preprocessed scalp EEG data segments;

[0099] A feature extraction module is used to extract features from the preprocessed scalp EEG data segments;

[0100] A deep learning recognition model training module is used to input the features of the patient scalp EEG data segments containing both focal seizures and non-focal seizures extracted by the feature extraction module and the hypergraph obtained by the hypergraph construction module into the deep learning recognition model for training, thereby obtaining a trained focal epileptic seizure classification model;

[0101] The focal epileptic seizure identification module is used to identify the probability that the scalp EEG data segment to be detected is a focal epileptic seizure based on the trained focal epileptic seizure classification model.

[0102] As for the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment, and the implementation methods of the remaining modules will not be repeated here. The system embodiment described above is only illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Ordinary technicians in this field can understand and implement it without paying any creative work.

[0103] Embodiments of the system of the present invention can be applied to any device with data processing capabilities, such as a computer or other device. System embodiments can be implemented through software, hardware, or a combination of software and hardware. For example, a software implementation, as a logical device, is implemented by a processor of any device with data processing capabilities, reading corresponding computer program instructions from non-volatile memory into internal memory and executing them.

[0104] Finally, it should be noted that the above examples are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples and is subject to numerous variations. All variations that can be directly derived or conceived by a person of ordinary skill in the art from the disclosure of the present invention are considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying focal epileptic seizures based on scalp EEG signals, characterized by: The following steps are involved: Step 1: Collect scalp EEG data from patients with focal seizures and patients with non-focal seizures, and segment the scalp EEG data; Step 2: Preprocess the scalp EEG data segments obtained in step 1; Step 3: Construct a hypergraph for the focal epileptic seizure recognition task based on the preprocessed scalp EEG data segments; The hypergraph described in step 3 is a combination of a basic hypergraph and a class hypergraph; The hypergraph construction method includes: Step 3-1: In the construction of the basic hypergraph, for all pre-processed scalp EEG data segments of patients with both focal and non-focal seizures, the mutual information is used to calculate the pairwise correlation between all electrode channels. The five electrode pairs with the largest correlation are selected to form the hyperedges of the basic hypergraph. Step 3-2: In the construction of the category hypergraph, for all preprocessed patient scalp EEG data segments containing only focal seizures, the mutual information is used to calculate the pairwise correlation between all electrode channels. The calculated correlation matrix is used as the input of the DBSCAN clustering algorithm to cluster all electrodes. The electrodes in each cluster are a set of hyperedges in the category hypergraph; Similarly, for all preprocessed scalp EEG data segments containing only non-focal seizures, we used mutual information to calculate the pairwise correlations between all electrode channels. The calculated correlation matrix was used as the input of the DBSCAN clustering algorithm to cluster all electrodes. The electrodes in each cluster were a set of hyperedges in the category hypergraph. Step 3-3: Combine the basic hypergraph and the category hypergraph to form a hypergraph for the focal epileptic seizure recognition task; Step 4: Extract features from the preprocessed scalp EEG data segments; Step 5: The features of the patient's scalp EEG data segments containing both focal and non-focal seizures extracted in step 4 and the hypergraph constructed in step 3 are input into the deep learning recognition model for training to obtain a trained focal epileptic seizure classification model; Step 6: Obtain the scalp EEG data segment to be detected and extract features, and identify the probability of focal epileptic seizures through the focal epileptic seizure classification model obtained in step 5.

2. The method for identifying focal epileptic seizures based on scalp EEG signals according to claim 1, characterized in that: The specific steps of step 1 are as follows: Step 1-1: Collect scalp EEG data from patients with focal seizures and those with non-focal seizures according to the 10-20 international electrode standard; Step 1-2: Use the EdfReader tool to read the collected raw scalp EEG signals and intercept the scalp EEG data during the preictal period, which is defined as 30 minutes before the onset of the attack. Step 1-3: Divide all collected preictal scalp EEG data into 12-second scalp EEG data segments.

3. The method for identifying focal epileptic seizures based on scalp EEG signals according to claim 1, characterized in that: The specific steps of step 2 are as follows: Step 2-1: The obtained scalp EEG data segments are sent to a 1-100 Hz bandpass filter to filter out high-frequency noise; Step 2-2: Feed the scalp EEG data segment output by the 1-100 Hz bandpass filter into a 50 Hz notch filter to filter out the AC noise from the power grid; Step 2-3: Normalize the scalp EEG data segments output by the 50 Hz notch filter.

4. The method for identifying focal epileptic seizures based on scalp EEG signals according to claim 1, characterized in that: The feature extraction process in step 4 is as follows: the preprocessed scalp EEG data segment is divided into several 1s non-overlapping data windows, the data in each data window is Fourier transformed to extract the spectrum features, and the time-frequency spectrum features of the scalp EEG data segment are obtained after recombination.

5. The method for identifying focal epileptic seizures based on scalp EEG signals according to claim 1, characterized in that: The deep learning recognition model described in step 5 is a hypergraph convolutional neural network, which includes an input layer, an output layer, a first convolutional layer located between the input layer and the output layer, several hypergraph convolutional blocks consisting of one-dimensional convolutional layers and hypergraph convolutional layers, a readout layer connected to each hypergraph convolutional block, and a second convolutional layer connected to all readout layers; The first convolution layer and several hypergraph convolution blocks are connected in sequence, and the output of the first convolution layer is connected to the output of each hypergraph convolution block through a skip layer, and the output of the previous hypergraph convolution block is connected to the output of each subsequent hypergraph convolution block through a skip layer; the outputs of all hypergraph convolution blocks are read by the Readout layer, adjusted to the same dimension and spliced, and then passed through the second convolution layer and the output layer to obtain the final output.

6. The method for identifying focal epileptic seizures based on scalp EEG signals according to claim 5, characterized in that: The features of the patient's scalp EEG data segment are used as input data of the hypergraph convolutional neural network input layer. At the same time, the hypergraph constructed in step 3 is used as an additional input of the hypergraph convolutional layer in each hypergraph convolutional block.

7. The method for identifying focal epileptic seizures based on scalp EEG signals according to claim 1, wherein the specific steps of step 6 are as follows: Step 6-1: Divide the scalp EEG data to be tested into scalp EEG data segments and perform preprocessing to extract data segment features; Step 6-2: The data segment features obtained in step 6-1 and the hypergraph obtained in step 3 are fed into the focal epileptic seizure classification model trained in step 5 for classification and recognition.

8. A focal epileptic seizure identification system based on scalp EEG signals, characterized in that: include: An EEG data acquisition module, which is used to acquire scalp EEG data of patients with focal seizures and / or non-focal seizures, and to segment the scalp EEG data; An EEG data preprocessing module, which is used to preprocess the scalp EEG data segments obtained by the EEG data acquisition module; A hypergraph construction module, which is used to construct a hypergraph for focal epileptic seizure recognition task based on preprocessed scalp EEG data segments; The hypergraph is a combination of a basic hypergraph and a class hypergraph; In the construction of the basic hypergraph, for all pre-processed scalp EEG data segments of patients with both focal and non-focal seizures, the mutual information was used to calculate the pairwise correlation between all electrode channels. The five electrode pairs with the largest correlation were selected to form the hyperedges of the basic hypergraph. In constructing the category hypergraph, we used mutual information to calculate the pairwise correlations between all electrode channels for all preprocessed scalp EEG data segments containing only focal seizures. The resulting correlation matrix was used as input to the DBSCAN clustering algorithm to cluster all electrodes. The electrodes in each cluster were then represented as a set of hyperedges in the category hypergraph. Similarly, for all preprocessed scalp EEG data segments containing only non-focal seizures, we used mutual information to calculate the pairwise correlations between all electrode channels. The calculated correlation matrix was used as the input of the DBSCAN clustering algorithm to cluster all electrodes. The electrodes in each cluster were a set of hyperedges in the category hypergraph. The basic hypergraph and the category hypergraph are combined to form the hypergraph used in the focal epileptic seizure recognition task; A feature extraction module is used to extract features from the preprocessed scalp EEG data segments; A deep learning recognition model training module is used to input the features of the patient scalp EEG data segments containing both focal seizures and non-focal seizures extracted by the feature extraction module and the hypergraph obtained by the hypergraph construction module into the deep learning recognition model for training, thereby obtaining a trained focal epileptic seizure classification model; The focal epileptic seizure identification module is used to identify the probability that the scalp EEG data segment to be detected is a focal epileptic seizure based on the trained focal epileptic seizure classification model.

Images