Epilepsy prediction method and system based on graph convolutional network, and storage medium

Abstract

The invention discloses an epilepsy prediction method and system based on a graph convolutional network, and a storage medium, relates to the technical field of medical artificial intelligence, and solves the problems that in the prior art, analysis is often carried out based on multi-dimensional features in multiple channels in electroencephalogram signals, the spatial topological relation between the channels is ignored, the position of a focus cannot be accurately recognized in epilepsy prediction, and the epilepsy prediction accuracy is poor. And the accuracy of the epilepsy prediction method is low. A dynamic electroencephalogram topological graph is constructed based on patient data; inputting the dynamic electroencephalogram topological graph into an epilepsy prediction model to obtain epilepsy prediction data; an early warning signal is generated based on epilepsy prediction data, and an electroencephalogram topological graph of a patient is dynamically adjusted within a time window range based on electroencephalogram signal data of the patient, so that during epilepsy analysis, not only multi-dimensional features in channels but also a connection relationship between the channels can be concerned, and an epilepsy occurrence position and a probability condition can be accurately predicted; and the accuracy of the epilepsy prediction method is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical artificial intelligence, and particularly relates to a seizure prediction method and system based on a graph convolution network and a storage medium. BACKGROUND

[0002] Seizure is a common neurological disease characterized by sudden and temporary brain dysfunction caused by repeated abnormal brain electrical activity. Seizure often has suddenness and unpredictability, which brings great inconvenience to the life of patients, and may even cause accidental injury or life danger. The core of seizure prediction is to analyze biological data such as electroencephalogram signals, identify specific electroencephalogram patterns or physiological changes that may occur before a seizure, and thus give an early warning of a possible seizure. Therefore, how to accurately predict seizures has become an important research topic in the field of neuroscience and clinical medicine in recent years.

[0003] The prior art often analyzes the multi-dimensional features within the multi-channel in the electroencephalogram signal when predicting seizures, ignores the spatial topological relationship between the channels, and thus cannot accurately identify the lesion location, resulting in a low accuracy of the seizure prediction method. Therefore, the seizure prediction method still needs to be further improved. SUMMARY

[0004] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application provides a seizure prediction method and system based on a graph convolution network and a storage medium, which are used to solve the technical problem that the prior art often analyzes the multi-dimensional features within the multi-channel in the electroencephalogram signal, ignores the spatial topological relationship between the channels, and thus cannot accurately identify the lesion location, resulting in a low accuracy of the seizure prediction method.

[0005] To achieve the above-mentioned purpose, the first aspect of the present application provides a seizure prediction method based on a graph convolution network, comprising: obtaining patient data; the patient data includes patient ID, body data and electroencephalogram signal data; constructing a dynamic electroencephalogram topology graph based on the patient data; inputting the dynamic electroencephalogram topology graph into a seizure prediction model to obtain seizure prediction data; the seizure prediction model is constructed by a graph convolution network model, and is used to analyze the dynamic electroencephalogram topology graph of the patient to obtain the seizure prediction data of the patient; generating an early warning signal based on the seizure prediction data.

[0006] The application dynamically constructs and adjusts the electroencephalogram topology graph of the patient based on the electroencephalogram signal data of the patient within the time window range, so that when analyzing epilepsy, not only the multidimensional characteristics within a single channel are considered, but also the functional connection relationship between different channels is further considered, thereby more comprehensively reflecting the spatiotemporal variation characteristics of the brain network, and the accuracy and reliability of the epilepsy prediction method are effectively improved through joint prediction of the epilepsy seizure position and occurrence probability.

[0007] Further, the constructing a dynamic electroencephalogram topology graph based on patient data comprises: extracting electroencephalogram signal data in the patient data within a time window range; the electroencephalogram signal data comprises a plurality of channel IDs and corresponding channel parameters; the time window range refers to the range of the electroencephalogram signal data in terms of time length; generating a dynamic adjacency matrix based on the channel IDs; generating a dynamic feature matrix based on the body data and the channel parameters; The dynamic electroencephalogram topology graph comprises a dynamic adjacency matrix and a dynamic feature matrix.

[0008] Further, the generating a dynamic adjacency matrix based on the channel IDs comprises: extracting the phase locking value PLV, the phase lag index PLI, the Euclidean distance OJ between each channel ID, and the channel signal-to-noise ratio SNR corresponding to each channel ID; the phase locking value is a quantitative stability of the phase difference between two channel IDs; the phase lag index is used to suppress zero-lag synchronization and highlight the directionality between channel IDs; the Euclidean distance refers to the physical distance between two channel IDs; and the channel signal-to-noise ratio refers to the high-frequency band energy proportion of the channel ID; calculating the connection weight between each channel ID by a formula The calculation formula of the connection weight satisfies: wherein i and j represent the numbers corresponding to the channel IDs; represents the average neighbor distance corresponding to the i th channel ID, k represents the number of channel IDs connected to the i th channel ID, and K represents the total number of channel IDs connected to the i th channel ID; represents a phase synchronization function, the phase synchronization function is constructed by a nonlinear relationship between the phase locking value PLV, the phase lag index PLI, and the phase synchronization coefficient, and the phase synchronization function satisfies the following calculation formula: wherein i and j represent the numbers corresponding to the channel IDs, represents a frequency band difference penalty coefficient, ∈(0, 1); a channel dominant frequency band difference between the ith channel ID and the jth channel ID; a phase synchronization coefficient between the channel IDs can be obtained by substituting the phase locking value and the phase lag index between the channel IDs into the phase synchronization function; extracting the connection weight between the channel IDs, and creating an adjacency matrix based on the total number of channel IDs; a dynamic adjacency matrix can be obtained by substituting the connection weight between the channel IDs into the corresponding position in the adjacency matrix.

[0009] Further, the dynamic feature matrix is generated based on the body data and the channel parameters, comprising: extracting the influence features in the channel parameters; the influence features refer to the features that affect the prediction of epilepsy; integrating the body data, the channel IDs, and the influence features corresponding to the channel IDs into channel weight analysis data; obtaining the feature weight coefficients corresponding to the channel IDs by substituting the channel weight analysis data into a channel weight generation model; the channel weight generation model is constructed by an artificial intelligence model, and is used to analyze the comprehensive data of the patient to predict the influence degree of the influence features in each channel ID on the prediction of epilepsy; fusing the influence features in the same channel ID by the feature weight coefficients corresponding to the channel ID to obtain the channel fusion features corresponding to the channel ID; creating a feature matrix based on the total number of channel IDs; obtaining a dynamic feature matrix by substituting the channel fusion features corresponding to the channel ID into the corresponding position in the feature matrix.

[0010] Further, the channel weight generation model is constructed by an artificial intelligence model, comprising: obtaining a plurality of historical channel weight analysis data, historical channel IDs corresponding thereto, and historical feature weight coefficients corresponding to the historical channel IDs; dividing the plurality of historical channel weight analysis data, the historical channel IDs corresponding thereto, and the historical feature weight coefficients corresponding to the historical channel IDs into training data, validation data, and test data; and performing data preprocessing on the training data, the validation data, and the test data to obtain a training set, a validation set, and a test set; selecting an artificial intelligence model as a base model; training the base model by the training set, and adjusting the learning rate and other hyperparameters on the validation set to obtain a pre-trained model; verifying the pre-trained model on the test set to finally obtain a channel weight generation model with the channel weight analysis data as the input and the feature weight coefficients corresponding to the channel IDs as the output.

[0011] The application analyzes several influence characteristics of each channel ID in the electroencephalogram signal data in real time, and dynamically generates the corresponding feature weight coefficients of the several influence characteristics in each channel ID through the pre-trained channel weight generation model, so that the several influence characteristics in each channel have different influence degrees in different states, the feature expression ability of the several channel IDs is more abundant, the input of the epilepsy prediction model is provided with strong data support, and the accuracy and efficiency of the epilepsy prediction method are improved.

[0012] Further, the epilepsy prediction model is constructed by a graph convolution network model, comprising: obtaining several historical dynamic electroencephalogram topological graphs and corresponding historical epilepsy data; dividing the several historical dynamic electroencephalogram topological graphs and corresponding historical epilepsy data into training data, verification data and test data; and performing data preprocessing on the training data, verification data and test data to obtain a training set, a verification set and a test set; selecting a graph convolution network model as a basic model; training the basic model through the training set, and adjusting the learning rate and other hyperparameters on the verification set to obtain a pre-trained model; verifying the pre-trained model on the test set to finally obtain an epilepsy prediction model with dynamic electroencephalogram topological graph as input and epilepsy prediction data as output.

[0013] Further, the pre-warning signal is generated based on the epilepsy prediction data, comprising: obtaining body data in the patient data; extracting the epilepsy prediction probability in the epilepsy prediction data; when the epilepsy prediction probability is greater than or equal to D times the personalized probability threshold, generating a severe epilepsy attack warning signal; the personalized probability threshold is generated by the body data; when the epilepsy prediction probability is less than D times the personalized probability threshold and greater than or equal to the personalized probability threshold, generating an epilepsy attack warning signal; wherein D is a proportion coefficient, D>1.

[0014] Further, the personalized probability threshold is generated by the body data, comprising: extracting several historical epilepsy prediction probabilities LDYG of the body data in a historical time window, and a historical prediction number LYS of several historical epilepsy prediction probabilities greater than or equal to E times the baseline probability threshold; the historical time window refers to a period of time in the past; wherein E represents a proportion coefficient, E∈(0, 1); the baseline probability threshold satisfies the following calculation formula: ; wherein PG represents the average probability of epilepsy of a plurality of epilepsy patients having the same epilepsy state as the patient, calculated by the formula ; h represents a sensitivity coefficient, h>0; the specific value is set according to experience, and in this embodiment, h is set to 2; n represents the number of the other epilepsy patients, and N represents the total number of the other epilepsy patients, represents the probability of epilepsy of the nth epilepsy patient; The personalized probability threshold satisfies the following calculation formula: ; wherein β represents a dynamic sensitivity coefficient, β∈(0, 1); represents the risk accumulation intensity, calculated by the formula ; wherein m represents the number corresponding to a plurality of historical epilepsy prediction probabilities in a historical time window, M represents the total number of the plurality of historical epilepsy prediction probabilities in the historical time window; max() represents a maximum value operation; BY represents a volatility suppression factor, calculated by the formula, which satisfies ; wherein m and a both represent the number corresponding to a plurality of historical epilepsy prediction probabilities in a historical time window.

[0015] The application dynamically adjusts the threshold corresponding to the epilepsy prediction probability according to the historical body data of the patient, so that the probability threshold can have a personalized correlation with the patient, provides a personalized probability threshold corresponding to each patient, avoids false positives or false negatives, and improves the accuracy of the warning signal and the efficiency of signal generation.

[0016] The second aspect of the application provides a seizure prediction system based on a graph convolution network, comprising a data acquisition module, a data analysis module and a warning module; the data acquisition module and the data analysis module are connected; the data analysis module and the warning module are connected; The data acquisition module acquires patient data through a data acquisition device; The data analysis module comprises a dynamic topology construction unit and a signal generation unit; The dynamic topology construction unit constructs a dynamic electroencephalogram topology graph based on the patient data; The signal generation unit inputs the dynamic electroencephalogram topology graph into a seizure prediction model to obtain seizure prediction data; the seizure prediction model is constructed by a graph convolution network model, used to analyze the dynamic electroencephalogram topology graph of the patient to obtain the seizure prediction data of the patient; and a warning signal is generated based on the seizure prediction data; The warning module makes a corresponding prompt based on the warning signal.

[0017] Another aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program realizes a graph convolution network based epilepsy prediction system according to the second aspect of the present application when executed by a processor.

[0018] Compared with the prior art, the beneficial effects of the present application are: 1、The present application constructs a dynamic electroencephalogram topology graph based on patient data; inputs the dynamic electroencephalogram topology graph into an epilepsy prediction model to obtain epilepsy prediction data; generates an early warning signal based on the epilepsy prediction data, dynamically adjusts the patient's electroencephalogram topology graph within the time window range based on the patient's electroencephalogram signal data, so that when analyzing epilepsy, not only the multi-dimensional features within the channel are concerned, but also the connection relationship between the channels is concerned, the position and probability of epilepsy occurrence are accurately predicted, and the accuracy of the epilepsy prediction method is improved.

[0019] 2、The present application analyzes each channel, extracts relevant data reflecting the connection degree of the channels, and quantifies the connection degree between each channel, so that when the electroencephalogram signal data changes, not only the change of the features is considered, but also the change of the channel connection is considered, so that more accurate data support is provided when analyzing by the epilepsy prediction model, and the accuracy of epilepsy prediction is improved. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 A graph convolution network based epilepsy prediction method flowchart of the present application; Figure 2 A graph convolution network based epilepsy prediction system principle schematic diagram of the present application. DETAILED DESCRIPTION

[0022] The technical solutions of the present application will be described in detail below in conjunction with the embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0023] Please refer to Figure 1 The first aspect of the present application provides a graph convolution network based epilepsy prediction method, which comprises: obtaining patient data; the patient data comprises patient ID, body data and electroencephalogram signal data; constructing a dynamic electroencephalogram topology graph based on the patient data; inputting the dynamic electroencephalogram topology graph into an epilepsy prediction model to obtain epilepsy prediction data; the epilepsy prediction model is constructed by a graph convolution network model, and is used for analyzing the dynamic electroencephalogram topology graph of the patient to obtain the epilepsy prediction data of the patient; generating an early warning signal based on the epilepsy prediction data.

[0024] In the embodiment, the dynamic electroencephalogram topology graph is constructed based on the patient data, which comprises: extracting electroencephalogram signal data in the patient data within a time window range; the electroencephalogram signal data comprises a plurality of channel IDs and corresponding channel parameters; the time window range refers to the range of the electroencephalogram signal data in the time length, and the specific value is set according to experience; in the embodiment, the time window range is set to 4 seconds; generating a dynamic adjacency matrix based on the channel IDs; generating a dynamic feature matrix based on the body data and the channel parameters; The dynamic electroencephalogram topology graph comprises the dynamic adjacency matrix and the dynamic feature matrix.

[0025] In the embodiment, the dynamic adjacency matrix is generated based on the channel IDs, which comprises: extracting phase locking value (PLV), phase lag index (PLI), Euclidean distance (OJ) between each channel ID, and channel signal-to-noise ratio (SNR) corresponding to each channel ID; the phase locking value is a quantitative stability of the phase difference between two channel IDs; the phase lag index is used to suppress zero-lag synchronization and highlight the directionality between channel IDs; the Euclidean distance refers to the physical distance between two channel IDs; the channel signal-to-noise ratio refers to the proportion of high-frequency band energy of the channel ID; calculating the connection weight between each channel ID by a formula The calculation formula of the connection weight satisfies: wherein i and j represent the number corresponding to the channel ID; represents the average neighbor distance corresponding to the i th channel ID, k represents the number of channel IDs connected with the i th channel ID, and K represents the total number of channel IDs connected with the i th channel ID; represents a phase synchronization function, which is constructed by a nonlinear relationship between the phase locking value (PLV), the phase lag index (PLI) and the phase synchronization coefficient, and satisfies the following calculation formula: wherein i and j represent the number corresponding to the channel ID, represents a frequency band difference penalty coefficient, ∈(0, 1), the specific value is set according to experience, and in the embodiment, the value is set to 0.15. is set to 0.15. is expressed as the channel dominant frequency band difference between the ith channel ID and the jth channel ID. The phase synchronization coefficient between the respective channel IDs can be obtained by substituting the phase locking value and the phase lag index between the respective channel IDs into the phase synchronization function. The connection weight between the respective channel IDs is extracted, and an adjacency matrix is created based on the total number of channel IDs. The connection weight between the respective channel IDs is set in the corresponding position in the adjacency matrix to obtain a dynamic adjacency matrix.

[0026] The embodiment analyzes the correlation between the respective electroencephalogram channels in depth, extracts relevant features that can reflect the connection strength between the channels, and quantitatively represents the connection degree between the channels accordingly. When the electroencephalogram signal data changes, this method not only focuses on the change of individual features, but also synchronously reflects the dynamic evolution of the connection mode between the channels, thereby providing more comprehensive and accurate data support when the epilepsy prediction model is analyzed, and effectively improving the accuracy and stability of epilepsy prediction.

[0027] The dynamic feature matrix based on the body data and the channel parameters in the embodiment includes: Extracting the impact features in the channel parameters; the impact features refer to the features that affect epilepsy prediction; Integrating the body data, the plurality of channel IDs, and the plurality of impact features corresponding to the plurality of channel IDs into channel weight analysis data; Obtaining the feature weight coefficients corresponding to the plurality of channel IDs by using the channel weight generation model on the channel weight analysis data; the channel weight generation model is constructed by using an artificial intelligence model, and is used to analyze the comprehensive data of the patient to predict the influence degree of the plurality of impact features in each channel ID on epilepsy prediction; Fusing the plurality of impact features in the same channel ID by using the feature weight coefficients corresponding to the channel ID to obtain the channel fusion features corresponding to the channel ID; Creating a feature matrix based on the total number of channel IDs; Placing the channel fusion features corresponding to the channel ID in the corresponding position in the feature matrix to obtain a dynamic feature matrix.

[0028] The channel weight generation model in the embodiment is constructed by using an artificial intelligence model, including: Obtaining a plurality of historical channel weight analysis data, and the corresponding historical channel IDs, historical feature weight coefficients corresponding to the historical channel IDs; The historical channel weight analysis data, the corresponding historical channel ID, and the historical feature weight coefficient corresponding to the historical channel ID are divided into training data, verification data, and test data; and the training data, the verification data, and the test data are preprocessed to obtain a training set, a verification set, and a test set; the ratio between the training set, the test set, and the verification set is 7:2:1; The artificial intelligence model is selected as the base model; the artificial intelligence model includes a BP model and the like; The base model is trained by using the training set, and the learning rate and other hyperparameters are adjusted on the verification set to obtain a pre-trained model; The pre-trained model is verified on the test set, and finally a channel weight generation model with the channel weight analysis data as the input and the feature weight coefficient corresponding to the channel ID as the output is obtained.

[0029] In the embodiment, the influence features of each channel ID in the electroencephalogram signal data are analyzed in real time, and the pre-trained channel weight generation model is used to dynamically generate the feature weight coefficient corresponding to each influence feature in each channel ID, so that the influence features of each channel in different states can reflect differentiated influence, thereby enhancing the expression ability and adaptability of the features in each channel ID, providing more accurate and differentiated data support for the input of the epilepsy prediction model, improving the prediction accuracy, and enhancing the operation efficiency of the model.

[0030] The epilepsy prediction model in the embodiment is constructed by using a graph convolution network model, including: A plurality of historical dynamic electroencephalogram topological graphs and corresponding historical epilepsy data are obtained; The plurality of historical dynamic electroencephalogram topological graphs and the corresponding historical epilepsy data are divided into training data, verification data, and test data; and the training data, the verification data, and the test data are preprocessed to obtain a training set, a verification set, and a test set; the ratio between the training set, the test set, and the verification set is 7:2:1; The graph convolution network model is selected as the base model; The base model is trained by using the training set, and the learning rate and other hyperparameters are adjusted on the verification set to obtain a pre-trained model; The pre-trained model is verified on the test set, and finally an epilepsy prediction model with the dynamic electroencephalogram topological graph as the input and the epilepsy prediction data as the output is obtained.

[0031] In the embodiment, the pre-warning signal is generated based on the epilepsy prediction data, including: Body data in the patient data is obtained; The epilepsy prediction probability in the epilepsy prediction data is extracted; generate a severe seizure warning signal when the seizure prediction probability is greater than or equal to D times the personalized probability threshold; the personalized probability threshold is generated by the body data; generate a seizure warning signal when the seizure prediction probability is less than D times the personalized probability threshold and greater than or equal to the personalized probability threshold; where D is a proportionality coefficient, D>1; the specific value is set according to experience, and D is set to 1.4 in the embodiment.

[0032] The personalized probability threshold in the embodiment is generated by the body data, including: extract a number of historical seizure prediction probabilities LDYG of the body data in a historical time window, and a number of historical prediction probabilities LYS greater than or equal to E times the baseline probability threshold; the historical time window refers to a period of time in the past, which is set to 5 minutes in the embodiment; where E represents a proportionality coefficient, E∈(0, 1); the specific value is set according to experience, and E is set to 0.7 in the embodiment; The baseline probability threshold satisfies the following calculation formula: ; where PG represents the average of the seizure probabilities of a number of other epilepsy patients with the same epilepsy status as the patient, which is calculated by the formula ; h represents a sensitivity coefficient, h>0; the specific value is set according to experience, and h is set to 2 in the embodiment; n represents the number of the other epilepsy patients, and N represents the total number of the other epilepsy patients, represents the seizure probability of the nth epilepsy patient; the embodiment selects the epilepsy data of the other epilepsy patients to calculate the baseline probability threshold, which can effectively avoid the situation of false negative or false positive caused by blindly setting the seizure probability threshold when the new patient lacks historical data; when selecting the other epilepsy patients, the consistency between the other epilepsy patients and the current patient in age and gender is considered, so that the calculated baseline seizure threshold has commonality; The personalized probability threshold satisfies the following calculation formula: ; where β represents a dynamic sensitivity coefficient, β∈(0, 1); the specific value is set according to experience, and β is set to 0.5 in the embodiment; represents the risk accumulation intensity, which is calculated by the formula ; where m represents the number corresponding to a number of historical seizure prediction probabilities in the historical time window, and M is the total number of the number of historical seizure prediction probabilities in the historical time window; max() represents the maximum value operation; BY represents a volatility suppression factor, which is calculated by the formula ; where m and a both represent the number corresponding to a number of historical seizure prediction probabilities in the historical time window.

[0033] The embodiment dynamically adjusts the threshold corresponding to the epilepsy prediction probability by combining the historical physical data of the patient, so that the threshold can be individually associated with the individual characteristics of the patient, effectively avoids the false alarm or missed alarm problem caused by the unified threshold, significantly improves the accuracy and reliability of the early warning signal, and improves the efficiency of the early warning generation, thereby providing strong support for precise and efficient epilepsy prediction.

[0034] Please refer to Figure 2 The second aspect embodiment of the present application provides a graph convolution network-based epilepsy prediction system, comprising a data acquisition module, a data analysis module and an early warning module; the data acquisition module is electrically and / or communicatively connected with the data analysis module; the data analysis module is electrically and / or communicatively connected with the early warning module; The data acquisition module: acquires patient data through a data acquisition device; the data acquisition device includes a plurality of sensors and the like; The data analysis module includes a dynamic topology construction unit and a signal generation unit; The dynamic topology construction unit: constructs a dynamic electroencephalogram topology map based on the patient data; The signal generation unit: inputs the dynamic electroencephalogram topology map into an epilepsy prediction model to obtain epilepsy prediction data; the epilepsy prediction model is constructed by a graph convolution network model, is used for analyzing the dynamic electroencephalogram topology map of the patient to obtain the epilepsy prediction data of the patient, and generates an early warning signal based on the epilepsy prediction data; The early warning module: makes a corresponding prompt based on the early warning signal; the early warning signal includes a severe epilepsy attack alarm signal and an epilepsy attack early warning signal.

[0035] Another aspect embodiment of the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement a graph convolution network-based epilepsy prediction system according to the second aspect embodiment of the present application.

[0036] Some data in the above formula is calculated by removing the dimension and taking the numerical value, and the formula is obtained by software simulation of a large amount of collected data to obtain a formula closest to the real situation; the preset parameters and the preset threshold in the formula are set by a person skilled in the art according to the actual situation or obtained by a large amount of data simulation.

[0037] The working principle of the present application is: acquiring patient data; constructing a dynamic electroencephalogram topology graph based on the patient data; inputting the dynamic electroencephalogram topology graph into an epilepsy prediction model to obtain epilepsy prediction data; generating an early warning signal based on the epilepsy prediction data, and dynamically adjusting the patient's electroencephalogram topology graph within the time window range based on the patient's electroencephalogram signal data, so that when analyzing epilepsy, not only the multi-dimensional features within the channel are focused on, but also the connection relationship between channels is focused on, the position and probability of epilepsy occurrence are accurately predicted, the accuracy of the epilepsy prediction method is improved, and the problem that the prior art often analyzes based on the multi-dimensional features in the multi-channel of the electroencephalogram signal, ignores the spatial topology relationship between channels, and the epilepsy prediction cannot accurately identify the lesion position, thereby leading to low accuracy of the epilepsy prediction method, is solved.

[0038] The above embodiments are only used to illustrate the technical method of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. An epilepsy prediction method based on graph convolutional network, characterized in that: include: Acquire patient data; the patient data includes patient ID, physical data and EEG signal data; Construct dynamic EEG topology maps based on patient data; Inputting the dynamic EEG topology map into an epilepsy prediction model to obtain epilepsy prediction data; the epilepsy prediction model is constructed by a graph convolutional network model, which is used to analyze the patient's dynamic EEG topology map to obtain the patient's epilepsy prediction data; Generate early warning signals based on epilepsy prediction data.

2. The epilepsy prediction method based on graph convolutional network according to claim 1, characterized in that: The method of constructing a dynamic EEG topology map based on patient data includes: Extracting EEG signal data from patient data within a time window; the EEG signal data includes a number of channel IDs and their corresponding channel parameters; Generate dynamic adjacency matrix based on channel ID; generating a dynamic feature matrix based on body data and channel parameters; The dynamic EEG topology map includes a dynamic adjacency matrix and a dynamic feature matrix.

3. The epilepsy prediction method based on graph convolutional network according to claim 2, characterized in that: Generating a dynamic adjacency matrix based on the channel ID includes: Extract the phase lock value PLV, phase lag index PLI, Euclidean distance OJ between each channel ID and the channel signal-to-noise ratio SNR corresponding to each channel ID; Calculate the connection weight between each channel ID through the formula ; The calculation formula of the connection weight satisfies: ; Where i and j are the numbers corresponding to the channel ID; It is represented as the average neighbor distance corresponding to the i-th channel ID, k is represented as the number of the channel ID connected to the i-th channel ID, and K is represented as the total number of channel IDs connected to the i-th channel ID; It is expressed as a phase synchronization function, which is constructed by the nonlinear relationship between the phase lock value PLV, the phase lag index PLI and the phase synchronization coefficient. The phase synchronization function satisfies the following calculation formula: ; Among them, i and j represent the numbers corresponding to the channel ID, Expressed as the frequency band difference penalty coefficient, ∈(0,1); It is expressed as the channel dominant frequency band difference between the i-th channel ID and the j-th channel ID; Substituting the phase lock value and phase lag index between each channel ID into the phase synchronization function can obtain the phase synchronization coefficient between each channel ID; Extract the connection weights between each channel ID and create an adjacency matrix based on the total number of channel IDs; The dynamic adjacency matrix can be obtained by resetting the connection weights between each channel ID into the corresponding positions in the adjacency matrix.

4. The epilepsy prediction method based on graph convolutional network according to claim 2, characterized in that: The generating of a dynamic feature matrix based on body data and channel parameters includes: Extracting influencing features from channel parameters; Integrate body data, several channel IDs, and several influencing features corresponding to the several channel IDs into channel weight analysis data; The channel weight analysis data is used to generate a channel weight model to obtain feature weight coefficients corresponding to several channel IDs; the channel weight generation model is constructed using an artificial intelligence model to analyze the patient's comprehensive data to predict the degree of influence of several influencing features in each channel ID on epilepsy prediction; The channel fusion feature corresponding to the channel ID is obtained by fusing several influencing features within the same channel ID through the feature weight coefficient corresponding to the channel ID; Create a feature matrix based on the total number of channel IDs; The channel fusion features corresponding to the channel ID are placed in the corresponding position in the feature matrix to obtain the dynamic feature matrix.

5. The epilepsy prediction method based on graph convolutional network according to claim 4, characterized in that: The channel weight generation model is constructed through an artificial intelligence model, including: Obtain several historical channel weight analysis data and their corresponding historical channel IDs, and historical feature weight coefficients corresponding to the historical channel IDs; Divide a number of historical channel weight analysis data and their corresponding historical channel IDs, and the historical feature weight coefficients corresponding to the historical channel IDs into training data, verification data, and test data; and perform data preprocessing on the training data, verification data, and test data to obtain a training set, a verification set, and a test set; Select an artificial intelligence model as the base model; Train the basic model using the training set, and adjust the learning rate and other hyperparameters on the validation set to obtain the pre-trained model; By verifying the pre-trained model on the test set, we finally obtain a channel weight generation model whose input is channel weight analysis data and output is several feature weight coefficients corresponding to the channel ID.

6. The epilepsy prediction method based on graph convolutional network according to claim 1, characterized in that: The epilepsy prediction model is constructed using a graph convolutional network model, including: Obtain several historical dynamic EEG topological maps and their corresponding historical epilepsy data; Dividing several historical dynamic EEG topological maps and their corresponding historical epilepsy data into training data, verification data, and test data; and performing data preprocessing on the training data, verification data, and test data to obtain a training set, a verification set, and a test set; Select the graph convolutional network model as the base model; Train the basic model using the training set, and adjust the learning rate and other hyperparameters on the validation set to obtain the pre-trained model; By verifying the pre-trained model on the test set, we finally obtained an epilepsy prediction model whose input is a dynamic EEG topology map and whose output is epilepsy prediction data.

7. The epilepsy prediction method based on graph convolutional network according to claim 1, characterized in that: The generating of the warning signal based on the epilepsy prediction data comprises: Obtaining physical data from patient data; Extracting epilepsy prediction probability from epilepsy prediction data; When the predicted probability of epilepsy is greater than or equal to D times the personalized probability threshold, a severe epilepsy alarm signal is generated; the personalized probability threshold is generated by physical data; When the predicted probability of epilepsy is less than D times the personalized probability threshold and the predicted probability of epilepsy is greater than or equal to the personalized probability threshold, an epilepsy onset warning signal is generated; where D is the proportional coefficient, D>1.

8. The epilepsy prediction method based on graph convolutional network according to claim 7, characterized in that: The personalized probability threshold is generated by physical data, including: Extracting several historical epilepsy prediction probabilities LDYG of the body data within the historical time window and several historical prediction numbers LYS of epilepsy prediction probabilities greater than or equal to E times the baseline probability threshold; where E represents a proportional coefficient, E∈(0,1); The baseline probability threshold satisfies the following calculation formula: PG is the mean probability of epilepsy of several other epileptic patients with the same epileptic state as the patient, which is expressed by the formula Calculation is performed; h represents the sensitivity coefficient, h>0; n represents the number of other epilepsy patients, N represents the total number of other epilepsy patients, It is expressed as the probability of epilepsy for the nth epilepsy patient; The personalized probability threshold satisfies the following calculation formula: ; Where β represents the dynamic sensitivity coefficient, β∈(0,1); Expressed as the cumulative risk intensity, through the formula Calculation is performed, where m represents the number corresponding to several historical epilepsy prediction probabilities within the historical time window, M represents the total number of several historical epilepsy prediction probabilities within the historical time window; max() represents the maximum value operation; BY represents the volatility suppression factor, which is calculated by the formula, and the formula satisfies ; Among them, m and a both represent the numbers corresponding to several historical epilepsy prediction probabilities within the historical time window.

9. An epilepsy prediction system based on a graph convolutional network, applied to an epilepsy prediction method based on a graph convolutional network according to any one of claims 1 to 8, characterized in that: include: Interconnected data acquisition module and data analysis module; The data acquisition module is used to acquire patient data through a data acquisition device; The data analysis module includes a dynamic topology construction unit and a signal generation unit; The dynamic topology construction unit is used to construct a dynamic EEG topology map based on patient data; The signal generating unit: inputs the dynamic EEG topology map into the epilepsy prediction model to obtain epilepsy prediction data; the epilepsy prediction model is constructed through a graph convolutional network model, which is used to analyze the patient's dynamic EEG topology map to obtain the patient's epilepsy prediction data; and generates a warning signal based on the epilepsy prediction data.

10. A computer-readable storage medium, applied to the epilepsy prediction system based on graph convolutional network according to claim 9, characterized in that: The computer-readable storage medium stores a computer program.

Images