Epilepsy prediction system based on multivariate weighted joint recursion and graph attention network

Abstract

The application provides an epilepsy prediction system based on multi-element weighted joint recursion and graph attention network, which can be applied to the fields of biomedical signal processing and artificial intelligence technology. The system comprises a brain function imaging module configured to acquire electroencephalogram signal data of a target object under authorization of the target object; a processor comprising a multi-element weighted joint recursion processing unit configured to obtain a phase space trajectory vector of each channel of the electroencephalogram signal data according to the electroencephalogram signal data, construct a recursion graph according to the phase space trajectory vector of each channel, obtain a channel correlation coefficient between any two channels based on the recursion graph of each channel, and construct a weighted adjacency matrix representing the brain function of the target object according to the channel correlation coefficient between any two channels; and a graph attention network processing unit configured to perform spatiotemporal feature processing according to the weighted adjacency matrix and the electroencephalogram signal data and output an epilepsy seizure prediction result.

Description

Technical Field

[0001] This application relates to the fields of biomedical signal processing and artificial intelligence, and more specifically, to an epilepsy prediction system based on multivariate weighted joint recursion and graph attention networks. Background Technology

[0002] Epilepsy is a chronic brain disorder caused by abnormal electrical activity in the brain's neurons. Electroencephalography (EEG) signals can record information about the activity of these neurons. In the event of an epileptic seizure, the abnormal electrical discharges in the brain's neurons will exhibit specific waveform changes in the EEG signal. Therefore, by recording changes in the electrical signals of brain neurons using EEG signals, it is possible to predict whether a subject will experience an epileptic seizure. However, the accuracy of predicting epileptic seizures using EEG signals in related technologies is relatively low. Summary of the Invention

[0003] In view of this, this application provides an epilepsy prediction system based on multivariate weighted joint recursion and graph attention networks.

[0004] One aspect of this application provides an epilepsy prediction system based on multivariate weighted joint recursion and graph attention networks. The system includes: a brain functional imaging module configured to acquire electroencephalogram (EEG) signal data of the target object upon authorization; and a processor including: a multivariate weighted joint recursion processing unit configured to obtain the phase space trajectory vector of each channel of the EEG signal data based on the EEG signal data, construct a recursive graph corresponding to each channel based on the phase space trajectory vector of each channel, obtain the channel correlation coefficient between any two channels based on their respective recursive graphs, and construct a weighted adjacency matrix representing the brain function of the target object based on the channel correlation coefficient between any two channels; and a graph attention network processing unit configured to perform spatiotemporal feature processing on the weighted adjacency matrix and the EEG signal data, and output a prediction result of epileptic seizures for the target object.

[0005] According to an embodiment of this application, the aforementioned multi-weighted joint recursive processing unit includes: a phase space reconstruction subunit, configured to perform dimensional mapping on the aforementioned electroencephalogram (EEG) signal data according to a predetermined embedding dimension and a delay time, to obtain the aforementioned phase space trajectory vector of each of the aforementioned channels of the aforementioned EEG signal data, wherein the aforementioned embedding dimension characterizes the mapping dimension of the aforementioned EEG signal data, and the aforementioned delay time represents the time interval between data selected from the aforementioned EEG signal data.

[0006] According to an embodiment of this application, the multi-weighted joint recursive processing unit further includes a recursive graph generation subunit, configured to construct a recursive graph corresponding to each of the channels based on a pre-set recursion rate threshold and the distance between any two points in the phase space trajectory vector of each of the channels.

[0007] According to an embodiment of this application, the aforementioned multivariate weighted joint recursive processing unit further includes: a matrix generation subunit, configured to obtain a joint recursive graph between any two channels based on the recursive graphs corresponding to each of the two channels, obtain the channel correlation coefficients between any two channels based on the joint recursive graph, and construct a weighted adjacency matrix representing the brain function of the target object based on the channel correlation coefficients between any two channels.

[0008] According to an embodiment of this application, the graph attention network processing unit includes: a graph attention subunit, configured to determine a second channel that is a neighbor channel of the first channel based on the weighted adjacency matrix and a first channel among the plurality of channels, to perform spatiotemporal feature extraction on the electroencephalogram (EEG) signal data based on the first channel and the second channel to obtain EEG spatiotemporal feature data, and to determine the attention weight between the first channel and the second channel based on the EEG spatiotemporal feature data.

[0009] According to an embodiment of this application, the graph attention subunit is configured to: determine a second channel that is a neighboring channel of the first channel based on the weighted adjacency matrix and a first channel among the plurality of channels; determine EEG feature channel data corresponding to the first channel and the second channel respectively from the EEG signal data; perform spatiotemporal feature extraction on the EEG feature channel data corresponding to the first channel and the second channel respectively to obtain EEG spatiotemporal feature data of the first channel and the second channel respectively; and determine the attention weight between the first channel and the second channel based on the EEG spatiotemporal feature data.

[0010] According to an embodiment of this application, the graph attention network processing unit includes a multi-head attention subunit, comprising multiple attention head networks connected in parallel, configured to input the attention weights between the first channel and the second channel and the electroencephalogram spatiotemporal feature data of the first channel and the second channel respectively into the multiple attention head networks, and output the channel features of the first channel.

[0011] According to an embodiment of this application, the graph attention network processing unit includes a result classification subunit configured to output the epileptic seizure prediction result for the target object based on the channel characteristics of the first channel.

[0012] According to an embodiment of this application, the processor further includes a preprocessing unit configured to, upon authorization by the target object, acquire initial EEG signal data of the target object, and segment the initial EEG signal data according to a preset time interval to obtain the EEG signal data.

[0013] According to an embodiment of this application, the processor is further configured to acquire EEG signal sample data and EEG signal label data of the sample object, input the EEG sample data into an initial model to obtain the epilepsy prediction sample result of the sample object, and train the initial model based on the target loss function and the EEG signal label data to obtain a trained epilepsy prediction model based on a multivariate weighted joint recursive and graph attention network.

[0014] According to embodiments of this application, a recursive graph for each channel of the EEG signal data is constructed through a multi-weighted joint recursive processing unit. Based on the recursive graphs of any two channels, channel correlation coefficients are obtained. Based on the channel correlation coefficients, a weighted adjacency matrix representing the brain function of the target object is constructed. Then, a graph attention network processing unit processes the weighted adjacency matrix and the EEG signal data to output the epilepsy prediction result of the target object. By generating a nonlinear recursive graph and a weighted adjacency matrix from the EEG signal data, the connection relationship between each channel of the target object's EEG signal data can be better reflected, and the brain functional state can be more accurately represented. This provides a more reliable and information-rich input for spatiotemporal feature processing, overcomes the deficiency of incomplete feature extraction of EEG signals in related technologies, and improves the accuracy of the epilepsy prediction result of the target object. Attached Figure Description

[0015] The above and other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0016] Figure 1 A block diagram of an epilepsy prediction system based on a multivariate weighted joint recursive and graph attention network according to an embodiment of this application is shown.

[0017] Figure 2 A block diagram of the multivariate weighted joint recursive processing unit of an epilepsy prediction system based on multivariate weighted joint recursion and graph attention network according to an embodiment of this application is shown.

[0018] Figure 3 A block diagram of the graph attention network processing unit of an epilepsy prediction system based on multivariate weighted joint recursion and graph attention network according to an embodiment of this application is shown.

[0019] Figure 4A block diagram of an epilepsy prediction system based on a multivariate weighted joint recursive and graph attention network according to another embodiment of this application is shown. Detailed Implementation

[0020] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0021] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0022] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0023] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0024] In the embodiments of this application, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security and network security.

[0025] In the embodiments of this application, the user's authorization or consent was obtained before obtaining or collecting the user's personal information.

[0026] EEG-based seizure prediction is mainly divided into traditional methods and deep learning methods. Traditional methods typically rely on time-domain, frequency-domain, or time-frequency-domain features in EEG signals for analysis. The feature extraction process of traditional methods requires manual design and selection, which is not only time-consuming and labor-intensive but also highly dependent on the operator's expertise and prior assumptions. This limits the breadth and automation of the analysis and makes it difficult to discover deeper, more complex patterns in the data.

[0027] To achieve automatic feature extraction, deep learning models, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, have been introduced in recent years. Deep learning methods are primarily used to process Euclidean structured data (such as time series or images), while EEG signals are acquired through multiple channels, and the distribution of electrodes on the scalp for each channel exhibits a non-Euclidean topological network structure. Deep learning models cannot effectively utilize the spatial relationships and functional connectivity information between electrodes, i.e., they ignore the inherent spatial characteristics of EEG signals.

[0028] To address the aforementioned issue of neglecting spatial information, Graph Neural Networks (GNNs), particularly Graph Convolutional Networks (GCNs), have been proposed. In one example, a brain functional network is constructed, and then GCNs are used to extract spatiotemporal features from EEG signals for epileptic seizure prediction. While these methods consider spatial information, they still suffer from the following problems: 1. Fixed neighbor node aggregation weights: When aggregating neighbor node information, GCNs typically use fixed weights predefined by the graph structure (e.g., based on the adjacency matrix or its normalized form). This approach cannot adaptively adjust the importance of different neighbor nodes according to their own characteristics and task context, reducing the model's flexibility and expressive power, and may fail to accurately capture the dynamic and complex interactions between key brain regions in the target subject during epileptic states. 2. Relatively simplistic functional network construction methods: Some methods use methods such as Phase Locking Values ​​(PLVs) to construct brain functional networks. While these can reflect the correlation between channels, they may not fully capture the complex generalized synchronicity when processing non-stationary and nonlinear EEG signals.

[0029] Therefore, how to construct a brain functional network that can more accurately reflect the nonlinear dynamic characteristics, and on this basis, adaptively integrate the spatiotemporal features of multi-channel EEG signals, is an urgent problem to be solved to improve the accuracy of predicting epileptic seizures.

[0030] In view of this, embodiments of this application provide an epilepsy prediction system based on multivariate weighted joint recursion and graph attention network. The system includes: a brain functional imaging module configured to acquire electroencephalogram (EEG) signal data of the target object with authorization; a processor including: a multivariate weighted joint recursion processing unit configured to obtain the phase space trajectory vector of each channel of the EEG signal data based on the EEG signal data, construct a recursion graph corresponding to each channel based on the phase space trajectory vector of each channel, obtain the channel correlation coefficient between any two channels based on the recursion graphs of any two channels, and construct a weighted adjacency matrix representing the brain function of the target object based on the channel correlation coefficient between any two channels; and a graph attention network processing unit configured to perform spatiotemporal feature processing on the weighted adjacency matrix and the EEG signal data, and output a prediction result of epileptic seizures for the target object.

[0031] Figure 1 A block diagram of an epilepsy prediction system based on a multivariate weighted joint recursive and graph attention network according to an embodiment of this application is shown.

[0032] like Figure 1 As shown, the epilepsy prediction system 100 based on multivariate weighted joint recursion and graph attention networks may include a brain functional imaging module 110 and a processor 120. The processor 120 includes a multivariate weighted joint recursion processing unit 121 and a graph attention network processing unit 122.

[0033] The brain functional imaging module 110 can be configured to acquire the electroencephalogram (EEG) signal data of the target object when authorized by the target object.

[0034] The multivariate weighted joint recursive processing unit 121 can be configured to obtain the phase space trajectory vector of each channel of the EEG signal data based on the EEG signal data, construct a recursive graph corresponding to each channel based on the phase space trajectory vector of each channel, obtain the channel correlation coefficient between any two channels based on the recursive graphs of any two channels, and construct a weighted adjacency matrix representing the brain function of the target object based on the channel correlation coefficient between any two channels.

[0035] The graph attention network processing unit 122 can be configured to perform spatiotemporal feature processing based on the weighted adjacency matrix and EEG signal data, and output a prediction result of epileptic seizures about the target object.

[0036] Electroencephalography (EEG) signal data can be obtained by recording the electrical activity of neurons in the brain of a target subject through electrodes, with the subject's authorization. EEG signal data can be multi-channel, with each channel corresponding to one electrode.

[0037] Since EEG signal data is nonlinear, and the Multivariate Weighted Joint Recurrence Network (MWJRN) is a modeling method for multivariate nonlinear dynamic systems based on recursive analysis, it can better extract the correlations between nonlinear data by integrating the joint recurrence matrix and network science theory. Therefore, MWJRN can be trained using multi-channel EEG sample data from sample subjects to obtain a multivariate weighted joint recurrence processing unit, which can then extract the nonlinear correlations between EEG signal data across channels.

[0038] Phase space reconstruction can be performed on the data of each channel in the EEG signal data to obtain the phase space trajectory vector, which can represent the trajectory vector of the EEG signal data in the high-dimensional phase space.

[0039] A recursive graph can be constructed based on the phase space trajectory vector of each channel. The recursive graph can represent the distance between any two points in the phase space trajectory vector. Based on the recursive graphs of any number of channels, the channel correlation coefficient between any two channels can be obtained. The channel correlation coefficient can represent the correlation between any two channels.

[0040] Based on the channel correlation coefficient between any two channels, a weighted adjacency matrix representing the brain function of the target object can be constructed. Each element in the weighted adjacency matrix represents the correlation between any two channels.

[0041] The graph attention network (GAT) processing unit 122 can be trained using the EEG sample data of the sample object to obtain the graph attention network processing unit 122. The graph attention network processing unit 122 can extract the spatiotemporal features of the EEG signal data from the weighted adjacency matrix, thereby outputting the epileptic seizure prediction result of the target object.

[0042] According to embodiments of this application, a recursive graph for each channel of the EEG signal data is constructed through a multi-weighted joint recursive processing unit. Based on the recursive graphs of any two channels, channel correlation coefficients are obtained. Based on the channel correlation coefficients, a weighted adjacency matrix representing the brain function of the target object is constructed. Then, a graph attention network processing unit processes the weighted adjacency matrix and the EEG signal data to output the epilepsy prediction result of the target object. By generating a nonlinear recursive graph and a weighted adjacency matrix from the EEG signal data, the connection relationship between each channel of the target object's EEG signal data can be better reflected, and the brain functional state can be more accurately represented. This provides a more reliable and information-rich input for spatiotemporal feature processing, overcomes the deficiency of incomplete feature extraction of EEG signals in related technologies, and improves the accuracy of the epilepsy prediction result of the target object.

[0043] Figure 2 A block diagram of the multivariate weighted joint recursive processing unit of an epilepsy prediction system based on multivariate weighted joint recursion and graph attention network according to an embodiment of this application is shown.

[0044] like Figure 2 As shown, the multi-weighted joint recursive processing unit 121 may include a phase space reconstruction subunit 1211, a recursive graph generation subunit 1212, and a matrix generation subunit 1213.

[0045] According to an embodiment of this application, the phase space reconstruction subunit 1211 can be configured to perform dimensional mapping on the EEG signal data according to a predetermined embedding dimension and delay time to obtain the phase space trajectory vector of each channel of the EEG signal data, wherein the embedding dimension represents the mapping dimension of the EEG signal data, and the delay time represents the time interval between data selected from the EEG signal data.

[0046] The EEG signal data for each channel can be represented as follows: ,in, Represents the first of P channels indivual, Indicates the first The EEG signal data from each channel contains One data point, Indicated The first of the data indivual, Indicates the first In each channel The first of the data Data points.

[0047] It can be done through the nearest neighbor method and Algorithms determine the embedding dimension and delay time. Alternatively, pseudo-nearest neighbor (FNN) methods or mutual information methods can be used to determine the embedding dimension and delay time. One-dimensional time-series EEG signal data can be mapped to phase space trajectory vectors in a high-dimensional phase space. These phase space trajectory vectors consist of multiple vector points, each of which can be represented by the following formula.

[0048] (1);

[0049] in, Indicates the first The phase space trajectory vectors of each channel are related to the EEG signal data. The t-th data in The corresponding vector point, Indicates the embedding dimension. Indicates the delay time. This indicates that the t-th data Starting from For m data points with a step size, .

[0050] According to an embodiment of this application, the recursion graph generation subunit 1212 can be configured to construct a recursion graph corresponding to each channel based on a pre-set recursion rate threshold and the distance between any two points in the phase space trajectory vector of each channel.

[0051] A recurrence plot can be constructed for the phase space trajectory vector of each channel. A pre-set recurrence rate threshold can be used; this threshold can be fixed or adaptively adjusted. For example, the recurrence rate threshold could be 0.1, 0.2, etc. The recurrence rate threshold determines whether any two points in the phase space trajectory vector are adjacent. The recurrence plot (RP) corresponding to each channel, constructed based on the recurrence rate threshold and the distance between any two points in the phase space trajectory vector of each channel, is represented by the following formula.

[0052] (2);

[0053] in, Indicates according to the first The first channel in the phase space trajectory vector Point and the The distance between points and the first Recursion rate threshold for each channel The resulting recursion graph, Indicates the first The first channel in the phase space trajectory vector point, No. The first channel in the phase space trajectory vector point, , , This represents the total number of vector points in the phase space trajectory vector. Let represent the Heaviside step function.

[0054] According to embodiments of this application, recursion graphs can visualize the recursive characteristics of dynamic systems and can more effectively obtain the internal correlations of nonlinear electroencephalogram (EEG) signal data.

[0055] According to an embodiment of this application, the matrix generation subunit 1213 can be configured to obtain a joint recursive graph between any two channels based on the recursive graphs corresponding to each of the two channels, obtain the channel correlation coefficient between any two channels based on the joint recursive graph, and construct a weighted adjacency matrix representing the brain function of the target object based on the channel correlation coefficient between any two channels.

[0056] For any two channels, their joint recurrence plot (JRP) can be calculated, which represents the events of recurrence occurring simultaneously in both channels. This can be expressed by the following formula.

[0057] (3);

[0058] in, Indicates the first Recursive graph of channels and the Recursive graph of channels The joint recursive graph between them Indicates according to the first The first channel in the phase space trajectory vector Point and the The distance between points and the first Recursion rate threshold for each channel The resulting recursion graph, Indicates according to the first The first channel in the phase space trajectory vector Point and the The distance between points and the first Recursion rate threshold for each channel The resulting recursion graph.

[0059] Based on the joint recurrence graph, the channel correlation coefficient between two channels can be calculated. The channel correlation coefficient can also be expressed as the joint recurrence rate (JRR), as shown in the following formula.

[0060] (4);

[0061] in, Indicates the first The first channel and the first The channel correlation coefficient between channels can also be called the Generalized Synchronization Index (GSI).

[0062] By calculating the channel correlation coefficients between all channels, a weighted adjacency matrix of size P is constructed, which can be represented as follows: Each element in the weighted adjacency matrix represents the channel correlation coefficient between the channel represented by that row and the channel represented by that column.

[0063] Compared with linear methods such as PLV, the weighted adjacency matrix constructed based on channel correlation coefficients can more robustly capture the generalized phase synchronization between non-stationary and nonlinear time series in EEG signal data, thereby more accurately constructing the functional connectivity between different brain regions in the target subject's epileptic seizure state.

[0064] Figure 3 A block diagram of the graph attention network processing unit of an epilepsy prediction system based on multivariate weighted joint recursion and graph attention network according to an embodiment of this application is shown.

[0065] like Figure 3 As shown, the graph attention network processing unit 122 may include a graph attention subunit 1221, a multi-head attention subunit 1222, and a result classification subunit 1223.

[0066] According to an embodiment of this application, the graph attention subunit 1221 can be configured to determine a second channel that is a neighbor channel of the first channel based on a weighted adjacency matrix and a first channel among multiple channels, extract spatiotemporal features of EEG signal data based on the first and second channels to obtain EEG spatiotemporal feature data, and determine the attention weight between the first channel and the second channel based on the EEG spatiotemporal feature data.

[0067] Since each element in the weighted adjacency matrix represents the channel correlation coefficient between the channel represented in that row and the channel represented in that column, the second channels that are neighbors of the first channel can be determined based on the size of each element in the weighted adjacency matrix. The first channel can be any of multiple channels. For example, if there are six channels: k1, k2, k3, k4, k5, and k6, channel k1 can be considered the first channel. The channel correlation coefficients between the first channel and channels k2, k3, k4, k5, and k6 are 4.3, 0.2, 5.8, 1.8, and 2.6, respectively. A channel correlation coefficient threshold of 2 can be set; channels with correlation coefficients greater than the threshold are considered neighbors, otherwise, channels k2, k4, and k6 are determined as the second channels that are neighbors of the first channel. The number of channels and the channel correlation coefficient threshold can also be other values, which are not limited here.

[0068] Spatiotemporal features of EEG signal data can be extracted using convolutional computation to obtain spatiotemporal feature data of EEG. Furthermore, convolutional computation can be used to calculate the attention weights between the first and second channels of the EEG spatiotemporal feature data. These attention weights characterize the degree of correlation between the first and second channels, and the channel features of the first channel can be calculated using these attention weights.

[0069] According to an embodiment of this application, the graph attention subunit 1221 can be configured to determine a second channel that is a neighbor channel of the first channel based on a weighted adjacency matrix and a first channel among multiple channels, determine the EEG feature channel data corresponding to the first channel and the second channel respectively from the EEG signal data, perform spatiotemporal feature extraction on the EEG feature channel data corresponding to the first channel and the second channel respectively to obtain the EEG spatiotemporal feature data of the first channel and the second channel respectively, and determine the attention weight between the first channel and the second channel based on the EEG spatiotemporal feature data.

[0070] The graph attention subunit takes a weighted adjacency matrix and EEG signal data as input. First, it processes the feature vector of the first channel in the EEG signal data through a shared, learnable linear transformation, generating a new set of feature vectors. This transformation aims to map the raw features in the EEG signal data to a more expressive feature space. For the first channel, the graph attention subunit uses the channel correlation coefficients in the adjacency matrix to find its neighboring channel, the second channel. Before calculating the attention weights for the first and second channels, the graph attention subunit performs spatiotemporal feature extraction on the EEG feature channel data corresponding to each channel. For example, it can concatenate the EEG feature channel data corresponding to the first and second channels to obtain their respective spatiotemporal EEG feature data. Based on these spatiotemporal EEG feature data, the attention weights for the first and second channels are then derived.

[0071] The first channel can be any channel; for example, it could be the... There are 10 channels, therefore, the feature vector of the corresponding EEG feature channel data can be... The second channel can be the first Each channel, and the feature vector of the corresponding EEG feature channel data, can be... ,right and Spatiotemporal feature extraction was performed to obtain the spatiotemporal feature data of the first channel of EEG and the spatiotemporal feature data of the second channel of EEG.

[0072] Therefore, for the first channel, the attention coefficient between it and the second channel... Calculate using the following formula.

[0073] (5);

[0074] in, This represents the shared linear transformation weight matrix. Indicates the relationship with the first The spatiotemporal feature vectors corresponding to the spatiotemporal feature data of each channel of EEG. Indicates the relationship with the first The spatiotemporal feature vectors corresponding to the spatiotemporal feature data of each channel of EEG. This indicates a splicing operation. Represents the weight vector of a single-layer feedforward neural network The transpose of .

[0075] Then, the softmax function is used to normalize all attention coefficients to obtain the final attention weights. The formula is as follows.

[0076] (6);

[0077] in, Indicates the first The set of all neighboring channels of a given channel.

[0078] Unlike GCN's fixed weights, GAT can adaptively learn the importance between channels based on data and tasks, enabling the network to focus on the functional connections that contribute the most to epileptic seizure prediction, thereby improving the accuracy of feature extraction and the network's expressive power.

[0079] According to an embodiment of this application, the multi-head attention subunit 1222 may include multiple attention head networks connected in parallel, configured to input the attention weights between the first channel and the second channel and the electroencephalogram spatiotemporal feature data of the first channel and the second channel respectively into the multiple attention head networks, and output the channel features of the first channel.

[0080] To make the learning process more stable and capture features in different channels, a multi-head attention mechanism is adopted. By inputting the attention weights between the first and second channels and the EEG spatiotemporal feature data of the first and second channels respectively into M independent attention head networks (each "head" has its own parameters), the channel features of the first channel are output.

[0081] (7);

[0082] in, Indicates the first Channel characteristics of each channel, Indicates the first The attention head network calculates the first... The first channel and the first Normalized attention weights between channels Indicates the first The attention head network calculates the first... The spatiotemporal feature vectors of each channel, This represents a non-linear activation function.

[0083] According to an embodiment of this application, the result classification subunit 1223 can be configured to output a prediction result of epileptic seizures of the target object based on the channel characteristics of the first channel.

[0084] The classification subunit may include pooling layers and fully connected layers. The channel feature set of the first channel, obtained after processing by multiple attention head networks, is aggregated into the features of the entire weighted adjacency matrix through a readout layer (such as average pooling). This is expressed by the following formula.

[0085] (8);

[0086] in, This represents the characteristics of the entire weighted adjacency matrix. Indicates the first Channel characteristics of each channel.

[0087] Finally, the characteristics of the weighted adjacency matrix are... The input is fed into a fully connected layer (MLP) and a softmax layer, and the output is the final epileptic seizure prediction result, such as "pre-seizure" or "interictal".

[0088] (9);

[0089] in, This indicates the prediction result of epileptic seizures. For example, in When the value is 1, the epileptic seizure prediction result can characterize the target subject as being in the pre-seizure stage. When the value is 0, the epileptic seizure prediction result can characterize the target object as the epileptic interictal period.

[0090] The classification subunit can further capture the temporal dynamics of the signal and obtain channel features by using recurrent neural networks (RNNs) or their variants LSTM and Gated Recurrent Units (GRUs). In addition to average pooling, the readout layer for the features of the weighted adjacency matrix can also employ max pooling or more complex weighted summation methods.

[0091] According to an embodiment of this application, the processor may further include a preprocessing unit configured to, upon authorization by the target object, acquire initial EEG signal data of the target object, and segment the initial EEG signal data according to a preset time interval to obtain EEG signal data.

[0092] The initial EEG signal data can be directly acquired, therefore it needs to be processed. This can be done by segmenting the initial EEG signal data according to a preset time interval, resulting in a one-dimensional time series of EEG signal data. For example, the preset time interval could be 5 seconds. Filtering and other processing can also be applied to the initial EEG signal data.

[0093] According to embodiments of this application, by segmenting the initial electroencephalogram (EEG) signal data, the data length is reduced, thereby reducing the amount of computation per data and improving computational efficiency.

[0094] According to embodiments of this application, the processor can also be configured to acquire EEG signal sample data and EEG signal label data of the sample object, input the EEG sample data into the initial model to obtain the epilepsy prediction sample result of the sample object, and train the initial model based on the target loss function and the EEG signal label data to obtain a trained epilepsy prediction model based on a multivariate weighted joint recursive and graph attention network.

[0095] It can acquire EEG signal sample data and EEG signal label data of the sample subjects, and divide them into preictal (positive samples) and interictal (negative samples). It can also segment the EEG signal sample data. For example, EEG signal sample data from 35 minutes to 5 minutes before a seizure can be defined as preictal, and EEG signal sample data from at least 4 hours before the onset of a seizure and at least 4 hours after the end of a seizure can be defined as interictal. The EEG signal sample data can be divided into fixed-length segments (e.g., 5 seconds), and the problem of imbalanced EEG signal sample data can be addressed (e.g., by increasing minority class samples through a sliding window or reducing majority class samples through random undersampling).

[0096] EEG sample data can be input into the initial model to obtain epilepsy prediction results for the sample object. The initial model includes an initial multivariate weighted joint recursive network and an initial graph attention network. Based on the target loss function and EEG signal label data, the initial model is trained, that is, the initial multivariate weighted joint recursive network and the initial graph attention network are trained to obtain a trained epilepsy prediction model based on multivariate weighted joint recursive and graph attention networks. The epilepsy prediction model based on multivariate weighted joint recursive and graph attention networks can be used by the processor to predict epileptic seizures of the target object.

[0097] As an example, EEG signal sample data and EEG signal label data from 18 channels common to all sample patients can be selected. The sampling frequency of the EEG signal sample data and EEG signal label data is 256Hz, and the length of the segmented EEG signal sample data is 5 seconds (1280 data points). The GAT network can be set up with 2 graph attention layers, each layer using 6 attention heads. During training, the Adam optimizer can be used with a learning rate of 10 per second. −3 The batch size is 64, the discard rate is 0.1, and the L2 regularization coefficient is 0.1.

[0098] Figure 4 A block diagram of an epilepsy prediction system based on a multivariate weighted joint recursive and graph attention network according to another embodiment of this application is shown.

[0099] like Figure 4As shown, the epilepsy prediction system 100 based on multivariate weighted joint recursion and graph attention networks may include a brain functional imaging module 110 and a processor 120. The processor 120 may include a multivariate weighted joint recursion processing unit 121, a graph attention network processing unit 122, and a preprocessing unit 123. The multivariate weighted joint recursion processing unit 121 may include a phase space reconstruction subunit 1211, a recursive graph generation subunit 1212, and a matrix generation subunit 1213. The graph attention network processing unit 122 may include a graph attention subunit 1221, a multi-head attention subunit 1222, and a result classification subunit 1223. Figure 4 For further explanation of the various parts, please refer to the descriptions of other embodiments of this application, which will not be repeated here.

[0100] The brain functional network constructed by the epilepsy prediction system based on multivariate weighted joint recursion and graph attention network in this application embodiment can more accurately reflect the real situation of brain functional connectivity in the epileptic state, providing a more reliable and information-rich input for subsequent feature extraction, and overcoming the shortcomings of traditional methods in capturing signal characteristics incompletely.

[0101] This application combines the MWJRN network and the GAT network. The MWJRN network provides high-quality graph structure input, while the GAT network performs efficient adaptive feature learning on this high-quality graph. The two complement each other, achieving deep fusion of the spatiotemporal features of EEG signals.

[0102] The MWJRN-GAT method of this application, along with Support Vector Machine (SVM), Bi-directional Long Short-Term Memory (Bi-LSTM), EEG-based network (EEGNet), GCN, and other methods, were tested on the dataset. The results are shown in Table 1 below.

[0103] Table 1

[0104]

[0105] The MWJRN-GAT method proposed in this application achieved an average accuracy of 98.26% and an average sensitivity of 99.01%. Compared with various methods such as SVM, Bi-LSTM, and GCN, its performance is significantly improved. The significant advantage in sensitivity, in particular, means that the system proposed in this application can more reliably predict the pre-seizure state with an extremely low false negative rate. This is crucial in clinical applications, effectively providing subjects with early warning and intervention time, and reducing the risk of seizures. Furthermore, the MWJRN-GAT method proposed in this application also demonstrates good generalization ability across trials.

[0106] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments of this application can be combined and / or combined in various ways. All such combinations and / or combinations fall within the scope of this application.

[0107] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. An epilepsy prediction system based on multivariate weighted joint recursive and graph attention networks, characterized in that, The system includes: The brain functional imaging module is configured to acquire the electroencephalogram (EEG) signal data of the target object when the target object authorizes the acquisition. Processor, including: The multivariate weighted joint recursive processing unit is configured to obtain the phase space trajectory vector of each channel of the EEG signal data based on the EEG signal data, construct a recursive graph corresponding to each channel based on the phase space trajectory vector of each channel, obtain the channel correlation coefficient between any two channels based on the recursive graphs of any two channels, and construct a weighted adjacency matrix representing the brain function of the target object based on the channel correlation coefficient between any two channels. The graph attention network processing unit is configured to perform spatiotemporal feature processing based on the weighted adjacency matrix and the electroencephalogram signal data, and output a prediction result of epileptic seizures for the target object.

2. The system according to claim 1, characterized in that, The multi-weighted joint recursive processing unit includes: The phase space reconstruction subunit is configured to perform dimensional mapping on the EEG signal data according to a predetermined embedding dimension and delay time to obtain the phase space trajectory vector of each channel of the EEG signal data, wherein the embedding dimension represents the mapping dimension of the EEG signal data, and the delay time represents the time interval between data selected from the EEG signal data.

3. The system according to claim 1, characterized in that, The multi-weighted joint recursive processing unit further includes: The recursion graph generation subunit is configured to construct a recursion graph corresponding to each channel based on a pre-set recursion rate threshold and the distance between any two points in the phase space trajectory vector of each channel.

4. The system according to claim 3, characterized in that, The multi-weighted joint recursive processing unit further includes: The matrix generation subunit is configured to obtain a joint recursive graph between any two channels based on the recursive graph corresponding to each of the two channels, obtain the channel correlation coefficient between any two channels based on the joint recursive graph, and construct a weighted adjacency matrix representing the brain function of the target object based on the channel correlation coefficient between any two channels.

5. The system according to claim 1, characterized in that, The graph attention network processing unit includes: The graph attention subunit is configured to determine a second channel that is a neighbor of the first channel based on the weighted adjacency matrix and a first channel among the multiple channels, extract spatiotemporal features from the EEG signal data based on the first channel and the second channel to obtain EEG spatiotemporal feature data, and determine the attention weight between the first channel and the second channel based on the EEG spatiotemporal feature data.

6. The system according to claim 5, characterized in that, The graph attention subunit is configured to: determine a second channel that is a neighboring channel of the first channel based on the weighted adjacency matrix and a first channel among the multiple channels; determine EEG feature channel data corresponding to the first channel and the second channel respectively from the EEG signal data; perform spatiotemporal feature extraction on the EEG feature channel data corresponding to the first channel and the second channel respectively to obtain EEG spatiotemporal feature data of the first channel and the second channel respectively; and determine the attention weight between the first channel and the second channel based on the EEG spatiotemporal feature data.

7. The system according to claim 6, characterized in that, The graph attention network processing unit includes: The multi-head attention subunit includes multiple attention head networks connected in parallel, configured to input the attention weights between the first channel and the second channel and the electroencephalogram spatiotemporal feature data of the first channel and the second channel respectively into the multiple attention head networks, and output the channel features of the first channel.

8. The system according to claim 7, characterized in that, The graph attention network processing unit includes: The result classification subunit is configured to output the epileptic seizure prediction result for the target object based on the channel characteristics of the first channel.

9. The system according to claim 1, characterized in that, The processor also includes: The preprocessing unit is configured to, upon authorization by the target object, acquire the initial EEG signal data of the target object, and segment the initial EEG signal data according to a preset time interval to obtain the EEG signal data.

10. The system according to claim 1, characterized in that, The processor is further configured to acquire EEG signal sample data and EEG signal label data of the sample object, input the EEG signal sample data into an initial model to obtain the epilepsy prediction sample result of the sample object, and train the initial model based on the target loss function and the EEG signal label data to obtain a trained epilepsy prediction model based on a multivariate weighted joint recursive and graph attention network.

Images