Electroencephalogram epilepsy detection method and system based on node adaptive graph neural network

Abstract

The invention discloses an electroencephalogram epilepsy detection method and system based on a node adaptive graph neural network, relates to a computer system based on a biological model, and provides the scheme for solving the problems of graph structure immobilization and the like in the prior art. The method comprises the following steps: an electroencephalogram signal acquisition and preprocessing step; constructing a hybrid EEG graph; optimizing a self-adaptive residual image; node specific diffusion convolution is carried out; modeling time sequence characteristics; and performing classified output. The system comprises a data acquisition module, a mixed graph construction module, a self-adaptive mapping module, a node specific convolution module, a time sequence modeling module and a classification output module. When the system runs, the steps of the method are executed, so that the electroencephalogram epilepsy detection function based on the node adaptive graph neural network is realized. The method has the technical advantages that (1) graph structure self-learning is carried out; (2) carrying out brain region personalized modeling, and strengthening region feature expression; and (3) combining space-time dependence modeling, and completely depicting the epilepsy dynamic process.

Description

TECHNICAL FIELD

[0001] The present application relates to a computer system based on a biological model, in particular to an electroencephalogram epilepsy detection method and system based on a node adaptive graph neural network. BACKGROUND

[0002] Epilepsy is a chronic neurological disease caused by abnormal discharge of brain neurons, affecting about 50 million people worldwide. Epilepsy seizures have various manifestations, ranging from brief loss of consciousness to severe tonic-clonic convulsions. Accurate identification of seizures and their types in the clinic is an important basis for guiding individualized antiepileptic drug therapy and surgical options.

[0003] Electroencephalogram (EEG) as a non-invasive neural electrical activity monitoring method has high temporal resolution, real-time performance, and repeatable acquisition, and is the core data source for seizure detection and type classification.

[0004] However, traditional EEG signal analysis still mainly relies on artificial visual interpretation by neurophysiological experts, which has the following obvious problems: (1) Strong subjectivity and low efficiency: artificial interpretation of EEG requires frame-by-frame browsing of signals for several hours or even tens of hours, which is a huge workload and has poor consistency; (2) Significant experience dependence: different experts have different standards for judging epileptic discharges (such as spikes and sharp waves), which affects the accuracy and repeatability of detection; (3) Difficult to achieve real-time monitoring: during long-term video EEG monitoring (VEEG), artificial review is extremely costly and difficult to meet the real-time seizure detection and clinical monitoring needs.

[0005] To improve the efficiency of EEG signal analysis, researchers have attempted to introduce machine learning and deep learning methods into EEG automatic analysis. Early methods were mostly based on handcrafted features (such as frequency band energy, sample entropy, Hjorth parameters, etc.) and combined with traditional classifiers such as support vector machines (SVM), decision trees, or random forests to achieve seizure detection. Although this method has improved the degree of automation to some extent, its feature design relies on experience, has poor generalization ability, and is difficult to maintain stable performance under different patients and acquisition conditions.

[0006] With the development of deep learning, Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN / LSTM) have been widely applied in automatic feature extraction and time series modeling of electroencephalogram (EEG). CNN can capture the local spatial and spectral features of EEG segments, while LSTM model can model the long-term dependence of the signal. However, such networks all assume that the input data is distributed on a regular Euclidean grid structure, while the spatial distribution of EEG electrodes on the scalp surface presents an irregular topological structure, and the functional connections between different brain regions also exhibit non-Euclidean relationships. Therefore, traditional CNN or RNN models are difficult to fully depict the topological dependence and functional interaction between EEG channels, limiting their modeling ability of dynamic activities between brain regions.

[0007] In recent years, the proposal of Graph Neural Network (GNN) has provided a new technical means for processing non-Euclidean structure data. Researchers regard EEG channels as graph nodes and define the spatial distance or signal correlation between channels as graph edges, and realize the propagation of information between brain regions through graph convolution operation. Typical methods include Diffusion Convolutional Recurrent Neural Network (DCRNN) and Spatio-Temporal Graph Convolutional Network (STGCN), etc. These methods have made certain progress in seizure detection tasks, but still have the following limitations: (1) Graph structure fixation: Existing models usually rely on static graph construction methods (such as generating a fixed adjacency matrix based on Euclidean distance or static Pearson correlation coefficient of electrodes), which cannot adaptively adjust the graph structure according to the differences between patients, state changes or dynamic characteristics of seizure stages, limiting the individualization and time-varying adaptation ability of the model.

[0008] (2) Node parameter sharing problem: Most GNN models use uniform convolution parameters for all nodes, without considering the differences in neural activity patterns and physiological functions of different brain regions, thus weakening the model's performance in capturing local brain region features and heterogeneous seizure patterns.

[0009] (3) Insufficient modeling of time series dependence: Some GNN methods only focus on spatial structure information and fail to fully couple with time series dynamic changes, making it difficult to fully depict the propagation rules of brain regions before and after seizures.

[0010] (4) Lack of clinical interpretability: The features output by existing models often lack a corresponding relationship with physiological brain regions or clinical indicators, limiting their application value in assisting diagnosis and clinical decision-making.

[0011] In summary, the existing epilepsy detection technology still has obvious deficiencies in adaptive modeling of brain region connection structure, node-specific parameter learning, and spatio-temporal joint feature extraction, which limits the detection accuracy, generalization ability and clinical interpretability.

[0012] Therefore, there is an urgent need for a new electroencephalogram intelligent analysis method that can adaptively learn the brain region connection relationship while maintaining physiological structure constraints, support node-level differentiated modeling, and realize spatio-temporal dependence joint representation to improve the accuracy, robustness and clinical practicability of seizure detection and type classification. SUMMARY

[0013] The present application aims to provide a node-adaptive graph neural network-based electroencephalogram epilepsy detection method and system to solve the problems existing in the prior art.

[0014] The node-adaptive graph neural network-based electroencephalogram epilepsy detection method in the present application comprises the following steps: S100: Electroencephalogram signal acquisition and preprocessing step: acquiring multi-channel electroencephalogram signals of the subject, and preprocessing the original signals; S200: Hybrid EEG graph construction step: establishing a hybrid graph that fuses anatomical structure and functional correlation between EEG channels to represent the spatial and functional dependence between brain regions; S300: Adaptive residual graph optimization step: for the dynamic changes of brain region connection patterns of different patients and time periods, an adaptive residual graph optimization mechanism is introduced to refine the initial graph structure through node embedding learning; S400: Node-specific diffusion convolution step: using a node-specific diffusion convolution mechanism to enable different brain regions to have independent convolution kernel parameters; S500: Temporal feature modeling step: introducing a diffusion graph convolution recurrent unit based on node-specific diffusion convolution to capture the dynamic characteristics of brain region activity changes over time and realize temporal modeling of the seizure process; the diffusion graph convolution recurrent unit embeds a diffusion convolution operator in a gated recurrent structure, enabling the model to simultaneously model the spatial diffusion between brain regions and the evolution of signals over time, and realizing joint representation of electroencephalogram signal features in the spatio-temporal domain; S600: Classification output step.

[0015] The node-adaptive graph neural network-based electroencephalogram epilepsy detection system in the present application comprises the following functional modules: Data acquisition module: for receiving multi-channel EEG signals input by an electroencephalogram acquisition device and completing preprocessing operations; Hybrid graph construction module: for calculating the spatial and functional similarity between electrodes and generating a hybrid initial adjacency matrix; Adaptive mapping module: residual optimization based on node embedding, dynamically adjusting the connection relationship of brain regions; Node-specific convolution module: perform node-specific diffusion convolution to realize differential modeling at the brain region level; Temporal modeling module: adopt diffusion convolution recurrent structure to capture the time evolution law of seizures; Classification output module: perform seizure detection or type identification based on global electroencephalogram features, and output prediction results; When the electroencephalogram seizure detection system is running, steps S100 to S600 in the method are sequentially executed, thereby realizing the electroencephalogram seizure detection function based on the node adaptive graph neural network.

[0016] The electroencephalogram seizure detection method and system based on the node adaptive graph neural network have the following technical advantages: (1) Graph structure self-learning, overcoming static limitations. Through the residual optimization mechanism driven by node embedding, the adaptive learning and dynamic adjustment of the connection relationship of brain regions are realized. The adjacency matrix structure can be automatically corrected according to the electroencephalogram features of different subjects and time segments during training, thereby overcoming the adaptability problem caused by relying on fixed graph structure in traditional methods.

[0017] (2) Brain region personalized modeling, strengthening regional feature expression. Introduce node-specific parameter learning mechanism in the diffusion graph convolution framework, so that each electroencephalogram channel can independently learn convolution weights and feature transformation functions. It fully reflects the physiological differences and functional heterogeneity between brain regions, which helps the model to accurately capture the feature patterns and participation degree of different brain regions in seizures.

[0018] (3) Joint spatiotemporal dependence modeling, complete description of the dynamic process of epilepsy. Introduce recurrent neural structure based on spatial diffusion convolution to build a spatiotemporal joint modeling framework. It can not only simulate the diffusion propagation law of electroencephalogram signals in space, but also capture the dynamic changes of time series before and after seizures, realizing unified modeling of epilepsy activity in both time and space dimensions. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flowchart of the electroencephalogram seizure detection method in the present application.

[0020] Figure 2 is a flowchart of step S200 of the electroencephalogram seizure detection method in the present application.

[0021] Figure 3 is a flowchart of step S300 of the electroencephalogram seizure detection method in the present application.

[0022] Figure 4is a flowchart of step S400 of the electroencephalogram epilepsy detection method described in the present application.

[0023] Figure 5 is a flowchart of step S500 of the electroencephalogram epilepsy detection method described in the present application.

[0024] Figure 6 is a structural diagram of the electroencephalogram epilepsy detection system described in the present application. DETAILED DESCRIPTION

[0025] As shown in Figures 1 to 5 , the electroencephalogram epilepsy detection method based on node adaptive graph neural network described in the present application includes the following steps: S100: Electroencephalogram signal acquisition and preprocessing step; Acquire the multi-channel electroencephalogram signal of the subject, and preprocess the original signal: 0.5-70Hz band-pass filtering and 50 / 60Hz notch filtering are used to remove the direct current drift and power frequency noise in the original signal; Independent component analysis (ICA) is used to remove artifacts such as electrooculogram and electromyogram.

[0026] Divide into segments according to a fixed time window, forming an electroencephalogram signal tensor

[0027] Wherein represents the number of electroencephalogram channels, represents the number of sampling points of each segment; represents the signal sequence of the th channel in the time window.

[0028] Z-score standardization is performed on each segment of signal to reduce the signal amplitude difference between different subjects, wherein is the mean value of channel , and is the standard deviation of channel .

[0029] S200: Hybrid EEG graph construction step; A hybrid graph that integrates anatomical structure and functional correlation is established between electroencephalogram channels to represent the spatial and functional dependence between brain regions. Specifically, it includes the following sub-steps: S201: Calculate the Euclidean distance between electrodes to obtain a spatial similarity matrix ; The three-dimensional space coordinates of the th channel corresponding electrode are represented as The spatial distance between any two channels is defined as:

[0030] The spatial similarity is defined as:

[0031] wherein denotes the channel and the spatial connection strength with channel j; is the scale parameter of the Gaussian kernel, used to control the distance decay rate.

[0032] S202: Calculate the functional similarity matrix between the electroencephalogram channels : The functional similarity is defined based on the Pearson correlation coefficient:

[0033] wherein is the channel signal at time step t, is the channel average value in the segment, the value range of denotes the channel , and the correlation strength between signals j.

[0034] S203: Fusion of spatial similarity and functional similarity to obtain a hybrid initial adjacency matrix :

[0035] wherein, is a learnable weight matrix, used to control the fusion proportion of the spatial graph; is a learnable weight matrix, used to control the fusion proportion of the functional graph; is a bias matrix; is a Sigmoid function, used for normalization output; is an element-wise multiplication.

[0036] S300: Adaptive residual graph optimization step In view of the dynamic changes of brain region connection modes of different patients and time periods, the application designs a residual adaptive graph optimization mechanism, and refines the initial graph structure through node embedding learning. Specifically, the following sub-steps are included: S301: Define the node embedding matrix wherein is the node embedding dimension, and the row vector represents the latent representation of the i-th node; th node. S302: Calculate node similarity matrix :

[0037] wherein is a linear rectification function to ensure the similarity non-negative.

[0038] S303: Project the similarity through a learnable mapping matrix to form the residual with the initial graph:

[0039] S304: Update to obtain the adaptive adjacency matrix:

[0040] wherein is an adaptive correction term, is the final adaptive adjacency matrix.

[0041] Through this residual form, the model only needs to learn the relative correction amount, thereby realizing the individual adjustment of the brain region connection relationship under the premise of ensuring the structural stability.

[0042] S400: Node-specific diffusion convolution step; To solve the problem of node parameter sharing in traditional graph convolution, the application proposes a node-specific diffusion convolution mechanism, which enables different brain regions to have independent convolution kernel parameters. Specifically, the following sub-steps are included: S401: Use the node row vector to input a lightweight multi-layer perceptron to generate a set of independent diffusion convolution kernel parameters for each node

[0043] wherein denotes the k-order diffusion weight matrix of node , is the input feature dimension, is the output feature dimension.

[0044] S402: Perform bidirectional diffusion convolution on the adaptive adjacency matrix :

[0045] wherein is the node output feature matrix, is the diffusion order, is the forward diffusion kernel, is the reverse diffusion kernel.

[0046] S500: a time sequence feature modeling step; To realize the time-dependent modeling of the electroencephalogram signal, the application introduces a diffusion graph convolution recurrent unit on the basis of node-specific diffusion convolution, which is used to capture the dynamic characteristics of the brain activity changing over time and realize the time sequence modeling of the epilepsy seizure process. The diffusion graph convolution recurrent unit, DCGRU, can model the spatial diffusion between brain regions and the evolution of the signal over time by embedding a diffusion convolution operator in a gated recurrent structure, so as to realize the joint representation of the electroencephalogram signal features in the time-space domain. Specifically, the following sub-steps are included: S501: update gate calculation; The update gate is used to control the proportion of the current brain region neural state from the historical state.

[0047] At the time step , the calculation formula of the update gate is:

[0048] wherein denotes the current time step, denotes the electroencephalogram input feature matrix at the current time, denotes the hidden state at the last time step, denotes the node-specific diffusion convolution operation performed on the adaptive adjacency matrix , , is the input weight matrix of the update gate, is the state weight matrix of the update gate, is a Sigmoid function, is the update gate matrix.

[0049] In the analysis of the electroencephalogram signal, the update gate controls the retention degree of the model to the old brain region state: when the electroencephalogram is in a stable stage, the value is small, and the model retains the historical state; the stable stage is the non-seizure period. When abnormal discharge occurs in the brain region, the value increases, and the model quickly updates the features, so as to capture the sudden changes of the epilepsy seizure; the abnormal discharge is the seizure period.

[0050] S502: reset gate calculation; The reset gate is used to adjust the influence degree of the state at the last time on the current candidate state, and the calculation formula is:

[0051] wherein is the input weight matrix of the reset gate, To reset the state weight matrix of the gate, other symbols are defined as above.

[0052] In the analysis of EEG signals, the value of the reset gate reflects the historical inhibitory relationship between brain regions: when a brain region is irrelevant to the current seizure, the corresponding value tends to 0, inhibiting the influence of irrelevant channels; when a brain region participates in seizure propagation, the value increases, enhancing the feature update strength of the region.

[0053] S503: Candidate hidden state calculation; The candidate hidden state is generated by the current input feature and the previous state adjusted by the reset gate, and its calculation formula is:

[0054] wherein, is the diffusion convolution weight matrix corresponding to the input signal, used to project the input feature X t at the current time into the hidden feature space; is the diffusion convolution weight matrix corresponding to the historical hidden state, used to perform state feature time propagation and update on the adaptive adjacency matrix ; represents element-wise multiplication; is the hyperbolic tangent function; represents node-specific diffusion convolution operation .

[0055] In EEG modeling, represents the potential neural activity pattern of the brain region at the current time. By introducing node-specific diffusion convolution operation , each brain region not only depends on its historical signal when updating its own state, but also combines the dynamic connection weight with other brain regions to realize feature diffusion and collaborative update between brain regions, thereby describing the propagation process of epileptic signals in the brain region network.

[0056] S504: Hidden state update; The hidden state update formula of the diffusion graph convolution recurrent unit is:

[0057] wherein represents the hidden state feature matrix at time , and other symbols are defined as above.

[0058] In the semantics of EEG signals, the two parts in the formula have the following meanings: represent the memory and retention of the model for the historical brain region state, represent the response and update of the model to the current sudden neural activity.

[0059] Through dynamic adjustment of the gating proportion, the model can flexibly balance historical memory and new activation according to the electroencephalogram features at different time stages, and effectively depict the occurrence, propagation and recovery process of the seizure.

[0060] S600: a classification output step; specifically comprising the following sub-steps: S601: after the time sequence feature modeling is completed, the classification output is performed.

[0061] The hidden state feature matrix of the time T is Global average pooling is performed to obtain a segment-level feature vector:

[0062] S602: the global feature is input into a fully connected layer, and classification is performed through a Softmax function, and a seizure detection or epilepsy type discrimination result is output:

[0063] wherein is an output probability vector; is the number of categories, in the detection task , in the classification task ; is an output layer weight; is a bias parameter.

[0064] The application also provides an electroencephalogram epilepsy detection system for implementing the above method, as shown in Figure 6 The system comprises the following functional modules: A data acquisition module: used for receiving multi-channel EEG signals input by an electroencephalogram acquisition device, and performing filtering, segmentation, standardization and other preprocessing operations; A hybrid graph construction module: used for calculating spatial and functional similarity between electrodes, and generating a hybrid initial adjacency matrix; An adaptive graph construction module: based on node embedding, residual optimization is performed to dynamically adjust the connection relationship between brain regions; A node-specific convolution module: performing node-specific diffusion convolution to realize differential modeling at the brain region level; A time sequence modeling module: adopting a diffusion convolution recurrent structure to capture the time evolution law of the seizure; A classification output module: based on global electroencephalogram features, seizure detection or type recognition is performed, and a prediction result is output.

[0065] Each module can be a software module or a hardware circuit module. When the electroencephalogram epilepsy detection system is running, steps S100 to S600 are sequentially executed, so as to realize the electroencephalogram epilepsy detection function based on the node adaptive graph neural network.

[0066] For those skilled in the art, other various corresponding changes and modifications can be made according to the above-described technical solutions and concepts, and all these changes and modifications should belong to the protection scope of the claims of the present application.

Claims

1. A method for detecting epilepsy via electroencephalography based on node adaptive graph neural networks, characterized in that, Includes the following steps: S100: EEG signal acquisition and preprocessing steps: Acquire multi-channel EEG signals from the subject and preprocess the raw signals; S200: Hybrid EEG Map Construction Steps: A hybrid map integrating anatomical structure and functional correlation is established between brain electrical channels to characterize the spatial and functional dependencies between brain regions. S300: Adaptive Residual Graph Optimization Steps: To address the dynamic changes in brain region connectivity patterns across different patients and time periods, an adaptive residual graph optimization mechanism is introduced, refining the initial graph structure through node embedding learning; S400: Node-specific diffusing convolution steps: Utilizing the node-specific diffusing convolution mechanism, different brain regions are equipped with independent convolution kernel parameters; S500: Temporal feature modeling steps: Based on node-specific diffusion convolution, a diffusion graph convolution recurrent unit is introduced to capture the dynamic features of brain region activity changing over time, thereby realizing the temporal modeling of the epileptic seizure process; the diffusion graph convolution recurrent unit embeds a diffusion convolution operator in a gated recurrent structure, enabling the model to simultaneously model the spatial diffusion between brain regions and the evolution of signals over time, thereby achieving a joint representation of EEG signal features in the spatiotemporal domain; S600: Classification output steps.

2. The method for detecting epilepsy based on node adaptive graph neural networks according to claim 1, characterized in that, Step S100 specifically involves: using a 0.5–70 Hz bandpass filter and a 50 / 60 Hz notch filter to remove DC drift and power frequency noise from the original signal; using independent component analysis to remove artifacts; and dividing the signal into segments according to a fixed time window to form an EEG signal tensor. ; in Indicates the number of brainwave channels. This indicates the number of sampling points for each segment; Indicates the first The signal sequence of each channel within this time window; Z-score normalization is performed on each signal segment. To reduce the difference in signal amplitude among different subjects, among which For channel The mean, For channel The standard deviation.

3. The method for detecting epilepsy based on node adaptive graph neural networks according to claim 2, characterized in that, Step S200 includes the following sub-steps: S201: Calculate the Euclidean distance between the electrodes to obtain the spatial similarity matrix. ; No. The electrodes corresponding to each channel are represented in three-dimensional space coordinates as follows: The spatial distance between any two channels is defined as: Spatial similarity is defined as: in Indicates channel Spatial connection strength with channel j; The scale parameter of the Gaussian kernel is used to control the distance decay rate; S202: Calculate the functional similarity matrix between EEG channels : Functional similarity, based on the Pearson correlation coefficient, is defined as follows: in For channel The signal at time step t, For channel The average value of this segment, The range of values ​​is , indicating channel The correlation strength between j signals; S203: Combine spatial similarity and functional similarity to obtain a hybrid initial adjacency matrix. : in, This is a learnable weight matrix used to control the fusion ratio of the spatial graph; It is a learnable weight matrix used to control the fusion ratio of the function graphs; This is the bias matrix; This is the Sigmoid function, used to normalize the output; This is element-wise multiplication.

4. The method for detecting epilepsy based on node adaptive graph neural networks according to claim 3, characterized in that, Step S300 includes the following sub-steps: S301: Define the node embedding matrix ,in Embedding dimension for nodes, row vector Indicates the first The potential representation of each node; S302: Calculate the node similarity matrix : in It is a linear rectified function used to ensure that the similarity is non-negative; S303: Through a learnable mapping matrix Projection similarity, and forming residuals with the initial image: S304: Update to obtain the adaptive adjacency matrix: in For adaptive correction terms, This is the final adaptive adjacency matrix.

5. The method for detecting epilepsy based on node adaptive graph neural networks according to claim 4, characterized in that, Step S400 includes the following sub-steps: S401: Using node row vectors Input lightweight multilayer perceptron Generate a set of independent diffusing convolution kernel parameters for each node. in Represents a node The k-th order diffusion weight matrix, For the input feature dimension, Output feature dimension; S402: In the adaptive adjacency matrix Perform bidirectional diffusing convolution on top: in Output the feature matrix for each node. Let the diffusion order be , Forward-diffused nuclei, It is a reverse-diffusion nucleus.

6. The method for detecting epilepsy based on node adaptive graph neural networks according to claim 5, characterized in that, Step S500 includes the following sub-steps: S501: Update gate calculation; The update gate is used to control the proportion of the current neural state in a brain region that comes from a historical state; At time step The formula for updating the gate is: in Indicates the current time step. This represents the EEG input feature matrix at the current moment. This represents the hidden state of the previous time step. In the adaptive adjacency matrix Node-specific diffraction convolution operation performed on , To update the input weight matrix of the gate, To update the state weight matrix of the gate, For the Sigmoid function, To update the gate matrix; In electroencephalogram (EEG) signal analysis, The degree to which the control model retains the state of previous brain regions: when the EEG is in a stable phase, the model retains the historical state; when abnormal discharges occur in the brain region, As the value increases, the model updates its features, thereby capturing sudden changes in epileptic seizures; S502: Reset door calculation; The reset gate is used to adjust the influence of the previous state on the current candidate state. The calculation formula is as follows: in To reset the input weight matrix of the gate, To reset the state weight matrix of the gate; In EEG signal analysis, the value of the reset gate Reflecting the historical inhibitory relationship between brain regions: when the brain region is unrelated to the current seizure, the corresponding The value approaches 0 to suppress the influence of irrelevant channels; When brain regions are involved in seizure propagation Increasing the value enhances the feature update efficiency of this region. S503: Calculation of candidate hidden states; The candidate hidden state is generated jointly by the current input feature and the previous state after being adjusted by the reset gate. Its calculation formula is as follows: in, This is the spread convolution weight matrix corresponding to the input signal, used to weight the input feature X at the current time step. t Projected into the hidden feature space; The spread convolution weight matrix corresponds to the historical hidden states and is used in the adaptive adjacency matrix. The temporal propagation and updating of execution state characteristics; This represents element-wise multiplication; It is the hyperbolic tangent function; Represents node-specific diffraction convolution operation ; S504: Hidden state update; The hidden state update formula output by the final convolutional recurrent unit of the diffusion graph is: in Indicates time The latent state feature matrix.

7. The method for detecting epilepsy based on node adaptive graph neural networks according to claim 6, characterized in that, Includes the following sub-steps: S601: After the temporal feature modeling is completed, perform classification output; The feature matrix of the hidden state at time T Perform global average pooling to obtain fragment-level feature vectors: S602: Global features are input into the fully connected layer and classified using the Softmax function, outputting results for epileptic seizure detection or epilepsy type determination. in This is the output probability vector; For the number of categories, in the detection task In classification tasks ; These are the output layer weights; This is the bias parameter.

8. A brainwave epilepsy detection system based on node adaptive graph neural network, characterized in that, Includes the following functional modules: Data acquisition module: Used to receive multi-channel EEG signals input from the EEG acquisition device and perform preprocessing operations; Hybrid graph construction module: used to calculate the spatial and functional similarity between electrodes and generate a hybrid initial adjacency matrix; Adaptive mapping module: Performs residual optimization based on node embedding to dynamically adjust the connectivity of brain regions; Node-specific convolution module: Performs node-specific diffusing convolution to achieve differential modeling at the brain region level; Temporal modeling module: Employs a diffusing convolutional loop structure to capture the temporal evolution of epileptic seizures; Classification output module: Detects seizures or identifies seizure types based on global EEG features and outputs prediction results; When the EEG epilepsy detection system is running, steps S100 to S600 of the method as described in any one of claims 1 to 7 are executed sequentially to realize the EEG epilepsy detection function based on node adaptive graph neural network.

9. The EEG epilepsy detection system based on node adaptive graph neural network according to claim 8, characterized in that, Each functional module can be either a software module or a hardware circuit module.

Images