Seizure detection system based on dual-branch spatio-temporal graph neural network

Abstract

The application discloses a seizure detection system based on a double-branch spatio-temporal graph neural network, which comprises an electroencephalogram signal collector, a signal preprocessor, a double-branch graph constructor, a double-branch graph neural network feature learner, a feature fusion device and a classifier; wherein the double-branch graph constructor comprises a time branch and a space branch; the time branch adopts a horizontal visual graph algorithm to map the down-sampled electroencephalogram sequence into a graph structure for each channel; the space branch constructs node features based on multi-scale visual graph network statistical features and constructs edge weights based on a dispersion index and a cross-channel phase amplitude coupling. The double-branch graph neural network feature learner automatically learns time dynamic features and space correlation features from the two types of graph structures respectively and aggregates multi-channel information through an attention mechanism. The feature fusion device adopts an attention mechanism to weight and fuse the double-branch features. The application significantly improves the accuracy of the seizure detection through the double-branch architecture to fuse the spatio-temporal features and the graph neural network automatic learning.

Description

Technical Field

[0001] This invention belongs to the field of medical signal processing and artificial intelligence, and relates to an epileptic seizure detection system based on a bibranch spatiotemporal graph neural network. Background Technology

[0002] Epilepsy is a common chronic neurological disorder characterized by recurrent seizures caused by abnormal electrical discharges in the brain. Electroencephalography (EEG), as an important tool for detecting brain electrical activity, directly reflects the electrophysiological state of neurons and is a crucial basis for the diagnosis and monitoring of epilepsy. However, traditional manual interpretation methods require professional physicians to analyze long-term EEG recordings segment by segment, which is not only time-consuming and laborious but also easily affected by subjective factors, making it difficult to meet the needs of real-time clinical monitoring.

[0003] In recent years, significant progress has been made in automated epilepsy detection methods based on machine learning. Complex network theory has provided a new perspective for EEG signal analysis, where visualization algorithms can map one-dimensional time series signals into complex network structures, thereby revealing the nonlinear dynamics of the signals. Furthermore, phase-amplitude coupling (PAC), as a cross-frequency coupling analysis method, can reveal the interactions between oscillations in different frequency bands, which is of great value in epilepsy detection. However, existing methods have the following limitations:

[0004] First, feature extraction from graphs relies on manual design. Traditional methods require researchers to predefine network statistical indicators, making it difficult to fully extract discriminative information from the graph structure.

[0005] Second, single-scale visualizations provide limited information. Natural visualizations and horizontal visualizations characterize signal properties from different angles, and using either alone is insufficient to fully represent the nonlinear structure of a signal.

[0006] Third, the fusion of spatiotemporal features is insufficient. Existing methods either focus on dynamic features in the time dimension or on channel correlation in the spatial dimension, lacking an effective cross-frequency coupling information fusion mechanism.

[0007] Fourth, the model lacks interpretability. Clinical applications require understanding which brain regions and time patterns contribute most to the test results, and existing methods struggle to provide this type of interpretable information.

[0008] Therefore, it is necessary to propose a solution to address the technical problems existing in the current technology, given its technical deficiencies. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention proposes an epilepsy seizure detection system based on a dual-branch spatiotemporal graph neural network. By fusing temporal dynamic features and spatial correlation features through a dual-branch architecture, it achieves end-to-end automatic feature learning and epilepsy detection.

[0010] In order to solve the technical problems existing in the prior art, the technical solution of the present invention is as follows:

[0011] An epileptic seizure detection system based on a bibranch spatiotemporal graph neural network includes an EEG signal acquisition unit, a signal preprocessor, a bibranch graph builder, a bibranch graph neural network feature learner, a feature fusion unit, and a classifier.

[0012] The EEG signal acquisition device is used to acquire multi-channel EEG signals, i.e., raw EEG data;

[0013] The signal preprocessor is used to receive the raw EEG data from the EEG signal acquisition device and perform data preprocessing, including bandpass filtering and fixed time window segmentation of the multi-channel EEG signal; wherein, the raw EEG data is bandpass filtered from 0.5Hz to 45Hz and the processed EEG signal is segmented into segments.

[0014] The dual-branch graph builder is used to construct a dual-branch graph structure from the signal preprocessed by the signal preprocessor. It includes a temporal branch graph construction unit and a spatial branch graph construction unit. The temporal branch graph construction unit independently performs downsampling and horizontal visualization construction for each channel to generate multiple temporal graphs. The spatial branch graph construction unit constructs node features based on the statistical features of the multi-scale visualization network and constructs edge weights based on the dispersion index and cross-channel phase amplitude coupling to generate a spatial graph.

[0015] The bi-branch graph neural network feature learner is connected to the bi-branch graph builder, including a temporal branch graph convolutional network and a spatial branch graph convolutional network, which are used to automatically learn feature representations from the temporal graph and the spatial graph, respectively, and perform feature aggregation through an attention mechanism;

[0016] The feature fusion unit is used to perform attention-weighted fusion of features from the temporal and spatial branches;

[0017] A classifier is used to classify the fused features to obtain the epileptic seizure detection results.

[0018] In the above technical solution, the dual-branch graph builder is used to construct a dual-branch graph structure for the preprocessed signal; the time branch graph construction unit independently performs downsampling and horizontal visualization construction for each channel to generate 20 time graphs; the spatial branch graph construction unit constructs node features based on the network statistical features of the multi-scale visualization, constructs edge weights based on the dispersion index and cross-channel phase amplitude coupling, and generates 1 spatial graph.

[0019] The dual-branch graph neural network feature learner consists of a temporal branch graph convolutional network and a spatial branch graph convolutional network, which are used to automatically learn feature representations from the temporal graph and the spatial graph, respectively.

[0020] The feature fusion unit uses an attention mechanism to weightedly fuse features from two branches, with the fusion weights automatically determined through network learning.

[0021] The classifier performs binary classification of epileptic seizure states based on fused features.

[0022] As a further improvement, the EEG signal acquisition device uses the international 10-20 electrode system to acquire 20 channels of EEG signals at a sampling rate of 512Hz.

[0023] As a further improvement, each EEG signal segment has a time span of 4 seconds and 2048 sampling points.

[0024] As a further improvement, the time branch graph construction unit adopts the horizontal visibility graph algorithm. The steps are as follows:

[0025] S1: Downsampling of single-channel EEG segments was performed, and the 2048 sampling points were reduced to 128 nodes using a segmented averaging method;

[0026] S2: Based on S1, calculate the 4-dimensional feature vector of each node;

[0027] S3: Based on S2, construct the adjacency matrix according to the rules of the horizontal visibility map;

[0028] S4: Based on S3, the edge weights are calculated using a radian weighted function.

[0029] As a further improvement, a spatial branching map construction unit is used to construct a brain region connectivity map. The steps are as follows:

[0030] S10: Using each EEG channel as a node, extract 4 network statistical features from both the natural view and the horizontal view to form an 8-dimensional node feature vector;

[0031] S20: Based on S10, calculate the dispersion index and cross-channel phase amplitude coupling;

[0032] S30: Based on S20, the edge weights are determined by a weighted fusion of the two.

[0033] As a further improvement, the temporal branch uses a multi-layer graph convolutional network to automatically learn temporal dynamic features from the visual image and aggregates multi-channel information through an attention mechanism; the spatial branch uses a multi-layer graph convolutional network to learn spatial correlation features, and the attention pooling mechanism can adaptively discover key brain regions related to epileptic seizures.

[0034] As a further improvement, the obtained fused features are input into the classifier for classification. End-to-end training is performed using the cross-entropy loss function and the Adam optimizer, and accuracy, sensitivity, specificity, and F1 score are calculated.

[0035] Compared with the prior art, the present invention has the following technical effects:

[0036] (1) The present invention adopts a dual-branch complementary architecture. The time branch captures the temporal dynamics and nonlinear characteristics of the single-channel signal, and the spatial branch integrates multi-scale visual images and cross-frequency coupling information to comprehensively characterize the spatiotemporal characteristics of epileptic seizures.

[0037] (2) The graph neural network is used to automatically learn graph structure features, which overcomes the limitations of traditional methods that rely on manually designed features and can mine deeper and richer discriminative information in the graph structure.

[0038] (3) By integrating the dispersion index and cross-channel phase amplitude coupling to construct spatial graph edge weights, the abnormal synchronization and enhanced cross-frequency coupling phenomena in brain regions during epileptic seizures can be effectively captured.

[0039] (4) The multi-level attention mechanism provides interpretability of the model, including channel importance, brain region importance and spatiotemporal branch contribution, which helps clinical understanding of the key brain regions involved in epileptic seizures. Attached Figure Description

[0040] Figure 1 This is a block diagram of a seizure detection system based on a bibranch spatiotemporal graph neural network.

[0041] Figure 2 This is a flowchart of the epileptic seizure detection process in an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram of the construction of a dual-branch graph in an embodiment of the present invention.

[0043] Figure 4 This is a schematic diagram of the feature fusion and classification module in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0046] To improve the accuracy and interpretability of epileptic seizure detection, this invention proposes an epileptic seizure detection system based on a bibranch spatiotemporal graph neural network. (See [link to relevant documentation]). Figure 1 The diagram shown illustrates its principle, which includes an EEG signal acquisition unit, a signal preprocessor, a bibranch graph builder, a bibranch graph neural network feature learner, a feature fusion unit, and a classifier.

[0047] The EEG signal acquisition device is used to acquire multi-channel EEG signals, i.e., raw EEG data;

[0048] The signal preprocessor is used to receive the raw EEG data from the EEG signal acquisition device and perform data preprocessing, including bandpass filtering and fixed time window segmentation of the multi-channel EEG signals.

[0049] The bibranch graph builder is used to construct a bibranch graph structure from the signal preprocessor; wherein, the temporal branch graph building unit independently constructs a horizontal visibility map for each channel, and the spatial branch graph building unit constructs a brain region connectivity map based on multi-scale visibility map features and cross-frequency coupling.

[0050] The bi-branch graph neural network feature learner is connected to the bi-branch graph builder, including a temporal branch graph convolutional network and a spatial branch graph convolutional network, which are used to automatically learn feature representations from the temporal graph and the spatial graph, respectively;

[0051] The feature fusion unit is used to perform attention-weighted fusion of features from the temporal and spatial branches;

[0052] A classifier is used to classify the fused features to obtain the epileptic seizure detection results.

[0053] The main advantage of the dual-branch architecture in the above technical solution is that it can simultaneously capture the temporal dynamic features and spatial correlation features of EEG signals, integrate multi-scale visual information and cross-frequency coupling information, and comprehensively characterize the spatiotemporal characteristics of epileptic seizures.

[0054] See Figure 2 The diagram shows the epilepsy seizure detection flowchart of this invention. The EEG signal acquisition device adopts the international 10-20 electrode system, including 20 standard electrode channels, with a sampling rate of 512Hz. The data processing procedure of the signal preprocessor is as follows:

[0055] S1: Bandpass filtering of 0.5Hz~45Hz is applied to the acquired raw EEG data to remove baseline drift and breathing artifacts below 0.5Hz, as well as power frequency interference and electromyography artifacts above 45Hz.

[0056] S2: Based on S1, the filtered EEG signal is segmented into 4-second time windows, with a sampling point count of 4 seconds × 512Hz = 2048 points, and a clean EEG segment is output with a shape of 20 channels × 2048 points.

[0057] See Figure 3 The diagram shown illustrates the construction of the dual-branch graph according to the present invention. The time branch graph construction unit independently constructs a horizontal visibility view for each channel, and the processing procedure is as follows:

[0058] S10: Downsampling. The goal is to reduce the number of nodes and lower computational complexity. The method is piecewise averaging, which reduces the original signal... The downsampled signal is obtained by averaging every 16 points. The number of points dropped from 2048 to 128, maintaining the basic shape but significantly reducing the number of points, while also reducing the impact of noise.

[0059] S20: Construct node features. Calculate 4-dimensional local features for each sampling point:

[0060]

[0061] in, For signal value, It is a local mean. For local variance, This represents the local slope. The calculation method uses points... Centered on, take the front and back sides Points:

[0062]

[0063]

[0064]

[0065] Output node feature matrix The shape is 128 rows × 4 columns.

[0066] S30: Based on S20, construct an adjacency matrix according to the horizontal visibility rule. The visibility condition is: point and points Visible if and only if for all intermediate points , The edge weights are calculated using a radian-weighted function.

[0067]

[0068] Physically, it refers to the degree of inclination of the connection lines, reflecting the severity of signal jumps. Output adjacency matrix. The matrix is ​​128×128 in shape and sparse, with only visible point pairs having non-zero weights.

[0069] S40: Output timing graph. Timing graph for a single channel. ,in The node feature matrix is ​​128×4. The adjacency matrix is ​​128×128. Repeat the above steps for all 20 channels to output 20 time maps. .

[0070] The spatial branching map construction unit constructs brain region connectivity maps based on multi-scale visual features and cross-frequency coupling. The processing procedure is as follows:

[0071] S100: Construct node features. Extract multi-scale visual features for each channel, constructing a Natural Visualization (NVG) and a Horizontal Visualization (HVG) respectively. Extract four network statistical features from each of the two visuals: weighted degree, clustering coefficient, average path length, and network entropy, forming an 8-dimensional feature vector. After processing 20 channels, the node feature matrix is ​​obtained. The shape is 20 rows (channels) × 8 columns (features).

[0072] S110: Construct edge weights. For each pair of channels... Calculate the connection strength, including the dispersion index. And calculation of cross-channel phase amplitude coupling .

[0073] Detailed steps for calculating the dispersion index:

[0074] Step 1 performs NCDF mapping, which normalizes the channel signal and maps it to the (0,1) interval using the normal cumulative distribution function;

[0075] Step 2 involves linear quantization ( (Number of categories), mapping the (0,1) interval to discrete categories. ;

[0076] Step 3: Construct embedding vectors A pattern consisting of three consecutive categories, such as ;

[0077] Step 4: Statistical analysis of pattern occurrence frequency. Possible patterns; Step 5: Calculate the dispersion index This represents the proportion of the same pattern occurring at the same time.

[0078] Detailed steps for cross-channel PAC calculation:

[0079] Step 1: Extract Channels low-frequency phase, The phase is obtained by Hilbert transform after bandpass filtering at 4-8 Hz (θ band). ;

[0080] Step 2: Extract Channels high frequency amplitude, The amplitude is obtained by performing a Hilbert transform after bandpass filtering in the 30-45Hz (γ band) frequency range. ;

[0081] Step 3: Calculate the coupling strength.

[0082]

[0083] Physical meaning: channel Does the slow rhythm (θ) modulate the channel? The rapid rhythm (γ) of the epileptic seizure is abnormally enhanced during a seizure.

[0084] The weighted fusion formula is:

[0085]

[0086] in, (like , ),or , These are learnable parameters. The adjacency matrix is ​​obtained after calculating for all 20×20 channel pairs. The shape is 20 rows × 20 columns, with a diagonal of 0 (no elements connect to themselves), and the matrix is ​​symmetric. Final output spatial diagram .

[0087] The processing procedure of the bi-branch graph neural network feature learner is as follows:

[0088] S200: Detailed workflow of Temporal Branching Graph Convolutional Network: Input 20 temporal graphs Each graph contains (128×4) and (128×128). Perform the following GNN processing on each channel, with all 20 channels sharing the same set of GNN parameters.

[0089] Figure 1 shows convolutional layer 1 (GCN Layer 1):

[0090]

[0091] Dimensional change: (128×4) → (128×32). Detailed process includes: Perform feature transformation. Neighbor feature aggregation is performed, ReLU is used for non-linear activation, and Dropout (0.3) is applied to prevent overfitting. Physically, this aggregates information from neighboring visible points for each sampling point.

[0092] Figure 2 shows convolutional layer 2 (GCN Layer 2):

[0093]

[0094] Dimensional change: (128×32) → (128×64). Physically, this means expanding the receptive field, allowing the viewer to see points within two jumps.

[0095] Figure 3 shows the GCN layer:

[0096]

[0097] Dimensional change: (128×64) → (128×128). Output The shape is (128×128), and each sampling point has a 128-dimensional high-level representation.

[0098] S210: Graph Readout: The goal is to aggregate the features of 128 nodes into a single vector. The method involves concatenating average pooling and max pooling.

[0099]

[0100]

[0101]

[0102] Output The shape is 256-dimensional and is a channel. The time representation. The above process is performed on all 20 channels, resulting in 20 256-dimensional vectors.

[0103] S220: Multichannel attention aggregation: The goal is to merge 20 vectors into one and learn the importance of channels.

[0104] Step 1: Calculate the importance score for each channel:

[0105]

[0106] The MLP structure is Linear(256→64) → ReLU → Linear(64→1).

[0107] Step 2: Perform Softmax normalization to convert the weights.

[0108]

[0109] all The sum is 1. Step 3: Weighted summation:

[0110]

[0111] Output The shape is 256-dimensional. Interpretability is reflected in... This indicates the importance of each channel, and visualization shows which channel contributes the most.

[0112] S230: Detailed flowchart of spatial branch graph convolutional network: Input 1 spatial graph ,in It is 20×8 (node ​​features). It is a 20×20 (adjacency matrix).

[0113] Figure 1 shows convolutional layer 1 (GCN Layer 1):

[0114]

[0115] Dimensional change: (20×8) → (20×32). Physically, each brain region aggregates information from other brain regions connected to it, and the connection strength (edge ​​weight) determines the weight of the aggregation.

[0116] Figure 2 shows the dimensionality change of GCN Layer 2 (20×32) to (20×64). The physical meaning is to expand the receptive field and capture two-hop relationships, that is, regions that are not directly connected but are associated through intermediate brain regions.

[0117] Figure 3 shows the GCN layer 3: dimensionality change from (20×64) to (20×128). The output $H^s$ has a shape of (20×128), with each brain region having a 128-dimensional high-level representation.

[0118] S240: Attention Graph Pooling: The goal is to aggregate the features of 20 nodes into a single vector, while simultaneously learning which brain regions are more important.

[0119] Step 1: Calculate the importance score for each brain region ;

[0120] Step 2:

[0121] Perform Softmax normalization ;

[0122] Step 3: Weighted summation Step 4: Concatenate to obtain the final output. The shape is 256-dimensional. Interpretability is reflected in the visualization of attention weights as a heatmap of brain region importance; brain regions with high attention weights may be the origin of epilepsy.

[0123] See Figure 4 The diagram shows a schematic of feature fusion and classification according to the present invention. The feature fusion unit uses an attention mechanism to fuse bi-branch features, and the processing procedure is as follows:

[0124] Detailed process of attention feature fusion: Input (256 dimensions) represents the time branch output. (256 dimensions) is the spatial branch output.

[0125] S300: Calculate branch importance:

[0126]

[0127]

[0128] The MLP structure is Linear(256→64) → ReLU → Linear(64→1).

[0129] S310: Perform Softmax normalization:

[0130]

[0131] Interpretability is reflected in the fact that these two weights represent the relative contributions of time to space.

[0132] S320: Weighted Fusion

[0133]

[0134] The shape has 256 dimensions.

[0135] S330: Splicing:

[0136]

[0137] The shape has 256 + 256 + 256 = 768 dimensions.

[0138] S340: Dimensional reduction and compression

[0139]

[0140] The MLP structure is Linear(768 → 256) → ReLU → Dropout(0.3) → Linear(256 → 128), with a shape of 128 dimensions.

[0141] The classifier classifies the fused features, and the processing procedure is as follows:

[0142] S350: Input the 128-dimensional fused features into the fully connected layer 1, Linear(128 → 64) followed by ReLU activation and Dropout(0.5), and output 64 dimensions.

[0143] S360: Based on S350, it outputs a 2D dimension through a fully connected layer 2, Linear(64 → 2). .

[0144] Softmax:

[0145]

[0146] in This represents the probability of a non-seizure event. This represents the probability of an attack.

[0147] Classification decision: If If the result is positive, the prediction is an epileptic seizure; otherwise, the prediction is a non-epileptic seizure.

[0148] The output includes: seizure probability Temporal branch attention weight (channel importance), spatial branch attention weight (brain region importance), and temporal / spatial fusion weight (which branch is more important).

[0149] The model training uses the cross-entropy loss function:

[0150]

[0151] The Adam optimizer was used with an initial learning rate of 0.001, a batch size of 32, and training for 100 epochs. An early stopping strategy was employed to prevent overfitting.

[0152] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A seizure detection system based on a dual-branch spatiotemporal graph neural network, characterized in that, It includes an EEG signal acquisition unit, a signal preprocessor, a bibranch graph builder, a bibranch graph neural network feature learner, a feature fusion unit, and a classifier. The EEG signal acquisition device is used to acquire multi-channel EEG signals, i.e., raw EEG data; The signal preprocessor is used to receive the raw EEG data from the EEG signal acquisition device and perform data preprocessing, including bandpass filtering and fixed time window segmentation of the multi-channel EEG signal; wherein, the raw EEG data is bandpass filtered from 0.5Hz to 45Hz and the processed EEG signal is segmented into segments. The dual-branch graph builder is used to construct a dual-branch graph structure from the signal preprocessed by the signal preprocessor. It includes a temporal branch graph construction unit and a spatial branch graph construction unit. The temporal branch graph construction unit independently performs downsampling and horizontal visualization construction for each channel to generate multiple temporal graphs. The spatial branch graph construction unit constructs node features based on the statistical features of the multi-scale visualization network and constructs edge weights based on the dispersion index and cross-channel phase amplitude coupling to generate a spatial graph. The bi-branch graph neural network feature learner is connected to the bi-branch graph builder, including a temporal branch graph convolutional network and a spatial branch graph convolutional network, which are used to automatically learn feature representations from the temporal graph and the spatial graph, respectively, and perform feature aggregation through an attention mechanism; The feature fusion unit is used to perform attention-weighted fusion of features from the temporal and spatial branches; A classifier is used to classify the fused features to obtain the epileptic seizure detection results; The process of building a horizontal view for each channel in the time branch graph construction unit is as follows: S1: Downsampling of single-channel EEG segments was performed, and the 2048 sampling points were reduced to 128 nodes using a segmented averaging method; S2: Based on S1, calculate the 4-dimensional feature vector of each node, including the signal value, local mean, local variance and local slope of that node; S3: Based on S2, construct an adjacency matrix according to the horizontal visibility rule. If the magnitude of all intermediate nodes k between node i and node j is less than the smaller of the magnitudes of the two nodes, then there is an edge connection between the two nodes. S4: Based on S3, calculate the edge weights and use the radian weighted function to characterize the inclination of the connecting lines; The process of constructing a brain region connectivity map using spatial branching map construction units is as follows: S10: Using each EEG channel as a node, construct a natural view and a horizontal view for each channel. Extract four network statistical features from each of the two views, including weighted degree, clustering coefficient, average path length and network entropy, to form an 8-dimensional node feature vector. S20: Based on S10, calculate the dispersion index and cross-channel phase amplitude coupling strength for each pair of channels; S30: Based on S20, the edge weights are determined by a weighted fusion of the dispersion index and the phase amplitude.

2. The epileptic seizure detection system based on a dual-branch spatiotemporal graph neural network according to claim 1, characterized in that, The EEG signal acquisition device uses the international 10-20 electrode system to acquire 20 channels of EEG signals with a sampling rate of 512Hz.

3. The epileptic seizure detection system based on a dual-branch spatiotemporal graph neural network according to claim 2, characterized in that, Each EEG signal segment spans 4 seconds and has 2048 sampling points.

4. The epileptic seizure detection system based on a dual-branch spatiotemporal graph neural network according to claim 1, characterized in that, In S4, the formula for calculating the edge weight is: ； in, and Let i and j be the magnitudes of the nodes, respectively. and These are the corresponding time indices.

5. The epileptic seizure detection system based on a dual-branch spatiotemporal graph neural network according to claim 1, characterized in that, In S20, the dispersion index is calculated as follows: First, the channel signal is mapped to the (0,1) interval by the normal cumulative distribution function, then linearly quantized into 6 discrete categories, and the frequency of occurrence of the embedded vector statistical pattern is constructed. Finally, the proportion of the two channels having the same pattern at the same time is calculated. The calculation process for cross-channel phase-amplitude coupling is as follows: extract the phase of the θ band (4-8Hz) of one channel and the amplitude of the γ band (30-45Hz) of another channel, and calculate the coupling strength. ； in, For the low-frequency phase of channel i, Let be the high-frequency amplitude of channel j.

6. The epileptic seizure detection system based on a dual-branch spatiotemporal graph neural network according to claim 1, characterized in that, The temporal branch graph convolutional network adopts a 3-layer graph convolutional structure with hidden dimensions of 32, 64 and 128 respectively. Each graph convolutional layer is followed by a ReLU activation function and a Dropout layer. After graph convolution, average pooling and max pooling are used in parallel to obtain graph-level features. Then, the features of 20 channels are weighted and aggregated into a 256-dimensional temporal feature vector through a multi-channel attention mechanism.

7. The epileptic seizure detection system based on a dual-branch spatiotemporal graph neural network according to claim 1, characterized in that, The spatial branch graph convolutional network adopts a 3-layer graph convolutional structure with hidden dimensions of 32, 64 and 128 respectively. After graph convolution, attention pooling is used to aggregate the features of 20 brain region nodes in a weighted manner. The attention weights represent the importance of each brain region and output a 256-dimensional spatial feature vector.

8. The epileptic seizure detection system based on a dual-branch spatiotemporal graph neural network according to claim 1, characterized in that, The feature fusion processor uses an attention mechanism to fuse features from two branches: First, the importance scores of the temporal and spatial branches are calculated, and the fusion weights are obtained by Softmax normalization. Then, the weighted features are concatenated with the original bi-branch features and compressed into 128-dimensional fusion features by multilayer perceptron dimensionality reduction.

Images