Intracranial EEG epileptic seizure warning method based on multi-path graph-brain network modeling

Abstract

The present invention discloses a method for early warning of intracranial electroencephalographic epileptic seizures based on multi-path graph brain network modeling, which belongs to the technical field of electroencephalogram analysis. The method includes proposing a multi-path graph brain network modeling framework for epileptic seizure prediction; constructing a multi-path brain map based on three cross-frequency coupling (CFC) relationships in neuroscience, wherein the frequency bands in each channel of the electroencephalogram signal form the nodes of the map, and the three CFC relationships form the edges of the map; designing a multi-path brain map learning network with a joint fusion module to learn the patterns in the multi-path brain map and form a final representation. The present invention can analyze iEEG signals, use the designed multi-path brain map learning network to mine the patterns therein and form a unified representation, thereby improving the accuracy of epileptic seizure prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram (EEG) analysis, and in particular relates to an intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling. Background Art

[0002] Epilepsy is a chronic neurological disorder that, according to the World Health Organization, affects approximately 50 million people worldwide as of February 2024. Epileptic seizures can impair consciousness and motor function, potentially leading to life-threatening consequences. Predicting epileptic seizures in advance allows patients and healthcare providers to take preventive measures, effectively reducing the damage caused by epileptic seizures and improving patients' quality of life. Electroencephalogram (EEG), as an important tool for measuring brain activity, is widely used in epilepsy research and clinical practice. The fundamental goal of epileptic seizure prediction is to use EEG signals to distinguish the preictal phase (the brief period before an epileptic seizure) from the interictal phase (the normal period between two epileptic seizures), which can be considered a binary classification problem. Compared to scalp EEG (sEEG), intracranial EEG (iEEG) can obtain purer and more accurate signals directly from deep within the brain tissue. Therefore, iEEG signals have become a more reliable foundation for epilepsy research.

[0003] In recent years, as epilepsy research has deepened, researchers have realized that epilepsy is a network disorder, and its predictive performance depends on effectively capturing the non-Euclidean structure of brain network connectivity. Consequently, the importance of brain network modeling for epilepsy prediction has become increasingly prominent. To this end, researchers have proposed various graph-based methods that represent brain networks by constructing brain graphs. These methods are typically constructed based on the relationships between electrode channels, where channels are considered nodes and the relationships between channels are considered edges. While these graph-based methods can capture brain network connectivity and improve epileptic seizure prediction performance, their modeling approaches remain relatively simple and fail to fully account for the complexity of brain networks, leaving much room for exploration. To better utilize iEEG signals to construct brain networks, two key points must be considered: First, brain networks are complex networks with multiple relationships. These diverse relationships reflect different aspects of the network, forming a heterogeneous and complementary structure. In contrast, existing graph learning methods primarily focus on modeling brain networks based on a single relationship, neglecting other relationships. Secondly, as mentioned above, iEEG signals provide more accurate and complete frequency information. However, existing methods process iEEG signals in a relatively crude way. Instead of extracting signals from individual frequency bands, they encode all frequency information together, obscuring the rich frequency information and correlations in iEEG signals. Therefore, it is necessary to develop a more effective method to address this issue and improve epileptic seizure prediction performance. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides an intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling, which is used for epileptic seizure prediction. It specifically solves the problems of modeling multiple relationships in brain networks and utilizing iEEG frequency information, designs a multi-path brain map learning network, and mines patterns in multi-path brain maps by combining a graph encoder and a joint fusion module.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for early warning of intracranial EEG epileptic seizures based on multi-path graph-brain network modeling, comprising:

[0007] Step 1: Obtain the original iEEG data and perform preprocessing and data enhancement to obtain the final time-frequency map;

[0008] Step 2: Extract amplitude and phase information from the final time-frequency diagram, divide it by frequency band and project it into amplitude and phase features;

[0009] Step 3: semantically encode the amplitude and phase features to obtain multi-path graph node features, construct multi-path graph edge features based on the similarity between the amplitude and phase features, and construct a multi-path brain map based on the multi-path graph node features and the multi-path graph edge features, wherein the multi-path brain map includes an amplitude-amplitude coupling subgraph, a phase-phase coupling subgraph, and a phase-amplitude coupling subgraph;

[0010] Step 4: Use a graph convolutional network to learn the relationship patterns within and between sub-graphs of the multi-path brain map, and perform feature fusion through a cross-attention mechanism to obtain the final representation, which is sent to the classification head for binary classification of pre-ictal and interictal periods.

[0011] On the other hand, the present invention provides an intracranial EEG epileptic seizure warning device based on multi-channel brain network modeling, comprising: a data acquisition and preparation module, a multi-channel brain map preparation module, a multi-channel brain map construction module and a multi-channel brain map learning module, wherein:

[0012] The data acquisition and preparation module is used to acquire raw iEEG data and perform preprocessing and data enhancement to obtain the final time-frequency map;

[0013] The multi-channel brain map preparation module is used to extract amplitude and phase information from the final time-frequency map, divide it into frequency bands, and project it into amplitude and phase features;

[0014] The multi-path brain map construction module is used to perform semantic encoding on the amplitude and phase features to obtain multi-path graph node features, construct multi-path graph edge features based on the similarity between the amplitude and phase features, and construct a multi-path brain map based on the multi-path graph node features and the multi-path graph edge features, wherein the multi-path brain map includes an amplitude-amplitude coupling subgraph, a phase-phase coupling subgraph, and a phase-amplitude coupling subgraph;

[0015] The multi-path brain map learning module is used to learn the relationship patterns within and between the sub-maps of the multi-path brain map, and perform feature fusion through the cross-attention mechanism to obtain the final representation, and send the final representation to the classification head for binary classification of the pre-ictal and interictal periods to obtain the classification results.

[0016] In a third aspect, the present invention provides an electronic device comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling.

[0018] The beneficial effects of the present invention are:

[0019] This paper proposes a multi-path brain network modeling framework for epileptic seizure prediction and early warning, which can better represent various relationships in the brain network and make full use of the frequency information in iEEG; designs a multi-path brain map construction process, which can construct semantically rich multi-path brain maps based on three CFC relationships; and proposes a multi-path brain map learning network, which can mine the relationship patterns within and between sub-graphs of the multi-path brain map and integrate them into a unified final representation, thereby improving the accuracy of epileptic seizure prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of the intracranial EEG epileptic seizure warning method based on multi-graph brain network modeling of the present invention;

[0021] Figure 2 This is a network structure diagram of the intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling of the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described below with reference to the accompanying drawings and examples.

[0023] To overcome the limitations of existing methods, this paper proposes a method for epileptic seizure prediction and early warning based on multi-path brain network modeling, which includes the following steps: multi-path brain map preparation, multi-path brain map construction, and multi-path brain map learning. The multi-path brain map preparation step extracts amplitude and phase representations from iEEG signals to prepare for map construction. The multi-path brain map construction step constructs three sub-graphs based on CFC to capture different brain network connections. Finally, the multi-path brain map learning step reveals patterns within and between sub-graphs and integrates them into a unified representation.

[0024] Specifically, CFC is defined as the interaction between neural oscillations in different frequency bands. Since amplitude and phase are fundamental properties of frequency, CFC primarily includes amplitude-amplitude coupling (AAC), phase-phase coupling (PPC), and phase-amplitude coupling (PAC). In a multi-channel brain map, the frequency band of each channel is considered a node, while the three types of CFC relationships are considered three types of edges. Therefore, a multi-channel brain map consists of three sub-graphs, each representing a CFC relationship. To construct the multi-channel brain map, amplitude and phase are first extracted from the iEEG signal using the short-time Fourier transform (STFT). These are then encoded to obtain semantic representations, further forming node features. Simultaneously, different CFC relationships are calculated and threshold filtered to form edges. After completing the construction of the multi-path brain map, a multi-path brain map learning network was introduced for graph representation learning. It uses a graph convolutional network (GCN) encoder to learn the relationship patterns within and between sub-graphs, and fuses them through a joint fusion module with a cross-attention mechanism to obtain the final representation, which is then sent to the classification head for binary classification of pre-ictal and interictal periods.

[0025] like Figure 1 FIG. 1 is a flow chart of an intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling according to the present invention, comprising:

[0026] Step 1: Prepare iEEG epilepsy data as input and perform data preprocessing and data augmentation; here we use the public Kaggle dataset (https: / / kaggle.com / competitions / seizure-prediction) as an example.

[0027] For data preprocessing, the original signal data is downsampled to 256 Hz, and the data is divided into non-overlapping segments of 5 seconds in length using a sliding window. STFT is used to generate a time-frequency graph as a training sample. channels and timestamped iEEG segments , STFT will fragment No. channels, i.e. , converted to have time windows and Channel time-frequency diagram of frequency components , which is a complex matrix:

[0028]

[0029] in, is the time offset, represents the Fourier transform, is the window function, The representative frequency is The complex exponential function, where is a natural constant, is pi, is an imaginary unit, Indicates time. The time-frequency diagrams of all channels are combined into segments Corresponding time-frequency diagram , the set of time-frequency graphs of all segments is taken as the sample set.

[0030] For the data augmentation operation, the samples are augmented to balance the number of samples in the interictal and preictal periods. Specifically, two samples are randomly selected from the preictal sample set after the preprocessing operation, and the augmented sample is created by connecting the low-frequency part of the first sample with the high-frequency part of the second sample. The boundary frequency of the low and high frequencies is set to 20 Hz because brain activity mainly occurs in the frequency range below 30 Hz. This augmentation process is repeated times, among which represents the number of preictal samples after preprocessing operations, is the enhancement factor. By adjusting , the balance between interictal and preictal samples can be controlled. Therefore, for each patient or each experimental dog, It is set individually based on the original ratio of interictal and preictal samples. The samples in the enhanced sample set are the final time-frequency graph. .

[0031] Step 2: Preparation of multi-channel mind map. Figure 2 As shown, the present invention first extracts amplitude and phase information from the iEEG signal processed according to step 1, and then divides it by frequency band and projects it into amplitude and phase features to prepare for the construction of a multi-channel brain map.

[0032] Step 2-1, amplitude and phase extraction. Amplitude The final time-frequency diagram The modulus is obtained, and the phase Corresponding to the final time-frequency diagram The main angles of , can be calculated as follows:

[0033]

[0034]

[0035] in, and In addition, in order to improve the training stability and speed up the convergence, and z-score normalization was applied.

[0036] Step 2-2, divide by frequency band and project. According to the frequency band division in the EEG signal, the normalized amplitude and phase Decomposition along the frequency dimension: (1-4 Hz), (4-8 Hz), (8-12 Hz), (12-30 Hz) and (>30 Hz). This decomposition produces the amplitude of a specific frequency band and phase ,in represents the frequency band index, represents the total number of frequency bands, Indicates the The bandwidth of the frequency band. (1-4 Hz), (4-8 Hz), (8-12 Hz), (12-30 Hz) and The bandwidths of these five frequency bands (>30 Hz) are different, and the amplitude and phase information of different frequency bands are different in dimension. Therefore, the amplitude and phase information of each frequency band are projected onto a dimension of The unified feature space of :

[0037]

[0038]

[0039] in, Indicates stacking along the second dimension, means flattening the last two dimensions, i.e. , are the learnable weights for the amplitude and phase projections of each frequency band, respectively. So far, the amplitude and phase features are expressed as and , which laid the foundation for the subsequent construction of multi-path mind maps.

[0040] Step 3: Multi-path mind map construction. Figure 2 As shown, the multi-path mind map constructed by the present invention inherits the core idea of ​​the multi-path map, and its characteristic is that there are various relationships between nodes. It can be expressed as:

[0041]

[0042] in, is the initial node feature, 、 、 is the adjacency matrix corresponding to the three CFC relationships. In addition, to ensure that different edges are supported by different types of node information, the initial node features Decompose and construct fine-grained features. Specifically, the amplitude and phase features are extracted and As the input of multi-way brain map construction, the multi-way brain map construction includes node construction, edge construction and the final multi-way brain map construction for brain network modeling.

[0043] Step 3-1: Semantic feature extraction. Amplitude and phase features and These are just preliminary representations that need to be encoded by the Transformer. The Transformer uses a multi-head self-attention mechanism and a feedforward network to model the interdependencies between features. This process generates semantically relevant amplitude and phase embeddings. , which can be expressed as:

[0044]

[0045]

[0046] Among them, the flattened amplitude and phase characteristics , Here Indicates that the first two dimensions are flattened, that is, (in ). It consists of multiple Transformer layers, whose embedding dimension is the same as same.

[0047] Step 3-2, node feature construction. Based on the extracted semantic amplitude and phase embedding, node features are customized for each subgraph to correspond to three different CFC relationships. Therefore, these node features are Specifically, since AAC captures the relationship between amplitudes and PPC captures the relationship between phases, the node features of AAC and PPC subgraphs are defined as and For the PAC subgraph representing the interaction between phase and amplitude, a PAC gate is used to combine the amplitude and phase embeddings. The gate value is calculated by applying a linear transformation to the concatenation of the amplitude and phase embeddings. :

[0048]

[0049] in, Indicates splicing along the node feature dimension, is the learnable weight of the PAC gate, and the superscript T represents the transpose. Then, the node features of the PAC subgraph It is derived as a weighted sum of semantically defined amplitude and phase embeddings:

[0050]

[0051] in, is element-wise multiplication.

[0052] Step 3-3: Edge construction,The edges in the multi-path brain map are constructed based on the similarity between the amplitude and phase features.,According to the definition of the three CFC relationships, the,adjacency matrix, is initialized as a similarity matrix. Specifically, It comes from the similarity between amplitude features, Due to the similarity between phase features, Derived from the similarity between phase features and amplitude features:

[0053]

[0054]

[0055]

[0056] in, represents the cosine similarity. Threshold filtering is then applied to the adjacency matrix to eliminate spurious relationships and ensure the sparsity of the subgraph, which can be expressed as:

[0057]

[0058] in, yes Middle Rank Elements of the column, is the threshold, set to 0.5. In addition, a cross-frequency mask is applied to the adjacency matrix to ensure that edges appear between different frequency bands, and positions in the same band are set to zero. The final adjacency matrix thus represents the edges in the corresponding subgraph, where non-zero positions indicate the presence of an edge with the corresponding weight, and zero positions indicate the absence of an edge.

[0059] Step 3-4: Multi-channel brain map construction. To construct this special multi-channel brain map, three sub-graphs are defined, representing the three relationships of AAC, PPC, and PAC. Since iEEG signals from the same channel and frequency band are regarded as one node, these sub-graphs share the same node set, but different node features and edge sets. Based on the settings of nodes and edges, the multi-channel brain map The build is as follows:

[0060]

[0061] in, 、 、 is a subgraph constructed using the corresponding node features and adjacency matrix:

[0062]

[0063]

[0064] .

[0065] As the name implies, these three subgraphs correspond to three types of CFC relationships. represents the coupling between the amplitudes, represents the coupling between phases, Represents the coupling between amplitude and phase. A variety of coupling types are covered overall to summarize the diverse relationships inherent in brain networks.

[0066] Step 4: Multi-channel mind map learning. Figure 2 As shown in Figure 2, the constructed multi-path brain map clearly represents the connections between nodes in the brain network, and multi-path brain map learning aims to reveal the underlying relationship patterns through model-based learning. Multi-path brain map learning mainly consists of two parts: graph encoding and feature fusion.

[0067] Step 4-1, graph encoding. The present invention uses GCN as the graph encoder. For a given node feature and the corresponding adjacency matrix , the GCN layer is defined as follows:

[0068]

[0069] in, is the layer index, is the degree matrix of the graph, It is layer GCN layer, It is The input node features of the layer, and , It is The learnable weights of the layers, is the ReLU activation function. This paper uses three independent two-layer GCN encoders to encode sub-graph embeddings respectively. ,in is the output dimension of the entire GCN encoder:

[0070]

[0071]

[0072]

[0073] The output dimensions of these two layers are set to twice and half of the feature dimension, respectively. The subgraph embedding contains the encoded CFC relations, reflecting the three aspects of the brain network.

[0074] Step 4-2: Feature fusion. After graph encoding, the three different subgraph embeddings have effectively learned the coupling relationship contained in their respective subgraphs. To this end, this paper designs a joint fusion module based on the cross-attention mechanism to fully utilize the information from all three subgraphs and integrate these embeddings into a unified final representation. First, the joint representation is obtained by summing the subgraph embeddings. , as the initial result of the fusion process:

[0075] ,

[0076] Then, the subgraph embeddings are refined using a cross-attention mechanism. In this process, As a query (query), and subgraph embedding As a key (key) and value (value). This can be expressed mathematically as:

[0077]

[0078]

[0079] in, are learnable weights, Represents the cross attention mechanism. By combining cross attention, residual connection and layer normalization, we get the refined embedding :

[0080]

[0081]

[0082]

[0083] in, The representation layer is normalized. The final representation is recorded as , which is obtained by additively aggregating the refinement embeddings:

[0084] .

[0085] Step 4-3: Output classification. The final representation is fed into a fully connected layer classification head to distinguish between interictal and preictal segments. If the output indicates interictal status, the patient is normal and no intervention is required. If the output indicates preictal status, the patient is about to have a seizure, and an early warning is required to prepare both the patient and the healthcare provider.

[0086] Preferably, cross entropy loss is used as the loss function, and backpropagation is used to update the model parameters until convergence; hyperparameters such as the learning rate, number of training rounds, feature dimension of the model, and number of layers and heads of the Transformer encoder are adjusted to achieve optimal model performance.

[0087] Preferably, a test set is used for evaluation. The evaluation strategy employs leave-one-out-of-seizure cross-validation (LOOCV). In each round, one seizure is set aside as the test set, while the remaining data serves as the training set. This process is repeated for each seizure, and the final performance is the average of these rounds. Evaluation metrics include AUC, sensitivity (Sn), false positive rate per hour (FPR / h), and p-value. AUC reflects classification performance for interictal and preictal periods; Sn refers to the ratio of successfully predicted seizures to the total number of seizures; FPR / h refers to the number of incorrect predictions per hour; and the p-value indicates the statistical significance of the model's improvement over a random predictor. A p-value less than 0.001 indicates that the model is statistically superior to a random predictor at a confidence level of at least 99.9%.

[0088] When the model output is converted into the actual output of the epileptic seizure warning system, in order to improve stability and avoid the impact of model output fluctuations, the warning system can only be used in a continuous There are at least Only when the model determines that the patient is in the early stage of the attack will an early warning be issued. and Adjust according to actual situation.

[0089] Through the above steps, the present invention can analyze iEEG signals, construct a multi-channel brain map to express the various relationships in complex brain networks and make full use of the frequency information in iEEG, and then use the designed multi-channel brain map learning network to mine the patterns and form a unified representation, thereby improving the accuracy of epileptic seizure prediction.

[0090] On the other hand, the present invention provides an intracranial EEG epileptic seizure warning device based on multi-channel brain network modeling, which includes various modules capable of implementing various steps of the aforementioned method, specifically including a data acquisition and preparation module, a multi-channel brain map preparation module, a multi-channel brain map construction module and a multi-channel brain map learning module, wherein:

[0091] The data acquisition and preparation module is used to acquire raw iEEG data and perform preprocessing and data enhancement to obtain the final time-frequency map;

[0092] The multi-channel brain map preparation module is used to extract amplitude and phase information from the final time-frequency map, and then divide it into frequency bands and project it into amplitude and phase features;

[0093] The multi-path brain map construction module is used to perform semantic encoding on the amplitude and phase features to obtain multi-path graph node features, construct multi-path graph edge features based on the similarity between the amplitude and phase features, and construct a multi-path brain map based on the multi-path graph node features and the multi-path graph edge features, wherein the multi-path brain map includes an amplitude-amplitude coupling subgraph, a phase-phase coupling subgraph, and a phase-amplitude coupling subgraph;

[0094] The multi-path brain map learning module is used to learn the relationship patterns within and between the sub-maps of the multi-path brain map, and perform feature fusion through the cross-attention mechanism to obtain the final representation, and send the final representation to the classification head for binary classification of the pre-ictal and interictal periods to obtain the classification results.

[0095] In a third aspect, the present invention provides an electronic device comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling.

[0096] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling.

[0097] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for early warning of intracranial EEG epileptic seizures based on multi-path graph-brain network modeling, characterized in that: include: Step 1: Obtain the original iEEG data and perform preprocessing and data enhancement to obtain the final time-frequency map; Step 2: Extract amplitude and phase information from the final time-frequency diagram, divide it by frequency band and project it into amplitude and phase features; Step 3: semantically encode the amplitude and phase features to obtain multi-path graph node features, construct multi-path graph edge features based on the similarity between the amplitude and phase features, and construct a multi-path mind map based on the multi-path graph node features and the multi-path graph edge features, wherein the multi-path mind map includes an amplitude-amplitude coupling subgraph, a phase-phase coupling subgraph, and a phase-amplitude coupling subgraph; including: Step 3-1: Transformer uses a multi-head self-attention mechanism and a feedforward network to model and encode the interdependencies between features and extract semantic amplitude and phase embeddings. Step 3-2: Based on the extracted semantic amplitude and phase embeddings, customize the node features for each subgraph so that it corresponds to the amplitude-amplitude coupling subgraph, the phase-phase coupling subgraph, and the phase-amplitude coupling subgraph; Step 3-3: construct a multi-path graph edge feature based on the similarity between the amplitude and phase features, use the adjacency matrix to represent the similarity, and apply a threshold filter to the adjacency matrix. A non-zero position in the adjacency matrix indicates the existence of an edge, and its weight is the corresponding value; Step 3-4: Define iEEG signals from the same channel and frequency band as one node, and subgraphs share the same node set, but different node features and edge sets. Construct a multi-path brain map based on the node and edge settings. : in, 、 、 The amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph are constructed using the corresponding node features and adjacency matrix: Where, 、 、 Represent the node features of the amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph, respectively. 、 、 Represent the adjacency matrices of the amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph respectively; Step 4: Use a graph convolutional network to learn the relationship patterns within and between subgraphs of multiple brain maps, and perform feature fusion through a cross-attention mechanism to obtain a final representation, and send the final representation to the classification head for binary classification of pre-ictal and interictal periods; the feature fusion through the cross-attention mechanism includes: summing the embeddings of the amplitude-amplitude coupling subgraph, the phase-phase coupling subgraph, and the phase-amplitude coupling subgraph to obtain a joint representation, refining the subgraph embedding based on the joint representation to obtain a refined embedding, and summing and aggregating the refined embeddings to obtain the final representation.

2. The intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling according to claim 1 is characterized in that: The step 1 includes collecting and downsampling iEEG data, dividing the data into non-overlapping segments using a sliding window, generating a time-frequency map using STFT and performing data enhancement to obtain a final time-frequency map as a training sample.

3. The intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling according to claim 1 is characterized in that: The step 2 includes: Step 2-1: extracting amplitude and phase based on the final time-frequency graph, wherein the amplitude is obtained by the modulus of the final time-frequency graph, the phase corresponds to the main argument of the final time-frequency graph, and the amplitude and phase are normalized; Step 2-2: According to the frequency band division in the EEG signal, the normalized amplitude and phase are decomposed along the frequency dimension, and the amplitude and phase information of each frequency band are projected into a unified feature space through linear transformation to obtain amplitude and phase features.

4. The intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling according to claim 1 is characterized in that: In step 3-2, the node features of the amplitude-amplitude coupling subgraph are defined as semantic amplitude embedding, the phase-phase coupling subgraph is defined as semantic phase embedding, and the phase-amplitude coupling subgraph is defined as the weighted sum of semantic amplitude and phase embeddings.

5. The intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling according to claim 1 is characterized in that: The step 4 comprises: Step 4-1: For a given node feature and the corresponding adjacency matrix, three independent two-layer GCN encoders are used to encode the amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph embeddings respectively; Step 4-2: Use a joint fusion module based on the cross-attention mechanism to integrate the sub-graph embeddings into a unified final representation; Step 4-3: The final representation is input into a fully connected layer classification head to distinguish between interictal and preictal segments.

6. The intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling according to claim 5 is characterized in that: The step 4-2 includes: The amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph are embedded and summed to obtain a joint representation; The subgraph embeddings are refined using the cross-attention mechanism to obtain refined embeddings, which are then aggregated to obtain the final representation.

7. An intracranial EEG epilepsy attack warning device based on multi-path graph-brain network modeling, characterized in that: include: Data acquisition and preparation module, multi-channel mind map preparation module, multi-channel mind map construction module and multi-channel mind map learning module, including: The data acquisition and preparation module is used to acquire raw iEEG data and perform preprocessing and data enhancement to obtain the final time-frequency map; The multi-channel brain map preparation module is used to extract amplitude and phase information from the final time-frequency map, divide it into frequency bands, and project it into amplitude and phase features; The multi-path mind map construction module is used to semantically encode amplitude and phase features to obtain multi-path graph node features, construct multi-path graph edge features based on the similarity between the amplitude and phase features, and construct a multi-path mind map based on the multi-path graph node features and the multi-path graph edge features, wherein the multi-path mind map includes an amplitude-amplitude coupling subgraph, a phase-phase coupling subgraph, and a phase-amplitude coupling subgraph; and includes: Transformer uses a multi-head self-attention mechanism and a feedforward network to model and encode the interdependencies between features and extract semantic amplitude and phase embeddings; Based on the extracted semantic amplitude and phase embeddings, node features are customized for each subgraph to make it correspond to the amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph; Based on the similarity between amplitude and phase features, a multi-path graph edge feature is constructed. The similarity is represented by an adjacency matrix. A threshold filter is applied to the adjacency matrix. A non-zero position in the adjacency matrix indicates the existence of an edge, and its weight is the corresponding value. iEEG signals from the same channel and frequency band are considered as one node, and subgraphs share the same node set, but with different node features and edge sets. Based on the settings of nodes and edges, a multi-path brain map is constructed. : in, 、 、 The amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph are constructed using the corresponding node features and adjacency matrix: Where, 、 、 Represent the node features of the amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph, respectively. 、 、 Represent the adjacency matrices of the amplitude-amplitude coupling subgraph, phase-phase coupling subgraph, and phase-amplitude coupling subgraph respectively; The multi-path brain map learning module is used to learn the relationship patterns within and between multi-path brain map sub-maps, and perform feature fusion through a cross-attention mechanism to obtain a final representation, and send the final representation to a classification head for binary classification of pre-ictal and interictal periods to obtain a classification result; the feature fusion through the cross-attention mechanism includes: summing the embeddings of the amplitude-amplitude coupling sub-map, the phase-phase coupling sub-map, and the phase-amplitude coupling sub-map to obtain a joint representation, refining the sub-map embedding based on the joint representation to obtain a refined embedding, and summing and aggregating the refined embeddings to obtain the final representation.

8. An electronic device, characterized in that: include: one or more processors; a memory for storing one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement the intracranial EEG epileptic seizure warning method based on multi-path graph brain network modeling as described in any one of claims 1-6.

Images