Epileptic seizure prediction method fusing multi-frequency coupling and multiple attention mechanisms

Abstract

The invention belongs to the technical field of epileptic seizure prediction, and particularly relates to an epileptic seizure prediction method fusing multi-frequency coupling and multiple attention mechanisms. The invention aims to realize high-precision prediction of epilepsy. The method comprises the following steps: preprocessing EEG (electroencephalogram) data to obtain preprocessed EEG data; the EEG data includes brain neuron electrical activity signals recorded by the electrodes. Dividing an epileptic seizure early stage and an epileptic seizure interval based on the preprocessed electroencephalogram data, and constructing a multi-frequency coupling feature; the multi-frequency coupling characteristics comprise coupling relation characteristics of phases and amplitudes among electroencephalogram signals of different frequency bands. And performing spatio-temporal feature extraction on the multi-frequency coupling features by adopting a multiple attention mechanism to obtain spatio-temporal features. The multiple attention mechanism includes at least one of mechanisms that fuse map attention, map exterior attention and sequential attention. And based on the extracted spatio-temporal features, distinguishing the early stage of attack and the interval of attack of the patient, and predicting the epileptic seizure.

Description

Technical Field

[0001] The present invention relates to the technical field of epileptic seizure prediction, and in particular to an epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms. Background Art

[0002] Epilepsy is a common neurological disease with sudden and unpredictable attacks, which seriously affects the patient's quality of life and even life safety. As an important means of recording the electrical activity of brain neurons, electroencephalogram (EEG) is widely used in the diagnosis and prediction of epilepsy because it can directly reflect the functional state of the brain. However, existing epileptic seizure prediction methods do not fully exploit the multi-frequency characteristics and spatiotemporal dynamic characteristics of EEG signals, making it difficult to capture potential key patterns in the data. They have poor adaptability to individual differences and cannot make accurate predictions. Summary of the Invention

[0003] The purpose of the present invention is to provide an epileptic seizure prediction method that integrates multi-frequency coupling and multiple attention mechanisms, aiming to achieve high-precision prediction of epilepsy.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: The present invention provides an epileptic seizure prediction method that integrates multi-frequency coupling and multiple attention mechanisms, comprising: S1: Preprocessing EEG data to obtain preprocessed EEG data. The EEG data includes signals of electrical activity of brain neurons recorded by electrodes. S2: Based on the preprocessed EEG data, the pre-ictal and interictal periods of the epileptic seizure are divided, and multi-frequency coupling features are constructed. The multi-frequency coupling features include the coupling relationship characteristics of the phase and amplitude between EEG signals in different frequency bands. S3: Using a multiple attention mechanism, spatiotemporal features are extracted from the multi-frequency coupling features to obtain spatiotemporal features. The multiple attention mechanism includes at least one of the mechanisms of integrating graph attention, external graph attention, and temporal attention, and is used to explore the spatial correlations and temporal dynamics in the multi-frequency coupling features. S4: Based on the extracted spatiotemporal features, the pre-ictal and interictal periods of the patient are distinguished to predict epileptic seizures.

[0005] The epileptic seizure prediction method that integrates multi-frequency coupling and multiple attention mechanisms provided in the embodiment of the present application effectively purifies the EEG signal through preprocessing, laying a reliable data foundation for subsequent analysis. When constructing the multi-frequency coupling feature, the phase and amplitude coupling relationship between EEG signals in different frequency bands is fully considered, which can more comprehensively reflect the state of brain activity than the method that only focuses on a single frequency band. The adopted multiple attention mechanism realizes the deep fusion of the spatial information and temporal information of the EEG signal, and improves the feature discrimination ability. By combining the multi-frequency coupling features and the spatiotemporal features for classification, the pre-ictal and interictal periods can be distinguished more accurately, effectively improving the accuracy and sensitivity of epileptic seizure prediction. At the same time, the model is more adaptable to the individual differences of different patients, reduces the prediction bias caused by the differences in individual EEG signal characteristics, and shows good generalization performance in cross-patient experiments.

[0006] In some embodiments, preprocessing includes: removing data corresponding to electrodes that have not collected valid EEG signals, bandpass filtering, notch filtering, removing ICA artifact components, and normalizing and standardizing brain neuron electrical activity signals.

[0007] In some embodiments, dividing the preictal and interictal periods includes constructing an epileptic seizure prediction period and an epileptic seizure occurrence period. The epileptic seizure prediction period is used to predict the time window of the seizure, and the epileptic seizure occurrence period indicates the time interval when the seizure actually occurs. EEG data before the epileptic seizure occurrence period and within the epileptic seizure prediction period are classified as the preictal period, and EEG data not in the preictal period or the epileptic seizure occurrence period are classified as the interictal period.

[0008] In some embodiments, constructing a multi-frequency coupling feature includes: using a bandpass filter to divide EEG data into multiple frequency bands. The multiple frequency bands include theta band, alpha band, beta band, gamma 1 band, and gamma 2 band. Phase-amplitude coupling data between any two of the multiple frequency bands is calculated to form a phase-amplitude coupling connection matrix. The phase-amplitude coupling connection matrix represents the coupling strength between the phase and amplitude of EEG signals in different frequency bands.

[0009] In some embodiments, calculating phase-amplitude coupling data between pairs of multiple frequency bands to form a phase-amplitude coupling connection matrix includes: obtaining phase and amplitude information of the EEG signal in each frequency band through a Hilbert transform. The phase information indicates the phase angle characteristics of the EEG signal oscillation, and the amplitude information indicates the amplitude characteristics of the EEG signal oscillation. Based on the phase and amplitude information, the coupling strength of the phase and amplitude between different frequency bands is calculated to generate the phase-amplitude coupling connection matrix.

[0010] In some embodiments, a multi-attention mechanism is used to extract spatiotemporal features from multi-frequency coupling features to obtain spatiotemporal features. This includes inputting the phase-amplitude coupling connection matrix as graph node features into a graph attention network. Graph node features are vectors representing the EEG signal characteristics of each electrode channel. The graph attention network calculates attention weights between nodes, learns the graph structure information within the connection matrix, and outputs spatial features.

[0011] In some embodiments, the attention weights between nodes are calculated using a graph attention network, including: performing a linear transformation on node features, calculating the degree of influence between nodes using a feedforward neural network, normalizing the influence using a softmax function, and obtaining the attention weights between nodes. Multiple attention weights are then fused using multiple attention heads to optimize spatial feature extraction results.

[0012] In some embodiments, a multi-attention mechanism is used to extract spatiotemporal features from multi-frequency coupling features to obtain spatiotemporal features. This also includes introducing an external graph attention structure. The external graph attention structure contains learnable external units independent of the input graph, which are used to store cross-graph shared information. The attention between the multi-frequency coupling features and the external units is calculated, and the attention is optimized through double normalization to mine hidden information between cross-frequency coupling networks, outputting optimized spatial features.

[0013] In some embodiments, a multi-attention mechanism is used to extract spatiotemporal features from multi-frequency coupling features to obtain spatiotemporal features. This also includes: inputting the spatial features output by the graph attention network and the graph external attention structure into an encoder based on the Transformer model. The multi-head attention sublayer in the encoder captures the temporal dependencies of the features to obtain temporal features. The temporal dependencies indicate the correlation between EEG features at different time points.

[0014] In some embodiments, the encoder of the Transformer model includes a fully connected feedforward network. After capturing the temporal dependencies of features through the multi-head attention sublayer in the encoder to obtain temporal features, the encoder further includes performing two linear transformations on the temporal features through the feedforward network, applying a nonlinear activation function to the two linear transformations, and outputting spatiotemporal features. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of an epileptic seizure prediction method that integrates multi-frequency coupling and multiple attention mechanisms, as provided in an embodiment of the present application; Figure 2 is a schematic diagram of an experimental result provided in an embodiment of the present application; Figure 3 is a schematic diagram of another experimental result provided in an embodiment of the present application; Figure 4 It is a schematic diagram of another experimental result provided in an embodiment of the present application. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] In the description of the invention, it should be understood that the terms "upper," "lower," "left," "right," "front," "back," "inner," "outer," and the like, indicating directions or positional relationships, are based on the directions or relative positional relationships shown in the accompanying drawings and are intended solely to facilitate the description of the invention and simplify the description. They do not indicate or imply that the devices or components referred to must have a specific direction, be constructed, or operate in a specific direction. Therefore, they should not be construed as limitations on the invention. Unless otherwise specified, the above-mentioned directions may be flexibly set in actual application, provided that the relative positional relationships shown in the accompanying drawings are met.

[0018] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the technical features being referred to. Thus, a feature specified as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0019] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connected," and "communicated" should be understood broadly. For example, they may refer to fixed connections, detachable connections, or integral connections. They may be directly connected, indirectly connected through an intermediary, or internally connected between two components. Those skilled in the art will understand the specific meanings of these terms in the present invention based on specific circumstances.

[0020] In embodiments of the present invention, the terms "comprises," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of other identical elements in the process, article, or apparatus comprising the element.

[0021] In the embodiments of the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present invention should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0022] Electroencephalogram (EEG) signals are nonlinear, multivariate, and complex signals that record the brain's electrical activity. As a commonly used clinical testing method, they are widely used in epilepsy diagnosis. However, as mentioned in the background article, traditional prediction methods suffer from limited cross-patient experimental models and low predictive performance due to the significant individual variability among epilepsy patients.

[0023] In view of this, the present invention provides an epileptic seizure prediction method that integrates multi-frequency coupling and multiple attention mechanisms. For example, Figure 1 The method includes: S1: Preprocess the EEG data to obtain preprocessed EEG data.

[0024] Among them, EEG data includes electrical activity signals of brain neurons recorded by electrodes.

[0025] As a possible implementation method, preprocessing includes: removing the data corresponding to the electrodes that did not collect valid EEG signals, bandpass filtering, notch filtering, removing ICA artifact components, and normalizing and standardizing the brain neuron electrical activity signals.

[0026] Bandpass filtering preserves EEG signals within a specific frequency range, while notch filtering removes specific power-frequency interference signals. EEG signals are susceptible to various interferences (such as myoelectricity, electrocardiogram, and power-frequency noise), resulting in a large amount of artifacts in the raw data, making it difficult to extract effective information. Removing ICA artifacts also removes non-brain-derived interference components identified through independent component analysis. Normalization scales the EEG signal amplitude to a uniform standard, while standardization adjusts the distribution characteristics of the EEG signal to conform to preset statistical characteristics.

[0027] S2: Based on the preprocessed EEG data, the pre-ictal and interictal periods of epilepsy are divided, and multi-frequency coupling features are constructed.

[0028] Multi-frequency coupling characteristics include the coupling relationship characteristics between phase and amplitude between EEG signals in different frequency bands.

[0029] In some embodiments, dividing the preictal and interictal periods includes constructing an epileptic seizure prediction period and an epileptic seizure occurrence period. The epileptic seizure prediction period is used to predict the time window of the seizure, and the epileptic seizure occurrence period indicates the time interval when the seizure actually occurs. EEG data before the epileptic seizure occurrence period and within the epileptic seizure prediction period are classified as the preictal period, and EEG data not in the preictal period or the epileptic seizure occurrence period are classified as the interictal period.

[0030] EEG signals are nonlinear, multivariate and complex time-varying signals that contain neural oscillation components in multiple frequency bands (such as θ, α, β, γ, etc.), and there are complex cross-coupling relationships between signals in different frequency bands. These coupling characteristics are closely related to the different stages of epileptic seizures.

[0031] In some embodiments, constructing a multi-frequency coupling feature includes: using a bandpass filter to divide EEG data into multiple frequency bands. The multiple frequency bands include theta band, alpha band, beta band, gamma 1 band, and gamma 2 band. Phase-amplitude coupling data between any two of the multiple frequency bands is calculated to form a phase-amplitude coupling connection matrix. The phase-amplitude coupling connection matrix represents the coupling strength between the phase and amplitude of EEG signals in different frequency bands.

[0032] In some embodiments, calculating phase-amplitude coupling data between pairs of multiple frequency bands to form a phase-amplitude coupling connection matrix includes: obtaining phase and amplitude information of the EEG signal in each frequency band through a Hilbert transform. The phase information indicates the phase angle characteristics of the EEG signal oscillation, and the amplitude information indicates the amplitude characteristics of the EEG signal oscillation. Based on the phase and amplitude information, the coupling strength of the phase and amplitude between different frequency bands is calculated to generate the phase-amplitude coupling connection matrix.

[0033] More specifically, the Hilbert transform is used to estimate 、 、 、 and The phase corresponding to the time series of each frequency band in several frequency bands and amplitude Information, calculate phase and amplitude The joint probability distribution of ,and and The marginal probability distribution of and , then the phase amplitude coupling is: Then the single-layer phase amplitude coupling connection matrix is ​​defined as: in, Indicates frequency band and frequency band Phase and amplitude coupling between and All taken from 、 、 、 and frequency band), and By brain area and They are calculated from the EEG signal data, which provide phase and amplitude information respectively.

[0034] S3: Use multiple attention mechanisms to extract spatiotemporal features from multi-frequency coupling features to obtain spatiotemporal features.

[0035] The multiple attention mechanisms include at least one of the fusion graph attention, external graph attention, and temporal attention mechanisms, which are used to mine the spatial correlation and temporal dynamics in multi-frequency coupling features.

[0036] In some embodiments, a multi-attention mechanism is used to extract spatiotemporal features from multi-frequency coupling features to obtain spatiotemporal features. This includes inputting the phase-amplitude coupling connection matrix as graph node features into a graph attention network. Graph node features are vectors representing the EEG signal characteristics of each electrode channel. The graph attention network calculates attention weights between nodes, learns the graph structure information within the connection matrix, and outputs spatial features.

[0037] More specifically, the calculation method of the graph attention network is: The graph node features ( ) and time The correlation matrix of is set as the initial input information of the graph attention network. Among them, The symbol representing the number of electrode channels, It represents the characteristic dimension of each node.

[0038] In some embodiments, the attention weights between nodes are calculated using a graph attention network, including: performing a linear transformation on node features, calculating the degree of influence between nodes using a feedforward neural network, normalizing the influence using a softmax function, and obtaining the attention weights between nodes. Multiple attention weights are then fused using multiple attention heads to optimize spatial feature extraction results.

[0039] For an input feature vector and For a pair of nodes, the nodes For Node Impact Expressed as: Among them, the weight matrix is the weight matrix, The calculation is limited to nodes ,in Representative Node Neighborhood, attention mechanism (based on Here, it is implemented in the form of a feedforward neural network.

[0040] Apply the softmax function to normalize the weights of all adjacent nodes: Attention coefficient Representation node and nodes The attention between is a node Nodes in the first-order neighbor node set of . Further, by introducing the LeakyReLU activation function and expanding the corresponding equation, the attention coefficient calculated by the attention mechanism is: in Represents a join operation.

[0041] In order to improve the accuracy of feature learning, a multiple attention mechanism is adopted here: in, Represents the attention head, Represents the sigmoid activation function.

[0042] In some embodiments, a multi-attention mechanism is used to extract spatiotemporal features from multi-frequency coupling features to obtain spatiotemporal features. This also includes introducing an external graph attention structure. The external graph attention structure contains learnable external units independent of the input graph, which are used to store cross-graph shared information. The attention between the multi-frequency coupling features and the external units is calculated, and the attention is optimized through double normalization to mine hidden information between cross-frequency coupling networks, outputting optimized spatial features.

[0043] For example, the method for calculating the external attention of the graph is as follows: Calculate the node features X of the input graph and the external unit The attention between: in, , , . is a learnable parameter independent of the input graph and can be viewed as a The external unit of each node is a shared memory unit of all input graphs.

[0044] In the external attention mechanism of the graph, Represents the similarity between the input graph nodes and the external units. Considering the sensitivity of the attention matrix to the input feature scale, The double normalization technique is used to normalize rows and columns separately: In practice, to improve the network's capabilities, two different external units are used for keys and values: in, It is a shared unit that stores connections between edges and nodes; It is the external key and value unit of the node.

[0045] Using multi-head attention calculation: in Indicates the Size, is the total number of heads, is the linear transformation matrix, Represents a shared memory unit for different heads.

[0046] In some embodiments, a multi-attention mechanism is used to extract spatiotemporal features from multi-frequency coupling features to obtain spatiotemporal features. This also includes: inputting the spatial features output by the graph attention network and the graph external attention structure into an encoder based on the Transformer model. The multi-head attention sublayer in the encoder captures the temporal dependencies of the features to obtain temporal features. The temporal dependencies indicate the correlation between EEG features at different time points.

[0047] For example, in the operation mechanism of the multi-head attention layer, scaled dot product attention is a core component, involving the query matrix , key matrix Sum Matrix .

[0048] The multi-head attention mechanism enables the model to 、 and The three matrices perform different linear transformation mappings and focus on different representation subspaces in parallel at different locations.

[0049] in, , , .

[0050] Use the Concat(⋅) function to aggregate the attention scores from multiple heads: in, .

[0051] In addition to the multi-head attention sublayer, each encoder module is equipped with a fully connected feedforward network (FFN), which can perform independent and identical operations on each position or node in the sequence. The FFN consists of two linear transformation links separated by a nonlinear activation function: In some embodiments, the encoder of the Transformer model includes a fully connected feedforward network. After capturing the temporal dependencies of features through the multi-head attention sublayer in the encoder to obtain temporal features, the encoder further includes performing two linear transformations on the temporal features through the feedforward network, applying a nonlinear activation function to the two linear transformations, and outputting spatiotemporal features.

[0052] S4: Based on the extracted spatiotemporal features, the patient's preictal and interictal periods are distinguished and epileptic seizures are predicted.

[0053] The embodiment of the present application provides an epileptic seizure prediction method that integrates multi-frequency coupling and multiple attention mechanisms. It effectively purifies EEG signals through preprocessing, laying a reliable data foundation for subsequent analysis. Considering that the cross-frequency coupling between the oscillation components of different neural signals can help distinguish the different stages of the epilepsy course, the model uses a cross-band phase-amplitude coupling method to learn the key information of multi-frequency coupling and capture the interaction relationship between EEG signals of different frequencies. When constructing multi-frequency coupling features, the phase and amplitude coupling relationship between EEG signals of different frequency bands is fully considered, which can more comprehensively reflect the state of brain activity than the method that only focuses on a single frequency band.

[0054] The multi-attention mechanism employed achieves a deep fusion of spatial and temporal information in EEG signals, improving the discriminative power of features. To further optimize model performance, an external-graph attention mechanism is introduced to conduct in-depth learning on a variety of constructed graph structures. By adaptively assigning attention weights, key information is highlighted, enabling better exploration of potential patterns in the data and overcoming the limitations of the self-attention mechanism in multi-graph learning.

[0055] By integrating multi-frequency coupling features with spatiotemporal features for classification, the model can more accurately distinguish between preictal and interictal phases, effectively improving the accuracy and sensitivity of epileptic seizure prediction. Furthermore, the model is more adaptable to individual patient differences, reducing prediction bias caused by individual EEG signal feature differences and demonstrating good generalization performance in cross-patient experiments.

[0056] For example, Figure 2 As shown, this application uses the public scalp EEG dataset CHB-MIT to obtain EEG data of 19 patients, and uses an epileptic seizure prediction method that integrates multi-frequency coupling and multiple attention mechanisms to distinguish between the pre-ictal and interictal periods of patients.

[0057] ACC (Accuracy) can reflect the accuracy of the overall prediction. SEN (Sensitivity / Recall / True Positive Rate): It is actually the proportion of pre-ictal samples (positive samples) that are correctly predicted by the model as pre-ictal samples, which can reflect the model's ability to identify pre-ictal samples. SPE (Specificity / True Negative Rate): It is actually the proportion of interictal samples (negative samples) that are correctly predicted by the model as interictal samples, which reflects the model's ability to identify interictal samples. FPR (False Positive Rate): It is actually the proportion of interictal samples (negative samples) that are incorrectly predicted as pre-ictal samples, reflecting the false alarm situation. F1-score (F1 score): It is the harmonic mean of precision (Precision) and sensitivity (SEN), which comprehensively measures the accuracy and recall ability of the model. According to Figure 2 It can be seen that ACC, SEN, SPE, and F1-score are high, while FPR is low. Therefore, the method provided in the embodiment of the present application has a high accuracy rate in predicting results.

[0058] In order to verify the generalization performance of the method, the present invention was experimentally verified on the Siena dataset, and the results are as follows: Figure 3 As shown. Figure 3 It can be seen that ACC, SEN, SPE, and F1-score are high, while FPR is low. Therefore, the method provided in the embodiment of the present application has a high accuracy rate in predicting results.

[0059] This application also conducts ablation experiments to verify the effectiveness of the method, for example, Figure 4 As shown, after the step of losing the temporal feature or the spatial feature, the ACC, SEN, SPE, and F1-score decreased, while the FPR increased. This shows that the use of a multiple attention mechanism to extract spatiotemporal features from multi-frequency coupling features and obtain spatiotemporal features plays an important role in the method provided in this application.

[0060] In the description of this specification, specific features, structures, materials or characteristics may be combined in an appropriate manner in any one or more embodiments or examples.

[0061] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for predicting epileptic seizures by integrating multi-frequency coupling and multiple attention mechanisms, characterized in that: include: S1: Preprocessing the EEG data to obtain preprocessed EEG data; wherein the EEG data includes brain neuron electrical activity signals recorded by electrodes; S2: Based on the pre-processed EEG data, the pre-ictal and interictal periods of epilepsy are divided, and a multi-frequency coupling feature is constructed; the multi-frequency coupling feature includes the coupling relationship characteristics of the phase and amplitude between EEG signals in different frequency bands; S3: Using a multiple attention mechanism to extract spatiotemporal features from the multi-frequency coupling features to obtain spatiotemporal features; the multiple attention mechanism includes at least one of a fusion graph attention mechanism, an external graph attention mechanism, and a temporal attention mechanism, for mining spatial correlations and temporal dynamics in the multi-frequency coupling features; S4: Based on the extracted spatiotemporal features, the patient's preictal and interictal periods are distinguished and epileptic seizures are predicted.

2. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 1 is characterized in that: The preprocessing includes: removing data corresponding to electrodes that have not collected effective EEG signals, bandpass filtering, notch filtering, removing ICA artifact components, and normalizing and standardizing the brain neuron electrical activity signals.

3. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 1 is characterized in that: The division of epilepsy into preictal and interictal phases includes: Constructing an epileptic seizure prediction period and an epileptic seizure occurrence period; wherein the epileptic seizure prediction period is used to predict the time window of the seizure, and the epileptic seizure occurrence period indicates the time interval when the seizure actually occurs; The EEG data before the epileptic seizure period and within the epileptic seizure prediction period are classified as the pre-ictal period, and the EEG data not within the pre-ictal period and the epileptic seizure period are classified as the interictal period.

4. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 1 is characterized in that: The constructing of multi-frequency coupling features includes: Using a bandpass filter to divide the EEG data into multiple frequency bands; the multiple frequency bands include theta band, alpha band, beta band, gamma 1 band and gamma 2 band; The phase-amplitude coupling data between any two frequency bands in the plurality of frequency bands are calculated to form a phase-amplitude coupling connection matrix; wherein the phase-amplitude coupling connection matrix is ​​a matrix of coupling strength between phase and amplitude of EEG signals in different frequency bands.

5. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 4 is characterized in that: The calculating the phase and amplitude coupling data between any two frequency bands in the plurality of frequency bands to form a phase and amplitude coupling connection matrix includes: Obtaining phase information and amplitude information of the EEG signal in each frequency band through Hilbert transform; wherein the phase information indicates the phase angle characteristics of the EEG signal oscillation, and the amplitude information indicates the amplitude characteristics of the EEG signal oscillation; Based on the phase information and amplitude information, the coupling strength of the phase and amplitude between different frequency bands is calculated to generate a phase-amplitude coupling connection matrix.

6. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 1 is characterized in that: The multi-attention mechanism is used to extract spatiotemporal features from multi-frequency coupling features to obtain spatiotemporal features, including: The phase-amplitude coupling connection matrix is ​​input into the graph attention network as a graph node feature; wherein the graph node feature is a vector representing the EEG signal characteristics of each electrode channel; The attention weights between nodes are calculated through the graph attention network, the graph structure information in the connection matrix is ​​learned, and the spatial features are output.

7. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 6, characterized in that: The calculation of the attention weights between nodes through the graph attention network includes: Perform linear transformation on node features and calculate the influence between nodes through feedforward neural network; The softmax function is used to normalize the influence degree and obtain the attention weight between nodes; By fusing multiple sets of attention weights through multiple attention heads, the spatial feature extraction results are optimized.

8. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 6, characterized in that: The method further comprises: extracting spatiotemporal features from multi-frequency coupling features using a multiple attention mechanism to obtain spatiotemporal features; Introducing an external graph attention structure; the external graph attention structure contains a learnable external unit independent of the input graph, which is used to store cross-graph shared information; The attention between multi-frequency coupling features and external units is calculated, the attention is optimized through double normalization processing, the hidden information between cross-frequency coupling networks is mined, and the optimized spatial features are output.

9. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 8, characterized in that: The method further comprises: extracting spatiotemporal features from multi-frequency coupling features using a multiple attention mechanism to obtain spatiotemporal features; Input the spatial features output by the graph attention network and the graph external attention structure into the encoder based on the Transformer model; The temporal dependencies of features are captured by the multi-head attention sublayer in the encoder to obtain temporal features; wherein the temporal dependencies indicate the association between EEG features at different time points.

10. The epileptic seizure prediction method integrating multi-frequency coupling and multiple attention mechanisms according to claim 9, characterized in that: The encoder of the Transformer model consists of a fully connected feedforward network; After capturing the temporal dependencies of features through the multi-head attention sublayer in the encoder to obtain temporal features, the method further includes: The time series features are linearly transformed twice through the feedforward network, and a nonlinear activation function is used to process the time series features during the two linear transformations to output spatiotemporal features.

Images