Seizure prediction method based on hypergraph convolution and shap feature feedback optimization

Abstract

The application discloses a seizure prediction method based on hypergraph convolution and SHAP feature feedback optimization, comprising: collecting and preprocessing multi-channel electroencephalogram and electrocardiogram combined signals, and extracting time-frequency domain features thereof; constructing a channel-level hypergraph fusing brain region priori and electrocardiogram information, obtaining a node feature matrix and an initialized hyperedge matrix; inputting the node feature and the initialized matrix into a hypergraph convolution neural network model, the hypergraph convolution neural network model aggregates information of each time window into a low-dimensional embedding vector; adopting a SHAP explainability analysis method to screen out TOP-K key markers with the largest contribution to prediction; fusing original waveform features and screened key marker features, constructing a double-branch prediction model, and outputting a seizure prediction result; the application introduces an explainable feedback optimization mechanism, significantly improves the accuracy and explainability of seizure prediction, and provides a reference basis for clinical auxiliary diagnosis and treatment decision.

Description

Technical Field

[0001] This invention relates to the field of convolutional learning technology, and in particular to a method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization. Background Technology

[0002] In the clinical management and neuroscience research of epilepsy, achieving accurate and prospective prediction of epileptic seizures is of great significance, as it helps in early intervention, reducing the risk of harm, and improving patients' quality of life. Currently, automated prediction methods based on multimodal physiological signals, especially scalp EEG, are the mainstream research direction. However, existing methods still face a series of significant shortcomings in practical applications. Mainstream deep learning models are often regarded as black boxes, lacking transparency in their decision-making processes, making it difficult for clinicians to understand and trust their warning results. Many models only utilize isolated features of EEG signals or simple pairwise connections, making it difficult to fully characterize the complex coordinated discharges and functional network reorganization phenomena among multiple brain regions before an epileptic seizure.

[0003] Furthermore, models that solely pursue predictive accuracy often overlook the physiological heterogeneity among different patients, resulting in limited generalizability of the extracted discriminant features and a significant decline in generalization performance on unseen patients or seizure types. These shortcomings collectively limit the reliable deployment and application of predictive systems in real-world clinical scenarios. Summary of the Invention

[0004] This application provides a method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization, which enables more refined and forward-looking capture and quantification of the evolution of brain network states, and predicts the timing of epileptic seizures.

[0005] This application provides a method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization, including:

[0006] S1 simultaneously acquires 19 standard scalp EEG signals and 1 ECG signal to form 20 channels of combined signal data, and adds the epileptic seizure start and end times marked by clinical experts to the combined data; S2 performs a 0.5-70Hz bandpass filter on the signal acquired by S1 to retain the effective physiological frequency band, and performs a power frequency notch filter to remove power supply interference; the signal is divided into standardized time segment samples by a 1-second window with no overlap. S3: For the scalp EEG signal channel of each time window, the power spectral density is estimated using the Welch method, and the power spectral density and differential entropy are calculated in five frequency bands: δ, θ, α, β, and γ. For the ECG signal channel, five time-domain statistics are calculated: standard deviation, peak-to-peak value, root mean square, skewness, and kurtosis. Five dedicated frequency bands are also calculated: heart rate baseband (0.5-5Hz), QRS main energy band (5-15Hz), QRS rising / falling edge band (15-25Hz), QRS high-frequency component band (25-40Hz), and EMG / high-frequency artifact band (40-70Hz). S4 uses 20 channels as nodes, integrates prior hyperedges based on anatomical brain regions and K-nearest neighbor data-driven hyperedges based on channel feature similarity, constructs a channel-level hypergraph that can characterize the high-order coupling relationship of multiple channels, and calculates its normalized correlation matrix. S5. Design a hypergraph convolutional neural network with channel features as node attributes and hypergraph association matrix as structure. Through graph convolution, channel attention pooling and projection, the information of each window is aggregated into a low-dimensional embedding vector. S6. Using the SHAP interpretability analysis method, the contribution of each dimension in the window-level embedding vector to the prediction result is calculated, and the top-K dimensions with the highest global importance are selected as key markers. S7 constructs a dual-branch fusion model. The main branch learns deep features from the waveform processed by S2, while the bypass branch receives Top-K key markers. The features of the two are fused and then classified to output the discrimination probability of the pre-seizure / interictal phase of epilepsy.

[0007] Preferably, the data forming the 20-channel combined signal specifically includes: the 20-channel combined signal data includes 19 channels of scalp EEG signals and 1 channel of synchronous ECG signal arranged based on the international 10-20 system; the 19 scalp EEG signal channels are specifically Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, T6, Fz, Cz, and Pz.

[0008] Preferably, the standardized time segment samples are obtained by removing interference from the joint signal through zero-phase finite impulse response bandpass filtering and multi-order power frequency notch filtering, and by segmenting the filtered signal using a fixed-length, non-overlapping sliding window; wherein the window length of the sliding window is 1 second, the step size is 1 second, and the sample data matrix shape of each time window is 20 channels multiplied by 500 sampling points.

[0009] Preferably, the method further includes: for each EEG signal channel in a time window, calculating the band power spectral density and band differential entropy in multiple preset frequency bands to obtain a multidimensional feature vector for each EEG channel; for the ECG signal channel, calculating multiple time-domain statistics and calculating the band power in multiple preset physiological frequency bands to obtain a multidimensional feature vector for the ECG channel; wherein the features extracted from all channels together constitute a feature matrix, the rows of which correspond to the channels, and the columns correspond to the feature dimensions of each channel.

[0010] Preferably, constructing a channel-level hypergraph capable of characterizing high-order coupling relationships across multiple channels specifically includes: defining multiple signal channels as nodes of the hypergraph; constructing a first type of hyperedge based on prior knowledge of anatomical brain regions, wherein nodes belonging to the same brain region or functional group are aggregated into a single hyperedge; constructing a second type of data-driven hyperedge based on channel feature similarity, wherein for each node, it is aggregated into a hyperedge with a predetermined number of other nodes whose features are most similar; constructing a node-hyperedge correlation matrix based on the hyperedge set composed of the first and second types of hyperedges, and calculating a normalized hypergraph correlation matrix based on this correlation matrix, which serves as the input to the hypergraph convolutional neural network.

[0011] Preferably, the step of aggregating the information of each window into a low-dimensional embedding vector includes: aggregating the information of each window into a low-dimensional embedding vector, including: inputting the feature tensor of the window into a hypergraph convolutional neural network; the hypergraph convolutional neural network propagating and aggregating information between channel nodes based on the normalized hypergraph correlation matrix through at least one layer of hypergraph convolution operation; calculating the attention weights of each channel through a channel attention pooling layer and performing weighted summation on the channel-level representation obtained after the hypergraph convolution operation to generate a single window-level representation; mapping the window-level representation to a fixed-dimensional low-dimensional space through an embedding projection layer to form the low-dimensional embedding vector of the window.

[0012] Preferably, the key biomarkers include: using interpretability analysis methods to calculate the contribution of each dimension in the window-level embedding vector to the prediction of epileptic seizure state; evaluating and ranking the global importance of each dimension based on the contribution; and selecting a preset number of dimensions with the highest global importance as the key biomarkers.

[0013] Preferably, the final classification is performed after the two features are fused. Specifically, the waveform backbone features are fed into a 1D-CNN-GRU model and then concatenated with the bypass feedback features along the feature dimension to form joint features. The joint features are then fed into a fully connected layer for processing, and the epileptic seizure prediction result is finally output. During model training, a learning rate scheduling strategy and regularization method are used for training, and the contribution of the bypass branch to performance improvement is verified by ablation experiments.

[0014] Preferably, the output of the discriminant probability of pre-seizure / interictal phase includes: S71, constructing a series of hypergraphs based on the inter-channel coupling strength within a short-term dense sliding analysis window to form a dynamic hypergraph sequence reflecting brain network fluctuations; S72, defining and calculating the rate of change of node centrality trajectory, hyperedge stability, and global topological indices from the dynamic hypergraph sequence to quantify the indices of temporal evolution of network topology and recording them as key dynamic features of network state transition; S73, using the key evolutionary indices extracted from the dynamic hypergraph sequence as new temporal feature vectors, concatenating them with the original signal output by the 1D-CNN-GRU model, and performing the final pre-seizure prediction through a fully connected layer; S74, incorporating an interpretability analysis framework to evaluate the contribution of each feature to the prediction.

[0015] Preferably, the key features of the network state transition recorded specifically include: the key features are obtained by extracting the temporal evolution features of node centrality, the stability and emergence features of hyperedges, and the temporal change features of global network topology indicators from the hypergraph sequence; the extracted node centrality evolution features, hyperedge features and global topology change features are summarized and post-processed to form a feature vector characterizing the network state transition process.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages: Existing hypergraph-based methods for predicting epileptic seizures typically employ static, fixed hypergraph structures, failing to characterize the dynamic evolution of brain functional networks in the pre-seizure phase. This results in models struggling to capture early, subtle network instability signals before state transitions. Furthermore, current methods lack sufficient depth in interpreting prediction results. While SHAP (Signal Significance Analysis) is introduced for feature importance analysis, it fails to incorporate dynamic network evolution patterns into the interpretability framework. This leads to a lack of dynamic, temporal explanation for the crucial clinical question of how brain networks gradually enter the pre-seizure state, limiting their credibility and practical value in clinical scenarios requiring prospective decision-making.

[0017] Based on the understanding of the above problems, this paper breaks through the limitations of existing technologies that only utilize static network features and introduces a novel design concept of dynamic time-varying network analysis. The core of this technical approach lies in constructing and analyzing a hypergraph sequence reflecting the instantaneous connectivity state of the brain network in real time from a high-temporal-resolution dense sliding time window sequence, thereby changing the basis for prediction from a static network to the temporal evolution process of the network topology.

[0018] Applying the concept of dynamic time-varying network analysis to epilepsy seizure prediction presents significant technical challenges: how to stably and reliably estimate the instantaneous functional connectivity strength between multiple channels to construct a meaningful hypergraph within an extremely short sliding time window; how to effectively filter out noise and extract robust evolutionary pattern features truly relevant to state transitions from dynamic network indicator sequences extracted from high-noise physiological signals, which exhibit huge fluctuations; and how to efficiently and losslessly fuse these high-dimensional, complex dynamic temporal features, whose scale differs from the original static features, with existing high-performance static prediction models. These challenges make the actual gain of the dynamic analysis path highly uncertain.

[0019] By designing a complete technical chain including steps S1 to S7, an optimal solution to the aforementioned problem is provided. Specifically, multimodal signals are acquired simultaneously and subjected to rigorous preprocessing and feature extraction (S1-S3); channel-level hypergraphs are constructed by fusing anatomical priors and data-driven information to characterize high-order coupling relationships among multiple channels (S4); a hypergraph convolutional neural network is designed to aggregate channel information from each time window into low-dimensional embedding vectors (S5); using the SHAP interpretability analysis method, the Top-K dimensions that contribute the most to prediction are selected from the embedding vectors as interpretable key markers (S6); finally, a two-branch fusion model is constructed, where the main branch learns deep features from the original waveform, and the bypass branch receives the aforementioned key markers, fusing the features of both for final classification (S7). This scheme successfully transforms complex multi-channel correlation patterns and deep features into quantifiable and interpretable key discriminant markers, and injects them into the prediction model through a feedback mechanism. This significantly enhances the transparency and clinical interpretability of the decision-making process while improving the model's discriminative performance, achieving synergistic optimization of performance and credibility. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the epileptic seizure prediction method based on hypergraph convolution and SHAP feature feedback optimization according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the custom brain region division of the present invention; Figure 3 A graph showing the results of 100 training iterations of a hyperconvolutional model using the loss function with training loss and validation loss; Figure 4 This is a schematic diagram illustrating the relationship between SHAP values ​​and feature importance. Figure 5 This is a schematic diagram of the training steps for the dual-branch fusion model in Example 1; Figure 6 The training process for the baseline model that does not incorporate higher-order topological associations and feature feedback; Figure 7The training process of the fusion model after introducing hypergraph convolution structure and SHAP feedback mechanism; Figure 8 This is a comparison curve of various evaluation indicators during the training process of the baseline model and the fusion model. Detailed Implementation

[0021] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terminology used herein includes any and all combinations of one or more of the associated listed items.

[0023] Example 1: Figure 1 This is a flowchart illustrating the epileptic seizure prediction method based on hypergraph convolution and SHAP feature feedback optimization according to an embodiment of the present invention.

[0024] like Figure 1 As shown, the epileptic seizure prediction method based on hypergraph convolution and SHAP feature feedback optimization includes the following steps: S1, simultaneously acquires 19 standard scalp EEG and 1 ECG signal to form 20 channels of combined signal data, and adds the epileptic seizure start and end time with clinical expert annotations to the combined data; Specifically, a publicly available epilepsy dataset containing long-term monitoring records was used, such as the epilepsy EEG dataset from the American University Medical Center in Beirut (AUBMC). This dataset is designed for multi-channel synchronous acquisition. The sampling frequency was set to 500Hz, while ensuring that the data included the required 20 combined electrophysiological signals. Specifically, this included: 19 channels of scalp EEG signals: based on the international 10-20 system, with channels specifically located at Fp1 (left frontal), Fp2 (right frontal), F7 (left anterior temporal), F8 (right anterior temporal), F3 (left frontal), F4 (right frontal), T3 (left middle temporal), T4 (right middle temporal), C3 (left central), C4 (right central), T5 (left posterior temporal), T6 (right posterior temporal), P3 (left parietal), P4 (right parietal), O1 (left occipital), O2 (right occipital), Fz (frontal midline), Cz (central midline), and Pz (parietal midline); and 1 channel of synchronized electrocardiogram (EKG) signal, acquired synchronously with the EEG signals.

[0025] The dataset used must provide the start and end times of each epileptic seizure precisely annotated by clinical experts (physicians). The data records should be organized, for example, by creating an index table based on patient-record number-seizure sequence. Simultaneously, segments with missing annotations or substandard recording quality should be removed.

[0026] It should be noted that when using other datasets or self-collected data in actual application scenarios, you only need to map the above channel names to the new data source.

[0027] S2 performs a 0.5-70Hz bandpass filter on the signal acquired by S1 to retain the effective physiological frequency band, and performs a power frequency notch filter to remove power supply interference; the signal is divided into standardized time segment samples by a 1-second window with no overlap.

[0028] It should be noted that this application constructs a group identification system for each patient's EEG data for tracing and subsequent analysis. Specifically, for a single seizure event, all 1-second windows in the preic episode corresponding to that seizure are assigned a unique seizure group identifier; for a single continuous interictal interval, all 1-second windows within that continuous interval are assigned a unique interictal interval group identifier. During the data partitioning phase, using a single 1-second window as the smallest allocation unit, all samples are divided into training and test sets at an 80:20 ratio. This is uniformly applied to both the training and test sets to ensure that statistical information from the test set is not leaked into the normalization process.

[0029] Specifically, the combined signal acquired in step S1 is subjected to a 0.5-70Hz zero-phase finite impulse response (FIR) bandpass filter to suppress low-frequency baseline drift and high-frequency white noise while retaining effective physiological frequency signals. Next, the signal is subjected to multi-stage notch filtering to remove interference from the 50Hz power frequency and its harmonics (such as 100Hz and 150Hz), thereby eliminating the influence of power supply noise.

[0030] The filtered full-length record is then segmented using a fixed-length, non-overlapping sliding window. Specific parameters include: window length. Seconds, step size Seconds. The sample data matrix for each time window has the shape of (20 channels, 500 sampling points).

[0031] The samples were grouped and the dataset was divided according to complete epileptic seizure events. All samples were initially divided into training and test sets in a 4:1 ratio. Within the training set, a validation set was further divided in a 4:1 ratio. The final training set ( The ratio of the validation set to the test set is approximately 64%:16%:20%.

[0032] The normalization operation calculates a scaling factor based on the training set samples. The specific calculation method is as follows: This means taking the maximum absolute value of all samples in the training set and dividing all samples in the training set, validation set, and test set by this scale factor.

[0033] S3: For the scalp EEG signal channel of each time window, the power spectral density is estimated using the Welch method, and then the power spectral density and differential entropy are calculated in the five frequency bands of delta wave, theta wave, alpha wave, beta wave and gamma wave respectively; for the ECG signal channel, the time domain statistics and the power of the five dedicated frequency bands are calculated.

[0034] Among the time-domain statistics and the five dedicated frequency band power parameters, the time-domain statistics are used to calculate the standard deviation, peak-to-peak value, root mean square, skewness, and kurtosis of the signal, which are used to characterize the QRS wave morphology. The five dedicated frequency band power parameters are: heart rate baseband (0.5-5Hz), QRS main energy band (5-15Hz), QRS rising / falling edge band (15-25Hz), QRS high-frequency component band (25-40Hz), and electromyography / high-frequency artifact band (40-70Hz). Specifically, for each of the 19 EEG channels, the following operations are performed: the power spectral density (PSD) of the signal in that channel is estimated using the Welch method, and two calculations are performed in the following five classic EEG frequency bands: δ band (0.5-4Hz), θ band (4-8Hz), α band (8-13Hz), β band (13-30Hz), and γ band (30-45Hz).

[0035] The power spectral density (PSD) and differential entropy (DE) of the frequency band are calculated. Specifically, the signal is segmented and windowed using the Welch method, and then an FFT is performed to calculate the power spectral density. .in The discrete Fourier transform of the signal segment. Let N be the sampling frequency and N be the number of FFT points. For each frequency band, calculate its power integral value. This yields 5-dimensional features. Based on the power of each frequency band... Using formula The frequency band differential entropy is calculated to obtain 5-dimensional features. It is a very small constant used to ensure numerical stability. This represents the differential entropy value of the frequency band calculated on the b-th frequency band. It is a feature dimension extracted for each EEG channel. 'b' represents the frequency band index, specifically referring to one of the five classic EEG frequency bands. Indicates taking and The larger of the two. is a mathematical constant and a component of the formula for the definition of differential entropy, where e is the base of the natural logarithm.

[0036] Each EEG channel ultimately consists of a 10-dimensional feature vector composed of a 5-dimensional frequency band PSD and a 5-dimensional frequency band DE.

[0037] For a single EKG channel, perform the following operations: directly calculate the following five time-domain statistical indicators for the EKG signal waveform within a 1-second window to characterize the QRS wave shape: standard deviation (std), peak-to-peak value (ptp), root mean square (rms), skewness (skew), and kurtosis (kurt).

[0038] Within the following five physiological frequency bands defined for EKG, the power of each band is calculated: heart rate baseband (covering the basic heart rhythm and P wave, T wave), QRS main energy band (containing the main energy of the QRS complex), QRS rising / falling edge band (reflecting changes in the slope of the QRS wave), QRS high-frequency component band (containing high-frequency details of the QRS wave), and electromyography / high-frequency artifact band (mainly reflecting electromyography and other activities).

[0039] The EKG channel ultimately consists of a 10-dimensional feature vector composed of 5-dimensional time-domain statistics and 5-dimensional dedicated frequency band power.

[0040] After feature extraction for all channels is completed, the data from a single 1-second time window is transformed into a feature matrix of shape (20, 10), where 20 rows correspond to the 20 channels and 10 columns correspond to the 10-dimensional features of each channel. Batch samples then form a third-order tensor of shape (N, 20, 10), where N represents the number of samples.

[0041] S4 uses 20 channels as nodes, integrates prior hyperedges based on anatomical brain regions and K-nearest neighbor data-driven hyperedges based on channel feature similarity, constructs a channel-level hypergraph that can characterize the high-order coupling relationship of multiple channels, and calculates its normalized correlation matrix.

[0042] Specifically, 20 signal channels (19 EEG channels and 1 EKG channel) are defined as 20 nodes of the hypergraph.

[0043] Two types of hyperedges are constructed based on different principles to integrate prior knowledge and data-driven information. The specific divisions are as follows: Frontal hyperedge e1 contains {Fp1, Fp2, F3, F4, F7, F8, Fz}, corresponding to the prefrontal and frontal lobe regions of the brain. Central hyperedge e3 contains {C3, C4, Cz}, parietal hyperedge e4 contains {P3, P4, Pz}, temporal hyperedge e2 contains {T3, T5, T4, T6, EKG}, and occipital hyperedge e5 contains {O1, O2}. Through this division, each prior hyperedge has a clear functional correspondence in an anatomical sense, thus providing a structural prior with neuroscientific basis for subsequent hypergraph neural networks. The specific partitioning of e1, e2, e3, e4, and e5 is as follows: Figure 2As shown.

[0044] The second type is data-driven K-nearest neighbor hyperedges. To capture abnormal functional connections across anatomical brain regions that may occur during seizures, hyperedges are adaptively constructed based on the training set data: On the training set, the average value of all sample features for each channel is calculated to obtain the feature prototype of that channel. The Euclidean distance between all channel prototypes is calculated as a measure of channel similarity. For each channel (as the central node), it and the k most similar (k=4 in this application) other channels are selected to form a hyperedge. Thus, 20 data-driven hyperedges are generated for 20 channels.

[0045] Based on the aforementioned hyperedge set, the following matrix operations are performed to obtain the normalized incidence matrix used for subsequent hypergraph convolution: Construct the node-hyperedge incidence matrix H, matrix... , where N=20 is the number of nodes, R is the set of real numbers, and |E| is the total number of hyperedges (prior hyperedges + data-driven hyperedges). Matrix element H[v, e] = 1 if and only if node v belongs to hyperedge e, otherwise it is 0.

[0046] Calculate the degree of the diagonal matrix Its diagonal elements [v, v] represents the number of hyperedges to which node v belongs. This is achieved by placing this N×1 vector (whose elements are the degree of each node) on the diagonal of an N×N matrix, with the remaining positions set to 0, thus forming a diagonal matrix of node degrees. Calculate the hyper-marginal diagonal matrix. Its diagonal elements [e, e] represents the number of nodes contained in the hyperedge e. This is achieved by placing this |E|×1 vector (whose elements are the degree of each hyperedge) on the diagonal of a |E|×|E| matrix, forming a hyperedge degree diagonal matrix. Define a diagonal matrix W for the hyperedge weights, and set W as the identity matrix, meaning that all hyperedges are considered to have the same importance (weight 1).

[0047] Calculate the normalized hypergraph incidence matrix G using the following formula: ; This matrix G is the Laplacian operator approximation on the hypergraph, and it will serve as the core operator for information propagation in the subsequent Hypergraph Convolutional Neural Network (HGCN).

[0048] S5. Design a hypergraph convolutional neural network with channel features as node attributes and hypergraph association matrix as structure. Through graph convolution, channel attention pooling and projection, the information of each window is aggregated into a low-dimensional embedding vector.

[0049] Specifically, a hypergraph convolutional neural network is designed to execute sequentially. The core components include: two hypergraph convolutional layers for information propagation and feature transformation on the hypergraph structure; a channel attention pooling layer to fuse information from all channels and generate a single window-level representation; an embedding projection layer to map the window-level representation to a fixed low-dimensional space; and (attached) a classification head: used only during the model training phase to provide supervision signals; not used when generating the final embedding vector.

[0050] The input is a tensor of shape (N,20,10), representing N time window samples, each sample has 20 channels, and each channel has 10 initial features.

[0051] Each layer performs core operations. ,in, It is the normalized hypergraph correlation matrix pre-calculated in step S4, which defines the higher-order connections and information propagation rules between nodes. It is the output result obtained after the node features are input into this layer and processed by this layer (including feature transformation and nonlinear activation). This is the learnable weight matrix for this layer. The first layer takes the initial channel features as input, and the second layer takes the output of the first layer as input. After two convolutional layers, a channel-level representation containing higher-order correlation information is obtained. .

[0052] Representation for each channel Calculate an attention weight The calculation formula is: MLP is a small neural network. It uses weights. We perform a weighted summation of all channel representations to obtain the aggregated representation of the entire time window: This step compresses the information from 20 channels into a single vector z. The window-level representation z is then input into a linear layer (i.e., an embedding projection layer), projecting it into a predefined 64-dimensional low-dimensional space. After passing through the ReLU activation function, the final window-level embedding vector e (64 dimensions) is obtained.

[0053] In the training configuration, the optimizer is Adam (learning rate 1e-3, weight decay 2e-3), the loss function is labeled smooth cross-entropy, and the training strategy allows for a maximum of 100 training epochs. Figure 3 The diagram shows the results of 100 training iterations of the hyperconvolution model using the loss function to adjust the training and validation losses.

[0054] After model training is complete, the classification head is removed. All samples from the training, validation, and test sets are then input into the trained HGCN network for forward propagation, collecting the embedding vector e output by each sample after the embedding projection layer. The embedding vectors of all samples are then stacked to form a shape of (…). The window-level embedding matrix (64) (This represents the total number of samples). This matrix will serve as input for step S6 (SHAP analysis).

[0055] S6 utilizes the SHAP interpretability analysis method to calculate the contribution of each dimension in the window-level embedding vector to the prediction result, and selects the Top-K dimensions with the highest global importance as key markers.

[0056] Specifically, the window-level embedding matrix generated in step S5 is used, along with clinically labeled binary classification labels (interictal and preictal phases) corresponding to the validation set samples. On the validation set, the TreeExplainer algorithm from the SHAP library is used to interpret the prediction model. The Hypergraph Convolutional Neural Network (HGCN) trained in step S5, or its top classifier, is considered the model to be interpreted. The algorithm calculates a SHAP value (denoted as SHAP) for each embedding dimension (64 dimensions in total) of each sample in the validation set. This value quantifies the magnitude and direction of the contribution of the d-th dimension feature of the sample to the model's prediction of the pre-seizure phase (positive values ​​promote prediction, negative values ​​inhibit prediction).

[0057] For each dimension d (d=1 to 64) of the embedding vector, calculate the sample mean of the absolute values ​​of its SHAP values ​​as the global importance score for that dimension. The calculation formula is: All 64 dimensions were scored according to their global importance. Sort from high to low, where M is the number of samples (time window) in the validation set used for SHAP analysis.

[0058] Based on a preset value K (K=10 in this application), the top K embedding dimensions by importance are selected from the ranking results of the previous step. The index set of these K dimensions... These are the selected Top-K key markers. They represent the subset of core features that contribute most to distinguishing seizure states, identified from the features learned from the hypergraph convolutional network.

[0059] The index set S of Top-K key biomarkers is stored. Simultaneously, the SHAP analysis process can generate bar importance plots and bee colony plots for visualization: the bar plots visually display the global importance ranking of the Top-K dimensions (or all dimensions); the bee colony plots show the distribution of SHAP values ​​for each Top-K feature across all samples, allowing observation of the relationship between feature values ​​and contribution direction (positive / negative). To further screen the biomarkers that contribute most significantly to epilepsy prediction, this invention uses the SHAP method to perform interpretability analysis on the model output, quantifying the marginal contribution of each feature to the prediction results. For example... Figure 4The diagram showing the relationship between SHAP absolute value and feature importance is illustrated. The horizontal axis represents the SHAP value of each feature (positive values ​​indicate positive prediction of pre-ictal phase, negative values ​​indicate negative prediction). The vertical axis is arranged in descending order of feature importance, with colors from blue to red representing feature original values ​​from low to high. This invention selects the top 10 features with the highest mean SHAP absolute value as the final input markers, in the following order: EKG_skew (skewness, measuring the symmetry of the ECG signal waveform distribution), EKG_ptp (peak-to-peak value, the difference between the maximum and minimum values ​​of the signal over a period of time), T4_DE_gamma (differential entropy of the γ band in the T4 channel (right mid-temporal region), F8_DE_gamma (differential entropy of the γ band in the F8 channel (right frontal region), and EKG_P_emg (power in the 40–70Hz band). EKG_P_hf2 (power in the 25–40 Hz band, corresponding to the energy of the high-frequency components of the QRS), F7_DE_gamma (differential entropy of the γ band in the F7 channel (left frontal region), F4_DE_beta (differential entropy of the β band in the F4 channel (right midfrontal region), EKG_rms (root mean square, the effective value of the signal, reflecting the overall average power level and stability of the ECG signal), and EKG_kurt (kurtosis, a measure of the sharpness of the ECG signal distribution).

[0060] S7 constructs a dual-branch fusion model. The main branch learns deep features from the waveform processed by S2, while the bypass branch receives Top-K key markers. The features of the two are fused and then classified to output the discrimination probability of the pre-seizure / interictal phase of epilepsy.

[0061] Specifically, the core of an enhanced epileptic seizure prediction model is a two-branch architecture that receives two types of inputs: a. Waveform Backbone Branch: Receives the raw EEG and EKG waveform data after preprocessing (filtering, normalization) in step S2. Input 19-channel EEG waveform. Extracts deep time-frequency and time-series features from the EEG using a deep network composed of multi-scale one-dimensional convolutional parallel layers, depthwise separable convolutions, ordinary convolutions, residual blocks, and bidirectional gated recurrent units. Input a 1-channel EKG waveform. Extract QRS morphology and heart rate dynamics features through one-dimensional convolution, batch normalization, activation, and pooling layers to obtain the feature set. .Will and By fusing the waveforms using a cross-attention mechanism, a unified waveform backbone feature is obtained. .

[0062] b. Receive the Top-K key markers selected in step S6, i.e., a specific dimension subset of the window-level embedding vector. First, batch normalize the specific dimension subset to regularize its data distribution. Then, pass it through a multilayer perceptron (to perform projection transformation) to obtain feedback features that match the dimensions of the backbone features. .

[0063] c. Identify the main characteristics of the waveform. Bypass feedback characteristics The concatenation is performed along the feature dimension. The concatenation formula is shown below: ; Where h∈ The main feature vector is learned from the original multi-channel EEG waveform by a temporal feature extraction module composed of 1D-CNN and Bi-GRU, with dimension d. h =64, containing rich local dynamic information in the time and frequency domains; e∈ The auxiliary feature vector is obtained by applying SHAP importance filtering and linear mapping to the node embeddings output by the hypergraph convolutional network, with dimension d. e =32, encoding high-order topological association information between electrode nodes; ⊕ indicates a direct vector concatenation operation, which connects two vectors end-to-end in dimensional order to form a higher-dimensional joint representation; λ is the scaling factor of the auxiliary features, used to adjust the relative contribution weights of the two types of features in the fused vector. In this method, λ=1, f The final fused vector has a dimension of d. h +d e =96. The concatenated features are fed into a classification head consisting of fully connected layers, which ultimately outputs the binary probabilities of the pre-ictal and interictal phases of epilepsy.

[0064] The parameter settings for the dual-branch fusion model are shown in Table 1.

[0065] Table 1

[0066] For example, to intuitively compare the convergence behavior and generalization ability of the baseline model and the model fused with handcrafted features during training, this invention experimentally compares the training process of the baseline model using only the original waveform input and the model enhanced by hypergraph feature feedback. Figure 6 and Figure 7 As shown, Figure 6 The training process for the baseline model that does not incorporate higher-order topological associations and feature feedback. Figure 7 The training process of the fusion model after introducing the hypergraph convolution structure and SHAP feedback mechanism is shown. By comparing the convergence speed and stability of the two curves, the significant advantages of the new model in improving training efficiency and preventing overfitting can be clearly verified. Figure 6It can be observed that the loss curve representing the training set (blue line) decreases extremely slowly and has not fully converged even after about 60 rounds of training, remaining in a state of obvious oscillation. Meanwhile, the loss curve representing the validation set (red line), although initially decreasing, subsequently exhibits sharp fluctuations and shows an upward trend. This indicates that the baseline model is highly prone to overfitting, has poor generalization performance, and cannot stably learn the deep features of the pre-seizure phase. Figure 7 This demonstrates the complete model training process after introducing a hypergraph convolutional structure and a SHAP feedback mechanism. Figure 6 In stark contrast, the training loss of this model (blue line) rapidly decreased and stabilized within a very short number of epochs (approximately 5-10 epochs) without significant oscillations. More importantly, its validation set loss (red line) also decreased rapidly and remained stable at a very low level, maintaining a good fit with the training set loss. This demonstrates that the fusion model not only significantly improves convergence speed but also possesses stronger generalization ability, effectively avoiding overfitting and thus providing more stable and reliable discrimination results in actual predictions.

[0067] Meanwhile, to intuitively verify the actual performance and stability of the hypergraph convolution and SHAP feedback optimization strategy proposed in this invention during the model training process, this invention compares the changing trends of various key evaluation indicators of the baseline model and the fusion model in 100 rounds of iterative training through ablation experiments. Figure 8 The training performance curves for two configurations are presented from five dimensions: accuracy (ACC), AUC, F1 score, specificity (Spec), and false alarm rate (FAR). The blue curve represents the baseline model without higher-order topology and feature feedback, while the red curve represents the fusion model with hypergraph convolution and SHAP feedback. Comparing the curves clearly shows that the fusion model not only outperforms the baseline model in all metrics but also remains stable throughout the training process, indicating that SHAP-based selection of key markers helps improve model performance. The baseline model has a relatively low overall performance, with an AUC of only 75% and an F1 score of 45%. The fusion model, by introducing a hypergraph structure and SHAP feedback mechanism, steadily improves the AUC to 82%, the F1 score to 53%, and the accuracy to 80%, while significantly reducing the false alarm rate to 14% and maintaining it at a low level. This demonstrates that utilizing higher-order topological associations and interpretable feature selection can effectively reduce the false alarm rate and improve overall accuracy, correcting discrimination errors.

[0068] In summary, the performance of Baseline1D-CNN-GRU and Enhanced1D-CNN-GRU (model with handcrafted feature enhancement) on the four evaluation metrics is as follows: the improvement results for accuracy (ACC), area under the curve (AUC), F1 score, specificity, and false alarm rate (FAR) are shown in Table 2. Overall, the introduction of selected handcrafted features improved all model metrics, indicating that this enhancement structure effectively improved the model's discriminative ability.

[0069] Table 2

[0070] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By constructing a channel-level hypergraph to model multi-channel high-order coupling and utilizing SHAP to screen interpretable key biomarkers for feedback to the prediction model, the accuracy and generalization performance of epileptic seizure prediction are significantly improved. It innovatively integrates cross-modal information from EEG and ECG, enabling the model to capture pre-seizure physiological pattern changes earlier. Simultaneously, the entire process possesses good interpretability, providing a reliable basis for clinical decision-making and effectively solving the problem of traditional deep learning methods struggling to balance performance and interpretability. To confirm that the performance improvement stems from the proposed feedback optimization chain, ablation experiments were conducted for comparison. Using the same backbone network model, the performance of two configurations was compared on the same test set: the baseline model without bypass branches; and the model with hand-crafted feature enhancements, bypass branches, and Top-K key biomarkers.

[0071] Example 2: In Example 1, a static channel-level hypergraph was constructed to characterize the high-order coupling relationships between multiple channels, and SHAP was used to screen key biomarkers and feed them back to the prediction model, effectively improving the performance and interpretability of seizure prediction. However, the hypergraph constructed in this example is static and fixed, reflecting a global or long-range average brain network connectivity pattern. The physiological processes before an epileptic seizure are essentially dynamic and evolving; the brain's functional networks undergo continuous temporal reconstruction during the interictal, preictal, and ictal periods. Static hypergraphs cannot capture this time-dependent dynamic fluctuation of network topology, causing the model to potentially lose crucial temporal prediction information about how the network warms up, when it reaches a critical point, and how it reorganizes, thus becoming insensitive to earlier and weaker network instability signals in the preictal period. To capture and quantify the evolution of brain network states more precisely and proactively, it is necessary to simultaneously consider the instantaneous strength of network connections, the temporal continuity of topological structure, and the evolutionary trends of dynamic indicators for further optimization and enhancement.

[0072] In some embodiments, the discriminant probability of the pre-seizure / inter-seizure phase is output, and step S7 further includes: S71, within a short-term dense sliding analysis window, constructs a series of data-driven hypergraphs arranged in chronological order in real time based on the instantaneous coupling strength between channels, forming a dynamic hypergraph sequence that reflects the rapid fluctuations of the brain network.

[0073] Specifically, on the preprocessed continuous multi-channel signal (from step S2), a set of short, highly overlapping, dense sliding time windows is defined for dynamic analysis. The window length is typically set shorter than the final prediction window (e.g., 250-500 ms) to achieve higher temporal resolution to capture rapid dynamic changes. The step size (or overlap rate) is set small enough (e.g., 90%-95% overlap) to ensure that the generated hypergraph sequence evolves continuously and smoothly in time.

[0074] For each signal segment within a dense sliding analysis window, calculate the instantaneous functional connectivity strength between all channel pairs (190 pairs in total). The calculation method must be applicable to short-time data. For example: phase lock value, used to assess the phase synchronization of rhythmic activity of different channels in a specific frequency band (such as the θ and α bands); amplitude envelope correlation, calculating the Pearson correlation coefficient of the amplitude envelope of each channel signal within a specific frequency band, reflecting the coherence of amplitude fluctuations. The output is a symmetric 20x20 instantaneous coupling matrix, which characterizes the pairwise connectivity strength between all channels within the current short-time window.

[0075] Using 20 channels as nodes, without relying on prior dissection, and based solely on the instantaneous coupling matrix of the current window, a data-driven hyperedge is constructed for each node (channel). The construction rule (taking K-nearest neighbors as an example) is as follows: For each node i, based on the i-th row (or column) of the coupling matrix, find the K other nodes (excluding itself) with the highest connection strength to node i. Merge node i with these K most similar neighbor nodes to form a hyperedge. Repeat this process 20 times for the current dense analysis window to generate an instantaneous hypergraph containing 20 data-driven hyperedges.

[0076] The above steps are repeated for each dense sliding analysis window in chronological order. By arranging all the instantaneous hypergraphs generated in chronological order, a dynamic hypergraph sequence is obtained. This sequence fully records the high-resolution fluctuations and evolution of the brain functional network topology over time.

[0077] S72 defines and calculates the rate of change of node centrality trajectory, hyperedge stability and global topology indices from dynamic hypergraph sequences, quantifies the indices of network topology temporal evolution, and records them as key dynamic features of network state transitions.

[0078] Specifically, on each instantaneous hypergraph of the dynamic hypergraph sequence, the eigenvector centrality value reflecting the global influence of each channel node is calculated to form the centrality time series curve of each node. Statistical features such as mean, trend slope, and fluctuation variance are extracted from the curve within a specific time period as quantitative indicators describing the dynamic evolution of node influence.

[0079] The lifespan of a hyperedge is analyzed by tracking whether it appears continuously in consecutive time windows, and the ratio of stable to transient hyperedges is statistically analyzed. At the same time, the matching degree is calculated by comparing the hyperedge sets of adjacent time windows, and abnormal hyperedges that appear or disappear suddenly near key prediction moments are detected. The channel combinations contained in these hyperedges are recorded as potential pathological network pattern markers.

[0080] For each instantaneous hypergraph, the global efficiency and average clustering coefficient of the overall network are calculated. After forming the time-series curves of these indicators, the first-order difference (instantaneous rate of change) and the second-order difference (acceleration of change) of the curves are further calculated. Key features such as extreme points and zero crossings are extracted from them to quantify the speed and intensity of the overall network topology reorganization.

[0081] All the node evolution features, hyperedge statistical features, and global dynamic features obtained from the above calculations are summarized and arranged in a fixed order to form a comprehensive dynamic graph feature vector. This vector is then subjected to post-processing such as standardization and filtering, and finally used as a dynamic feature to characterize the network state transition process and input into the subsequent prediction model.

[0082] S73 extracts key evolutionary indicators from the dynamic hypergraph sequence as new temporal feature vectors, which are then fused with the original static features and input into the classifier for final prediction of pre-seizure epilepsy.

[0083] Specifically, since the dynamic feature sequence originates from dense sliding windows with high temporal resolution, and the final prediction target (pre-ictal / interval) is a 1-second window for each standard, time alignment is required first. The dynamic feature vector sequence is integrated into a time range matching each 1-second prediction window according to its corresponding timestamp. For example, for a 1-second prediction window, the mean or concatenated statistical characteristics (such as maximum value or trend) of the dynamic feature vectors calculated from all dense analysis windows within its time interval are taken to form a fixed-dimensional aggregated dynamic feature vector corresponding to that 1-second window.

[0084] The aligned aggregated dynamic feature vector is then fused with the following two types of static features from the original Example 1: Static hypergraph features, which are low-dimensional embedding vectors generated in step S5 (static hypergraph convolutional neural network) that characterize the global channel correlation pattern within a 1-second window; and deep features of the original waveform, which are deep features extracted from the corresponding 1-second original waveform by the backbone branch (1D-CNN-GRU network) in step S7.

[0085] The dynamic feature vector, static hypergraph embedding vector, and waveform depth feature vector are directly concatenated along the feature dimension to form a longer joint feature vector. A lightweight attention module is designed to learn a weight for each of the dynamic, static hypergraph, and waveform features, and then perform a weighted summation, enabling the model to adaptively focus on important information from different sources. The dynamic features are first fused with the static hypergraph features, and the result is then fused with the waveform features to form a hierarchical feature representation.

[0086] Using the fused joint feature vector as input, a classifier (such as a multilayer perceptron) is constructed or reused (modified). Continuing with the model architecture of Example 1 (step S7), it is extended to a three-branch input model: Branch 1 (unchanged) receives the original waveform and outputs waveform depth features; Branch 2 (unchanged) receives static hypergraph embedding features (or key markers filtered by SHAP). The output features of the two branches are integrated using the above fusion strategy and then input into the final classification head.

[0087] Training is performed end-to-end. The loss function and optimizer can be set as in Example 1. To ensure effective learning of dynamic features, the weights of the waveform and static hypergraph branches can be fixed initially, and only the dynamic feature branch and the fusion classification head are trained. After the loss stabilizes, all parameters can be fine-tuned jointly.

[0088] The trained ensemble model synchronously receives the corresponding raw waveform data, static hypergraph key markers (or embeddings), and aggregated dynamic feature vectors for each 1-second time window. The model performs forward propagation and ultimately outputs the probability value that the time window belongs to the pre-seizure phase. By setting a threshold or applying sliding window smoothing decision logic, the continuous probability sequence is transformed into the final pre-seizure warning signal.

[0089] S74 incorporates dynamic graph indicators into the interpretability analysis framework to assess the contribution of each dynamic feature to the prediction.

[0090] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By constructing dynamic time-varying hypergraph sequences to capture the temporal evolution of brain functional network topology and extracting key dynamic features, this method can identify the dynamic process of brain network instability before epileptic seizures earlier and with greater sensitivity, thus significantly improving the predictive power and accuracy. Furthermore, by incorporating dynamic graph indicators into the interpretability analysis framework, this method enables interpretable tracking of dynamic pathological processes, further enhancing the clinical credibility and practicality of the model.

[0091] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization, characterized in that, include: S1 simultaneously acquires 19 standard scalp EEG signals and 1 ECG signal, forming a combined signal data of 20 channels; S2 performs a 0.5-70Hz bandpass filter on the signal acquired by S1 to retain the effective physiological frequency band, and performs a power frequency notch filter to remove power supply interference; the signal is divided into standardized time segment samples by a 1-second window with no overlap. S3. For the scalp EEG signal channel of each time window, the power spectral density is estimated using the Welch method, and the power spectral density and differential entropy are calculated in five frequency bands: delta wave, theta wave, alpha wave, beta wave, and gamma wave. For the ECG signal channel, five time-domain statistics are calculated: standard deviation, peak-to-peak value, root mean square, skewness, and kurtosis. Five dedicated frequency bands are also calculated: heart rate baseband (0.5-5Hz), QRS main energy band (5-15Hz), QRS rising / falling edge band (15-25Hz), QRS high-frequency component band (25-40Hz), and EMG / high-frequency artifact band (40-70Hz). S4 uses 20 channels as nodes, integrates prior hyperedges based on anatomical brain regions and K-nearest neighbor data-driven hyperedges based on channel feature similarity, constructs a channel-level hypergraph that can characterize the high-order coupling relationship of multiple channels, and calculates its normalized correlation matrix; specifically, 19 EEG channels and 1 EKG channel are defined as the 20 nodes of the hypergraph; two types of hyperedges are constructed based on different principles, and the prior knowledge and data-driven information are integrated to classify them, clarifying the functional correspondence of each prior hyperedge in an anatomical sense; S5. Design a hypergraph convolutional neural network with channel features as node attributes and hypergraph association matrix as network structure. Through graph convolution, channel attention pooling and projection, the information of each window is aggregated into a low-dimensional embedding vector. S6. Using the SHAP interpretability analysis method, the contribution of each dimension in the window-level embedding vector to the prediction result is calculated, and the top-K dimensions with the highest global importance are selected as key markers. S7 constructs a dual-branch fusion model. The main branch learns deep features from the waveform processed by S2, while the bypass branch receives Top-K key markers. The features of the two are fused and then classified to output the discrimination probability of the pre-seizure / interictal phase of epilepsy.

2. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 1, characterized in that, The data forming the 20-channel combined signal specifically includes: the 20-channel combined signal data includes 19 scalp EEG signals and 1 synchronous ECG signal based on the international 10-20 system; the 19 scalp EEG signal channels are specifically the positions of the EEG electrodes: Fp1 left frontal region, Fp2 right frontal region, F7 left anterior temporal region, F8 right anterior temporal region, F3 left frontal region, F4 right frontal region, T3 left middle temporal region, T4 right middle temporal region, C3 left central region, C4 right central region, T5 left posterior temporal region, T6 right posterior temporal region, P3 left parietal region, P4 right parietal region, O1 left occipital region, O2 right occipital region, Fz frontal midline, Cz central midline and Pz parietal midline.

3. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 1, characterized in that, The standardized time-segment samples are obtained by removing interference from the joint signal through zero-phase finite impulse response bandpass filtering and multi-order power frequency notch filtering, and by segmenting the filtered signal using a fixed-length, non-overlapping sliding window; wherein the window length of the sliding window is 1 second, the step size is 1 second, and the sample data matrix shape of each time window is 20 channels multiplied by 500 sampling points.

4. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 1, characterized in that, The method further includes: for each EEG signal channel in a time window, calculating the band power spectral density and band differential entropy in multiple preset frequency bands to obtain a multidimensional feature vector for each EEG channel; for the ECG signal channel, calculating multiple time-domain statistics and calculating the band power in multiple preset physiological frequency bands to obtain a multidimensional feature vector for the ECG channel; wherein, the features extracted from all channels together constitute a feature matrix, the rows of which correspond to the channels, and the columns correspond to the feature dimensions of each channel.

5. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 1, characterized in that, The construction of a channel-level hypergraph capable of characterizing high-order coupling relationships across multiple channels specifically includes: defining multiple signal channels as hypergraph nodes; constructing a first type of hyperedge based on prior knowledge of anatomical brain regions, where nodes belonging to the same brain region or functional group are aggregated into a single hyperedge; constructing a second type of data-driven hyperedge based on channel feature similarity, where for each node, it is aggregated into a hyperedge with a predetermined number of other nodes whose features are most similar to its own; constructing a node-hyperedge correlation matrix based on the hyperedge set composed of the first and second types of hyperedges, and calculating a normalized hypergraph correlation matrix based on this correlation matrix, which serves as the input to the hypergraph convolutional neural network model.

6. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 1, characterized in that, The step of aggregating the information of each window into a low-dimensional embedding vector includes: aggregating the information of each window into a low-dimensional embedding vector, including: inputting the feature tensor of the window into a hypergraph convolutional neural network; the hypergraph convolutional neural network propagating and aggregating information between channel nodes based on the normalized correlation matrix through at least one layer of hypergraph convolution operation; calculating the attention weights of each channel through a channel attention pooling layer and performing weighted summation on the channel-level representation obtained after the hypergraph convolution operation to generate a single window-level representation; and mapping the window-level representation to a fixed-dimensional low-dimensional space through an embedding projection layer to form the low-dimensional embedding vector of the window.

7. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 1, characterized in that, The key biomarkers include: using interpretability analysis methods to calculate the contribution of each dimension in the window-level embedding vector to the prediction of epileptic seizure state; evaluating and ranking the global importance of each dimension based on the contribution; and selecting the top-K features with the highest global importance as the key biomarkers.

8. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 1, characterized in that, The two features are fused for final classification. Specifically, the waveform backbone features are fed into a 1D-CNN-GRU model and then concatenated with the bypass feedback features along the feature dimension to form joint features. The joint features are then fed into a fully connected layer for processing, and the epileptic seizure prediction results are finally output. During model training, a learning rate scheduling strategy and regularization method are used for training, and the contribution of the bypass branch to performance improvement is verified by ablation experiments.

9. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 1, characterized in that, The output probability of pre-seizure / interictal epilepsy includes: S71, constructing a series of hypergraphs based on inter-channel coupling strength within a short-term dense sliding analysis window to form a dynamic hypergraph sequence reflecting brain network fluctuations; S72, defining and calculating the rate of change of node centrality trajectory, hyperedge stability, and global topological indices from the dynamic hypergraph sequence to quantify the indicators of temporal evolution of network topology and record them as key dynamic features of network state transition; S73, using the key evolutionary indices extracted from the dynamic hypergraph sequence as new temporal feature vectors, concatenating them with the original signal output by the 1D-CNN-GRU model, and performing the final pre-seizure prediction through a fully connected layer; S74, incorporating an interpretability analysis framework to evaluate the contribution of each feature to the prediction.

10. The method for predicting epileptic seizures based on hypergraph convolution and SHAP feature feedback optimization according to claim 9, characterized in that, The key dynamic features recorded for network state transitions specifically include: the key dynamic features are obtained by extracting the temporal evolution features of node centrality and the temporal change features of global network topology indicators from the dynamic hypergraph sequence; the extracted node features, hyperedge features and global topology change features are summarized and post-processed to form a feature vector, which is then input into the hypergraph convolutional neural network model for training.

Images