Figure 1 provides an overview of the overall process for classifying epileptic EEG signals using the MGCNA proposed in this article. First, intrachannels features are extracted from the raw EEG signals to create graph signals. Second, three different time-series-to-graph representation are employed and graph features are extracted through GCN. Third, a multi-head attention is utilized to learn dependencies of different graph features. Finally, the combined graph features are processed to output classification results.

Convolution has been proven to be highly effective in capturing features of EEG signals. Inspired by the EEGWaveNet (Thuwajit et al., 2021) model, this module utilizes depthwise convolution to compress the input signals’ resolution in a channel-wise manner and capture features. Depthwise convolution does not cross information between channels and extracts features independently. This module comprises k consecutive layers, where each layer captures valuable features within channels at half the scale of the resolution of the previous layer.

In this module, the input data X are shaped as ( n , 1 , C ) , where n and C represent the number of channels and the number of sampling points, respectively. n is set to 22, and C is defined as f × t , where f denotes the EEG’s sampling frequency, and t is the duration of each epoch’s segmentation. In each layer, depthwise convolution is applied to each channel with a kernel size of (1, 2), a stride of 2, and no padding. The output size of each k th layer is ( n , 1, C / 2 k ).

The module of extraction of intrachannels features can be represented as: F k = D W C o n v F k − 1 (1) where F 0 is the input signal X ∈ R n × 1 × C , D W C o n v ( ⋅ ) is the depthwise convolution. After reshape, the features extracted by this module N ∈ R n × ( C / 2 K ) will be fed into the next module to obtain multi-scale graph features. Figure 2 depicts the process of extracting intrachannels features.

To design an appropriate graph representation for EEG signals, it is necessary to construct a graph based on the physical distribution of channels and the spectral-temporal domain characteristics of the signals. Each channel was considered as a node and the connection of each node is characterized in terms of spatial position information, EEG signal similarity features, and a learnable graph representation to dynamically capture the topological relationship between EEG signal channels. These three types of graph signals are input into GCN, allowing us to leverage the structural characteristics of EEG signals to capture spatial dependencies between nodes. Figure 3 illustrates the overall process of multi-branch GCN. The graph convolution processes at the top, middle, and bottom are based on spatial distance graph representation, functional connectivity graph representation, and adaptive graph representation, respectively.

The attention mechanism, which is a deep learning model, emulates the pivotal information and critical elements that individuals concentrate on during observations. The multi-head attention mechanism makes the output of attention to incorporate encoding from different spatial locations, thereby enhancing the expressive power of the model. In standard multi-head attention with input matrix Z ∈ R m × n × D , the self-attention for each head are computed according to the following formula: H e a d Q , K , V = S o f t m a x Q K T d V (10) where Q , K , V represent the query vector, key vector and value vector, respectively. h is the number of heads, d = m / h . The original multi-head attention mechanism has a high computational cost, with a quadratic complexity of O ( L 2 d ) , where L is n × D .

Inspired by MobileViT v2 (Mehta and Rastegari, 2022) and HViT-DUL (Deng et al., 2023), the multi-head design is incorporated into the separable self-attention mechanism, which is applied to the proposed model for the reduction of computational overhead. The separable multi-head attention mechanism employed in this paper has a linear complexity of O ( L ) , which is lower compared to the standard multi-head attention mechanism. When multiple self-attention mechanisms are sequentially linked and subsequently subjected to a final projection, it results in the generation of the ultimate values for the multi-head attention mechanism. The spatial topological information and dynamic variation information of multi-channel electrodes are obtained through a multi-head attention mechanism to combine representations of multiple graph signals, thereby enhancing the discrimination ability. Figure 4 illustrates the process of separable multi-head attention. Q , K , V is calculated using the following formula: Q = C o n v 1 Z K = C o n v 2 Z V = C o n v 3 Z (11) where Z is the output of multi-branch GCN features. C o n v is two-dimensional convolution. We employ three 1 × 1 convolutional layers, with parameters not shared, to compute the Q K V of the input graph features. After the reshaping process, the outputs Q , K , V are in the following dimensions: Q ∈ R h × 1 × n × D , K ∈ R h × d × n × D and V ∈ R h × d × n × D .

To calculate self attention scores, we apply the S o f t m a x operation to the Q and taking the element-wise multiplication with K , After that, we perform element-wise multiplication with V after passing it through ReLU. Self attention scores A t t e n i are computed based on the following formula: a t t e n i = R e L U V ∗ K ∗ S o f t m a x Q (12) where a t t e n is the self-attention scores. R e L U is activation function. After concatenating the scores of multiple self-attention mechanisms, we perform a 3 × 3 convolution on the resulting features, normalize it using BN, and then apply ReLU to enhance generalization ability. The computed multi-head attention scores, matching the dimensions of the input Z , are added to the output of multi-branch GCN, expressed as: A t t e n = a t t e n 1 , a t t e n 2 , … , a t t e n h (13) M A = R e L U B N C o n v A t t e n (14) Z ′ = B N Z + M A (15) where [ ⋅ ] represents the concatenation.

A classifier module is placed after the multi-head attention mechanism for the final class inference. Initially, it involves two layers of 2D convolution, with a R e L U activation function applied after each convolution to enhance the robustness of model. This process can be described as: F ′ = R e L U C o n v R e L U C o n v Z ′ (16) where Z ′ is the output of multi-head attention mechanism. Global average pooling pools each channel of the convolution result, followed by a linear layer that reduces the feature dimensionality to match the number of classes. After passing through a sigmoid operation, the index with the highest probability is selected as the classification result. This process can be described as: P r e s = σ L i n e a r A v g P o o l F ′ (17) where A v g P o o l ( ⋅ ) denotes global average pooling, σ is sigmoid operation, L i n e a r ( ⋅ ) is the Linear operation, and P r e s is the final classification result. Figure 5 illustrates the structure of the classifier module.

In order to achieve optimal network parameters during the training process, we employ the backpropagation (BP) algorithm, which iteratively updates the network parameters until an optimal or suboptimal solution is reached. In this context, we introduce a loss function based on binary cross entropy, defined as follows: l o s s = c r o s s _ e n t r o p y P r e , P r e s + τ ‖ Ω ‖ (18) c r o s s _ e n t r o p y P r e , P r e s = − P r e ⁡ log ⁡ P r e S + 1 − P r e log 1 − P r e S (19) where P r e and P r e s represent the actual label and the predicted label, respectively. The binary cross entropy, denoted as c r o s s _ e n t r o p y ( P r e , P r e s ) , quantifies the disparity between the true labels and the predicted labels. τ represents the trade-off parameter, while ‖ ⋅ ‖ denotes the l 2-norm to prevent overfitting. Ω represents all the parameters within this model.

For adaptive graph learning, the adjacency matrix A3 is a trainable parameter within the network that optimizes with model optimization. The partial derivatives of the loss function with respect to the optimal adjacency matrix A3 and the loss are expressed as follows: ∂ l o s s ∂ A = c r o s s _ e n t r o p y P r e , P r e s ∂ A + τ ‖ Ω ‖ ∂ A (20)

The CHB-MIT dataset (Shoeb, 2009) was collected by Boston Children’s Hospital and is currently the most widely used public dataset. The multichannel scalp EEG signals consists of EEG signal recordings from 23 children with epilepsy at a sampling rate of 256 Hz. The EEG signals are collected using EEG electrodes placed according to the International 10–20 system. The dataset comprises approximately 983 h of EEG recordings, including 198 seizure onset events, with a total duration of 3 h 15 min.

In the CHB-MIT dataset, most records have 23 channels, while some records have missing or duplicated channels. In this study, 22 channels are selected in this study to maintain consistency of channels among all patients. The selected 22 channels are FP1-F7, F7-T7, T7-P7, P7-O1, FP1-F3, F3-C3, C3-P3, P3-O1, FP2-F4, F4-C4, C4-P4, P4-O2, FP2-F8, F8-T8, T8-P8, P8-O2, FZ-CZ, CZ-PZ, P7-T7, T7-FT9, FT9-FT10, FT10-T8. This dataset is classed into two categories, including interictal and ictal. To reduce the effect of noise, a fifth-order Butterworth band-pass filter ranging from 0.5 Hz to 70 Hz was used. The filtered EEG signals are segmented into 3-s windows. Due to the scarcity of ictal EEG data compared to interictal EEG data, the windows overlap by 50 % to obtain more ictal data.

For a comprehensive evaluation of the MGCNA, we train both a patient-independent model and a patient-specific model. In a patient-specific approach, the model is trained, validated, and tested using data from an individual subject. Interictal signals are randomly discarded for each subject, and the ratio of interictal signals to ictal signals is maintained at 5:1. As the number of ictal signals is significantly lower than that of interictal signals, which is detrimental to model training, we employ re-sampling to create a balanced training dataset. In each epoch of training, a random selection of an equal number of interictal signals is made to match the ictal signals in the training set, followed by random shuffling. Consequently, we obtain a balanced training subset for each epoch. We employ a 5-fold cross-validation approach and report the average performance. In a patient-independent approach, the model was trained and validated using data from multiple subjects, and then tested on data from individual subject. We utilized the leave-one-subject-out cross-validation as evaluation method. The data in both categories are kept balanced. The patient-specific approach emphasizes individual differences and personalization, and the patient-independent approach focuses more on overall trends and general patterns.

We employ the Adam optimizer with a learning rate of 3e-4, a weight decay of 1e-3.The dropout rate is experimentally set to 0.5. R 1 is set to 0.4 and R 2 is set to 0.25. k is set to 2. All the above experiments were performed and implemented by Pytorch 1.7.1 in the NVIDIA GTX3090 and CUDA11.0 environment.

We use eight different performance metrics to evaluate the performance of MGCNA including Accuracy (Acc), Sensitivity (Sen), Specificity (Spe), F 1 Score ( F 1 ) , and Area Under Curve (AUC). These metrics are obtained using true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) values.

Accuracy represents the proportion of samples correctly classified by the model out of the total sample count. Sensitivity and Specificity measure the model’s ability to classify positive and negative samples. F1 Score combines precision and recall to provide a comprehensive performance measure. AUC is based on the area under the ROC curve and provides a comprehensive performance metric for unbalanced category distributions.

Patient-specific experimental results on CHB-MIT dataset using the MGCNA are shown in Table 1. According to the table, the model demonstrates an average accuracy of 99.32 % , specificity of 98.4 % , sensitivity of 98.74 % , an F1 score of 97.76 % , and an AUC of 98.51 % . For the majority of patients, the recognition accuracy exceeds 97 % , with one patient achieving 100 % accuracy. The model achieves a specificity of 99 % for 12 patients, which represents 50 % of all test subjects, and a sensitivity of 99 % for 17 patients, covering 70 % of all test subjects. F1 scores for all patients are greater than 92 % , and the AUC is above 95 % . These results indicate the stability and high performance of MGCNA proposed in this study, which can aid medical professionals in diagnosis.

In contrast to the patient-specific experiments, patient-independent experiments involve training a general model for all patients. The experiment employed the leave-one-out method. Patient-independent experiments separate one patient’s signal as a test set, using the EEG data of other patients as the training set and Validation set. This allows the model to detect epileptic activity in the test set by learning common seizure characteristics from the training set. This experiment is more clinically meaningful but requires to address differences in EEG signals among different patients, which may result from physiological variations, equipment noise, etc. Compared to patient-specific experiments, patient-independent experiments make it more challenging for the algorithm to identify epileptic seizures and lead to poorer detection results.

Patient-independent experimental results on CHB-MIT dataset using the MGCNA are shown in Table 2. As seen in the table, the average values for Accuracy, Sensitivity, Specificity, and F1 are all above 80 % . There is a significant individual variation among different patients. Patients ♯ 4, ♯ 5, ♯ 7, ♯ 9, ♯ 10, ♯ 15, ♯ 17, ♯ 18, ♯ 19, ♯ 22, ♯ 23 achieved accuracy rates exceeding 90 % , whereas patients ♯ 8, ♯ 12, ♯ 13, and ♯ 24 had comparatively lower accuracy rates. Compared to patient-specific experiments, patient-independent experiments exhibited a 13.95 % lower accuracy rate. This is due to the differences in seizures between patients and the presence of different types of symptoms during seizures, which vary in frequency and duration.

To assess the contributions of different components in our model to classification performance, we conduct comparative experiments using model based only on similarity graph representation, model based only on distance graph representation, model based only on trainable graph representation, and model without multi-head attention. To ensure the fairness of the experiments, the model settings are kept consistent. The results are shown in Figures 6, 7.

Graph convolutional network uses only spatial distance graph learning, denoted as ∗h1. Graph convolutional network uses only functional connectivity graph learning, denoted as ∗h2. Graph convolutional network uses adaptive graph learning, denoted as ∗h3. Non-attention graph representation models refers to the model that does not employ self-attention mechanisms to obtain self-attention scores for the three graph representations. We directly concatenate the representations of three graphs into a graph feature, denoted as ∗self-attention.

From the Figures 6, 7, it is evident that MGCNA outperforms the other four ablation experiments. Compared to our model in this paper, the accuracy of the four ablation experiment models is lower by 5.22 % , 5.30 % , 3.37 % , and 3.99 % , respectively. As illustrated in Figure 6, the trainable graph representation model demonstrates superior performance compared to the other two graph representation learning methods, particularly for patients ♯ 4, ♯ 12, ♯ 13, ♯ 14, ♯ 17, and ♯ 21. Between spatial distance graph learning and functional connectivity graph learning, the latter more effectively captures the interrelationships among EEG signals.

The effect of the self-attention mechanism on epilepsy detection approximates that of the model using only distance graph representation. Figure 7 illustrates that allocating different attention weights to graph features through self-attention mechanism helps to capture dependencies between graph features and enhances the model’s ability to model relationships between graph features, which can achieve good performance in epilepsy detection. The model can effectively utilize the spatiotemporal relationships among EEG signals, supplement information through learnable graph representations, and obtain attention scores through self-attention mechanisms, and therefore the model’s learning capability has been significantly enhanced.

To further validate the superiority of MGCNA, t-SNE is applied to visualize and analyze the features extracted from the ablation study. As depicted in Figure 8, the t-SNE visualization in two-dimensional embedding space illustrates the interictal and ictal features for both patient-specific and patient-independent experiments. We can see that in both patient-specific and patient-independent experiments, our approach exhibits superior recognition capabilities compared to the ablated models. Particularly, models utilizing only one graph construction method tend to confuse some interictal and ictal features. In contrast, better discriminative features were obtained using MGCNA, mainly in terms of significant inter-ictal distances and dense intra-ictal distributions. These observations indicate that combining multi-branch GCN with self-attention mechanisms can yield the optimal performance for epileptic seizure classification.

In both functional connectivity graph learning and spatial distance graph learning methods, it is necessary to set a threshold for constructing an adjacency matrix. The constructed adjacency matrix must not only ensure the sparsity of the graph but also be capable of distinguishing temporal and spatial characteristics of different types of EEG data to enhance the accuracy of the model in identifying epileptic seizures.

Figure 9A depicts the average epileptic seizure detection results of the CHB-MIT dataset in patient-specific and patient-independent experiments, under varying thresholds R 1 for spatial distance graph learning. The threshold range selected spans from 0.2 to 0.7. Notably, at a threshold of 0.2, the patient-specific epileptic seizure detection accuracy is the lowest, as lower thresholds tend to introduce excessive irrelevant physical connections between unrelated nodes in the spatial distance graph. Therefore, to maintain the sparsity of the adjacency matrix and achieve optimal seizure detection results, configuring the threshold R 1 for spatial distance graph learning to 0.4 is advocated.

Figure 9B illustrates the average epileptic seizure detection results in both patient-specific and patient-independent experiments under various threshold values R 2 for functional connectivity graph learning. The threshold range selected spans from 0.1 to 0.5, with a stride of 0.05. From Figure 9, it is evident that R 2 has a more pronounced impact on patient-independent epileptic seizure detection results. In the patient-specific experiment, where the training and testing datasets originate from the same patient, they exhibit similar data distributions. However, in the patient-independent experiment, differences exist between various patients, leading to inconsistent effects of different thresholds on the patients. Appropriate thresholds are applied to eliminate unrelated channel signals, utilizing functional connectivity to capture neuronal synchronized discharge during epileptic seizures (Abbas et al., 2021). To maximize the spatial discriminative power of the functional connectivity graph and achieve optimal results for seizure detection, the threshold R 2 for functional connectivity graph learning is set at 0.25.

The spatial distance graph learning based on Equation 3 with the threshold R 1 set to 0.4 is illustrated in Figure 10. The distance between two bipolar derivations was determined by measuring the separation between the centers of the bipolar derivations. For instance, the coordinates of FP1-F3 are the centers of the FP1 coordinates and the F3 coordinates. After calculating the Euclidean distances, these distances are normalized to create the distance graph representation, and self-connections are added, i.e., the graph representation of the Euclidean distances plus the diagonal matrix.

Figure 11 depicts the functional connectivity graph learning of interictal and ictal signals under thresholds set at 0.25. In the figure, we have selected and described the functional connectivity graph learning that represents interictal and ictal signals with the highest degree of similarity. The values in the graph are calculated as the Pearson correlation coefficients, which are then normalized into a similarity adjacency matrix ranging from 0 to 1. From Figure 11, it can be observed that functional connectivity graph learning for interictal signals exhibits a strong degree of similarity, while the ictal signal displays weaker inter-correlations across many bipolar derivations. Graph representations based on functional connectivity can depict the functional connections between brain regions as interdependencies among EEG signals.

Table 3 presents a performance comparison between the proposed MGCNA and state-of-the-art epilepsy detection algorithms in the patient-specific experimental setting. Most of these methods involve feature extraction and deep learning algorithms for epilepsy detection. It is evident that the advancement of deep learning methods holds significant importance for signal recognition and detection.

Machine learning methods are classical approaches for epilepsy detection. Ein Shoka et al. (2021) conducted feature extraction after channel selection and evaluated the performance of different classifiers for epilepsy detection, with KNN exhibiting the best classification performance. Sukriti et al. (Sukriti et al., 2021) extracted multiscale spectral features (MSSFs) and employed a random forest classifier for epilepsy classification exhibits a higher specificity compared to MGCNA (99.17 % versus 98.4 % ). However, the MGCNA achieves higher accuracy and sensitivity than that method. Cimr et al. (2023) achieved EEG classification by normalizing input signals and an 8-layer depth CNN model. The MGCNA is superior to that method in terms of accuracy, specificity and sensitivity. Zhao et al. (2023a) proposed using CNNs to extract local features and transformers to capture global information. Although MGCNA outperforms this approach, the differences in preprocessing methods also affect the results of epilepsy detection.

GCN can analyze signals by considering the three-dimensional spatial positions of EEG electrodes, thereby compensating for potential spatial information loss in deep learning methods such as CNN. The introduction of GCN can be utilized to analyze the spatiotemporal correlations among channels. Methods based on GCN have already demonstrated excellent performance for the detection of epileptic signals. Wang et al. (2023b) introduced a two-stream graph-based framework for learning the Weighted Neighbour Graph (WNG) representation in both the frequency and time domains. However, this approach achieved inferior accuracy, sensitivity, and specificity, with values of 93.1 % , 91.8 % , and 96.3 % , respectively, in comparison to the method proposed in the current study. Zhao et al. (2023b) introduced a hybrid Attention Network that utilizes the GAT to extract spatial features and the Transformer to extract temporal features addresses the issue of imbalanced data. However, in the patient-specific experiment, it achieved slightly lower accuracy, specificity, and sensitivity compared to MGCNA. He et al. (2022) utilized the GAT to extract spatial features and employed a BiLSTM network to capture temporal features, achieving an accuracy of 98.52 % on the CHB-MIT dataset, slightly lower than MGCNA.

Table 4 presents comparisons of performance between the proposed method and state-of-the-art epilepsy detection algorithms in the patient-independent experimental setting. Due to the substantial variations in signals among different patients, the overall performance of the patient-independent experiment is lower than that of the patient-specific experiment. The accuracy of the patient-independent experiment based on SVM (Shoeb, 2009) only reaches 58.32 % , indicating a significant performance gap compared to MGCNA. Wei et al. (2019) proposed an epilepsy detection algorithm by combining CNN and Wasserstein Generative Adversarial Nets (WGANs), achieving lower sensitivity than MGCNA, i.e. 72.11 % vs. 83.54 % , respectively. Zhao et al. (2023b) introduced the HAN model, which performs epilepsy detection not only in patient-specific experiments but also in patient-independent experiments. In both types of experiments, the three performance metrics were lower than MGCNA. Additionally, the ablation studies conducted in the patient-independent experimental setting are showcased herein, demonstrating that the MGCNA method exhibits superior performance compared to each ablation study. Since the three graph learning models can better extract the structural information of EEG and perform graph feature fusion, the overall performance is better than other epilepsy detection algorithms.

Through the above analysis, the MGCNA framework offers several key advantages. First, it employs a multi-branch graph convolutional network structure that dynamically learns temporal correlations and spatial topological information, enhancing the ability to process complex EEG signals, particularly in capturing spatial and temporal dependencies in epileptic brains. Second, the use of three graph learning approaches allows for a comprehensive evaluation of connectivity and synchronization across multiple channels, improving adaptability in different patient scenarios. Additionally, the multi-head attention mechanism further strengthens the framework’s ability to handle local features and complex EEG patterns. Experimental results demonstrate that MGCNA outperforms other methods in both patient-specific and patient-independent tasks, highlighting its strong generalization capabilities. Lastly, as an end-to-end automatic seizure detection model, MGCNA can be applied in clinical decision-making, helping clinicians diagnose childhood epileptic seizures more quickly and accurately, providing significant practical value.

While the MGCNA achieved satisfactory results in both patient-specific and patient-independent experiments, the MGCNA has several limitations. Firstly, spatial distance graph learning requires the prior determination of channel locations and is not robust to variations in EEG channel count. If channels are missing or if channel locations change, it can adversely affect recognition performance. Secondly, our model has a relatively long computation time because it involves calculating Pearson correlation coefficients for each sample to extract features for constructing EEG graph representations. Finally, the performance of model may depend on the quality and diversity of the data it has been trained on, and it is crucial to validate it on larger, more diverse datasets to ensure its generalizability. Specifically, both functional connectivity graph learning and spatial distance graph learning employ thresholds.

In future work, it is essential to explore graph representation methods that are more suitable for epileptic signals. While many scholars have already employed various graph representations (Raeisi et al., 2022), these often involve manual feature engineering and have long computation times. In the future, it’s possible to explore alternative composition methods or utilize clustering algorithms to identify the most suitable adjacency matrix. In the future, we will explore alternative graph generation techniques, such as imposing appropriate constraints on trainable adjacency matrix or using clustering algorithms to identify the most suitable graph generator. Identifying the most suitable graph representation method for epileptic signal recognition is of paramount importance. In this study, patient-independent experiments hold more clinical relevance, and there is significant room for improvement in accuracy. To capture common seizure characteristics among different patients, techniques such as transfer learning will be considered to enhance the accuracy of patient-independent experiments. In current research, considerable attention has been devoted to the occurrence of seizures, yet there exist variations in seizure types among individuals. Consequently, in future investigations, emphasis will be placed on the analysis of seizure types in epilepsy research.