Epileptic seizures diagnosis and prognosis from EEG signals using heterogeneous graph neural network

Research contribution

This study’s main contributions are explained in more detail in the list format below.

• This study proposes a heterogeneous GNN model for epilepsy prediction using the CHB-MIT dataset that extends the capabilities of Conventional GNNs by accommodating various types of nodes and edges in the graph.

• This study performs exploratory data analysis and utilizes data preprocessing approaches, including signal segmentation, resampling, label encoder (one hot-encoding) and normalization (Z-score Normalization). Further, it employs a standard train-test split with stratified sampling to ensure class distribution and reduce bias.

• The research article uses several evaluation measurements such as accuracy, precision, recall and F1-score to analyze the model’s efficacy for identifying seizures and no-seizures. The results demonstrate that the proposed GNN approach correctly predicts epilepsy and surpasses the performance of other DL models.

Research organization

The study technique for classifying epilepsy is described as follows: A survey of the literature on DL and ML methodologies for classifying epilepsy detection is presented in “Related Work”. “Proposed Approach” provides the research approach for the proposed work, which utilizes the CHB_MIT dataset, a deep learning model, and data preprocessing. The conclusions and findings are discussed and explained in “Experimental Result and Analysis”. The work’s conclusion and recommendations for future direction are contained in “Conclusion”.

Experimental dataset and preprocessing

The first step in any detection system is selecting inputs for classification. Brain activity, measurable by electroencephalogram (EEG), begins in childhood and continues throughout life, reflecting overall health. For epilepsy detection, we used EEG signals from the CHB-MIT dataset, which includes scalp EEG recordings from ten pediatric patients (ages 1.5 to 15) with uncontrollable seizures (John, 2010; Deepa & Ramesh, 2021). This dataset, widely used for real-time automated seizure detection, supports our goal of developing a generalized EEG seizure sensor independent of specific patients or channels. Data preprocessing is crucial for improving model precision and efficiency. This section implements data preprocessing steps, including exploratory data analysis (EDA) and data cleaning, data splitting, normalizing the data using normalization (Z-score normalization) approaches, and turning the categorical data into numerical values.

Recurrent neural networks

RNNs are DL classifiers that process data sequentially. RNNs differ from standard feedforward neural networks in that they feature connectivity that forms internal cycles, enabling the network to remember prior data. Because of this, RNNs are especially effective for sequence-related tasks like speech recognition, natural language processing, and time series assessment. The basic idea behind RNNs is to use information from previous time steps to influence the current prediction or output. This enables the network to capture temporal dependencies in the data. However, conventional RNNs have drawbacks. For example, they have trouble learning long-term relationships because of problems like vanishing gradients. This study used a simple RNN layer with a single recurrently connected hidden layer. The number of units in the RNN layer is 32, which determines the size of the hidden state and the activation function used in the RNN layer, rectified linear unit (ReLU). The dense layer is used as the output layer, and the following parameters include num_classes, and the activation function is softmax.

Long short-term memory

LSTM was produced to successfully capture long-range relationships in data and solve the vanishing gradient issue that RNNs have. Applications combining time series data and natural language processing benefit greatly from the use of LSTMs (Alsubai et al., 2022). This study applied an LSTM model with two LSTM layers and two dropout layers. It then compiles the model with the sparse categorical cross-entropy loss function and the Adam optimizer. The model is trained using a custom training loop, where the gradients are computed and applied manually. The key parameters of an LSTM layer are units, return_sequences, dropout and recurrent_dropout. An LSTM model trains for 200 epochs with a batch size of 32.

Graph neural network

GNN is a framework developed to proceed and illustrate data in graphs or networks. Graphs here comprise nodes, also called vertices, connected by edges and links. GNNs are applied to obtain information and derive understandings from data (Scarselli et al., 2008; Zhou et al., 2020). Figure 2 depicts the architecture of the GNN model. The GNN model in this study consists of two dense layers. The first layer is a fully connected dense layer with 16 units. ReLU activation is chosen to introduce non-linearity and enable the model to learn complex patterns in the data. The second layer consists of num_classes units and softmax activation that are used to output the probability distribution over the classes for multi-class classification.

The loss function used is sparse categorical cross-entropy, and the optimizer is Adam. We used the sparse categorical cross-entropy loss function because it is well-suited for multi-class classification tasks where integers encode class labels. This choice helps the model optimize for the correct class predictions by minimizing the cross-entropy between the true and predicted distributions. The Adam optimizer was selected due to its adaptive learning rate properties, which provide efficient and robust training performance across a wide range of deep learning tasks. Adam combines the benefits of both AdaGrad and RMSProp, making it suitable for dealing with sparse gradients and noisy data. The number of epochs was set to 50, and the batch size was set to 32. These values were chosen based on preliminary experiments to balance training time and model performance. The epoch count ensures sufficient iterations for the model to converge, while the batch size balances computational efficiency and gradient estimate stability.

The proposed approach for epilepsy disease prediction on the CHB-MIT dataset is presented in Algorithm 1. Using CHB-MIT data as input, the system outputs predicted epilepsy disorders with seizure or no-seizure episodes. The three primary tasks of the algorithm are to assemble a dataset, train the model, and predict ratings. The main function combines the three functions to assemble the entire prediction system. Using EEG signals, the Assemble() function takes pertinent data from patient records. Following data retrieval, analysis and preprocessing are performed to detect missing values, eliminate duplicate data, encode labels, and apply the normalization procedure to transform the data into a standard format. After completing these procedures, the dataset is thoroughly cleaned and ready to be input into the DL model to build a prediction model. The preprocessed dataset is finally returned by the functions X and Y. The TrainingModel() function creates a DL model from the preprocessed dataset. The used dataset is divided into 20% and 80% weightage for testing and training data. The function finally provided the trained model in return. The trained function utilizes the model to categorize seizures and non-seizures. After receiving the preprocessed dataset, the function outputs the predicted classes and the trained model as inputs.

Evaluation measurements

The evaluation measures provided below are used to determine how effective the proposed model is. Among the measurements utilized to assess the classification problems are the accuracy provided in Eq. (2), the precision in Eq. (3), the recall in Eq. (4), and the F1-score in Eq. (5). A confusion matrix (CM) table can assess a classification model’s performance by contrasting its expected and actual outputs. Machine learning is frequently used to assess a categorization model’s efficacy. Each of the four quadrants of a CM—true positive (TP), false positive (FP), true negative (TN), and false negative (FN)—represents a potential result. The matrix columns illustrate the anticipated class labels, and the rows illustrate the actual ones. The matrix’s main diagonal represents the successfully categorized examples, and the off-diagonal entries illustrate the incorrectly classified samples. These measurements aid in determining the model’s advantages and disadvantages and can be applied to enhance the model’s performance.

(2) $$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

(3) $$Precision=\frac{TP}{TP+FP}$$

(4) $$Recall=\frac{TP}{TN+FN}$$

(5) $$F1-score=2\times \frac{Precision\times Recall}{Precision+Recall}.$$

The receiver operating characteristic (ROC) curve is a graphical representation used to evaluate the performance of binary classification systems. It plots the true positive rate (TPR) against the false positive rate (FPR) at various threshold settings, providing a comprehensive visualization of a model’s ability to distinguish between positive and negative classes. A model with a ROC curve closer to the top-left corner indicates a higher discriminative ability. In contrast, the area under the curve (AUC) summarizes the overall performance: a value of 1 represents perfect accuracy, while 0.5 suggests no better performance than random guessing. This section provides a detailed discussion of the experiment outcomes.

The classification performance results for RNN, LSTM, and GNN models are shown in Table 2. GNN has the highest accuracy, 98.00%, for both the “No-Seizure” and “Seizures” classes, indicating that it makes correct predictions more often than RNN and LSTM. LSTM achieves an accuracy of 97.00%, while RNN achieves 98%. LSTM achieves testing F1-scores of 98.00% for the “Seizures” class and “No-Seizure” class. RNN achieves F1-scores of 98.00% for the “No-Seizure” and “Seizures” classes, respectively.

Table 3 presents the time required for training and testing predictions for the LSTM, RNN, and GNN models. Among the three models, the GNN model is the fastest, with a training time of 0.0081 s and a testing time of 0.0054 s, demonstrating its computational efficiency. The LSTM model takes the longest time for training, requiring 0.1731 s, while its testing time is slightly shorter at 0.126 s. The RNN model has a moderate training time of 0.029 s and a testing time of 0.0144 s. These results highlight that the GNN model significantly outperforms RNN and LSTM in terms of computational speed, making it more suitable for applications requiring real-time predictions.

Figure 3 depicts the graphical representation of the RNN model by evaluation training, testing loss, accuracy, CM, and ROC. The left Fig. 3A shows the training and validation loss. Validation loss starts from $0^{th}$ epoch with value $0.18\mathrm{\%}$, and after some fluctuation increase and decrease, it decreases to $0.1\mathrm{\%}$ at ${50}^{th}$ epochs. The right graph in 3A displays the model’s training and validation accuracy. Validation accuracy starts from $0^{th}$ epoch with a value around $0.95\mathrm{\%}$. After some fluctuation increase and decrease, it stops at $0.98\mathrm{\%}$ at ${50}^{th}$ epoch. The left graph in Fig. 3B is a training confusion matrix that shows it performs better because the proposed strategy produces more continuous, better true positive (13,421) and negative (71) values and fewer false positive (365) and negative (13,175) results. The right graph in Fig. 3B is a testing confusion matrix. It also produces more continuous, better true positive (3,374) and negative (30) values and fewer false positive (89) and negative (3,266) results. Figure 3C displays the ROC curves for both classes. For Class 1, the train curve area is expressed by the blue line with a 0.98 value, and the orange line displays the test ROC curve with the same value. The train ROC curve for Class 2 is shown by the green line with a value of 0.98, and the red line shows the test ROC curve with the same value for Class 2.

Figure 4 displays the graphical representation of the LSTM model by evaluation training, testing loss, accuracy, CM, and ROC. Figure 4A displays the training and testing accuracy of the model. Training accuracy starts from $0^{th}$ epoch with value $0.940\mathrm{\%}$, and after some fluctuation, it increases to $0.972\mathrm{\%}$ at ${50}^{th}$ epoch. Testing accuracy starts from $0^{th}$ epoch with value $0.943\mathrm{\%}$, and after some fluctuation of increase and decrease, it increases to $0.975\mathrm{\%}$ at ${50}^{th}$ epoch. Figure 4B shows the training and testing loss, which decreases as the number of epochs increases. Training loss starts from $0^{th}$ epoch with value $0.18\mathrm{\%}$, and after some fluctuation increase and decrease, it decreases to $0.09\mathrm{\%}$ at ${50}^{th}$ epoch. Testing loss starts from $0^{th}$ epoch with value $0.17\mathrm{\%}$, and after some fluctuation increase and decrease, it decreases to $0.08\mathrm{\%}$ at ${50}^{th}$ epoch. Figure 4C is a training confusion matrix that performs better because of true positive (13,415) and negative (12,902) values and fewer false positive (77) and negative (639) results. The graph in Fig. 4D is a testing confusion matrix that performs better because of true positive (3,378) and negative (3,221) values and fewer false positive (26) and negative (134) results. Figure 4E displays the ROC curves for both classes in which the train ROC curve for classes 1 and 2 is 0.99, and the test ROC curve for Classes 1 and 2 is 0.99.

Figure 5 displays the graphical representation of the GNN model by evaluation training, testing loss, accuracy, CM, and ROC. The graph in Fig. 5A displays the training and testing accuracy of the model. Training accuracy starts from $0^{th}$ epoch with value $0.955\mathrm{\%}$, increasing to $0.978\mathrm{\%}$ at ${10}^{th}$ epoch. After some minor fluctuation of increase and decrease, training accuracy stops at $0.985\mathrm{\%}$ at ${50}^{th}$ epoch. Testing accuracy starts from $0^{th}$ epoch with value $0.956\mathrm{\%}$ increasing to $0.981\mathrm{\%}$ at ${10}^{th}$ epoch. After some minor fluctuation of increase and decrease, training accuracy stops at $0.983\mathrm{\%}$ at ${50}^{th}$ epoch. The graph in Fig. 5B shows the training and testing loss. Training loss starts from $0^{th}$ epoch with value $0.16\mathrm{\%}$, decreasing to $0.05\mathrm{\%}$ at ${50}^{th}$ epoch. Testing loss starts from $0^{th}$ epoch with value $0.16\mathrm{\%}$, decreasing to $0.06\mathrm{\%}$ at ${50}^{th}$ epoch. The graph in Fig. 5C is a training confusion matrix graph that performs better because the proposed strategy produces more continuous, better true positive (10,912) and negative (9,600) values and fewer false positive (4) and negative (1,381) results. The graph in Fig. 5D is a testing confusion matrix graph that performs better because the proposed strategy produces more continuous, better true positive (3,401) and negative (2,965) values and fewer false positive (3) and negative (390) results. The graph in Fig. 5E displays the ROC curves for both classes. For Class 1 and 2, the train and test ROC curve value is 0.98.

Comparison with conventional ML approaches

Table 4 compares different models for epileptic seizure prediction, focusing on the authors, classifiers used, dataset employed, and the corresponding performance. Wang et al. (2019) used the random forest (RF) classifier and achieved a performance of 84.00%. Varnosfaderani et al. (2021) utilized a two-layer LSTM using the Swish activation function classifier and achieved a performance of 85.1%. Rasheed et al. (2021) implemented a deep convolutional generative adversarial network (DCGAN) model for classification and achieved a performance of 88.21%. Dissanayake et al. (2020) employed one of the CNN architectures and achieved a performance of 88.81%. Yu et al. (2022) utilized a machine learning long short-term memory (MLSTM) model for epileptic seizure prediction and achieved a performance of 89.47%. Zhang, Liu & Chen (2022) used a Transformer-based model and achieved a performance of 89.50%. All models, including the proposed model, used the CHB-MIT dataset for epileptic seizure prediction. This research utilizes a GNN as a proposed model and achieves the highest performance with 98.90%. The proposed GNN model outperforms the other models in terms of predictive accuracy, achieving the highest performance. The performance differences between the models are relatively small, but even marginal improvements can be significant in medical applications.

Discussion and findings

Conventional deep learning methods, such as CNNs, require graph data to be transformed into fixed-size feature vectors, often leading to information loss and disregarding the relational structure. GNNs operate directly on graph structures, enabling them to capture relational dependencies and extract node embeddings that preserve local and global graph properties. Conventional methods need mechanisms for effectively aggregating information from neighbouring nodes in a graph, limiting their ability to exploit graph topology. GNNs leverage message-passing algorithms to propagate and aggregate information across graph nodes, enabling them to capture complex relationships and dependencies. Conventional methods struggle to handle variable graphs and often require fixed-size input representations, making them less adaptable to diverse graph structures.

GNNs are inherently flexible and can operate on graphs of varying sizes and structures, making them suitable for a wide range of graph-based tasks without requiring extensive preprocessing. Conventional methods like image classification may outperform GNNs in certain tasks with well-defined input-output mappings and sample-labelled data. GNNs excel in tasks involving graph data, including node classification, link prediction, and graph classification, where capturing relational dependencies is crucial for accurate predictions. GNNs offer several advantages over conventional DL methods for graph data analysis. By operating directly on graph structures and leveraging message-passing mechanisms, GNNs can effectively capture relational dependencies and extract valuable insights from complex graph data.

While this approach has demonstrated promising results, we recognize that using heterogeneous GNNs for epileptic seizure detection does not introduce novel improvements or new variant algorithms specifically tailored for this domain. This is an important consideration, as similar methodologies have been extensively researched and applied in other fields. Therefore, future research endeavours should focus on enhancing graph Neural Network models for epileptic seizure prediction by proposing innovative improvements or introducing new variant algorithms. This could include designing custom GNN layers tailored for EEG data or developing novel graph encoding techniques that better capture the unique characteristics of EEG signals.

Moreover, comprehensive comparative studies with existing heterogeneous GNN approaches across different domains should be conducted to highlight the unique benefits and potential adaptations for epileptic seizure detection. Empirical evidence, including ablation studies and evaluations using diverse performance metrics, is crucial to substantiate claims of improved performance or novelty. Future research should also explore novel applications of EEG data in seizure therapy and treatment detection. By addressing technical challenges, fostering interdisciplinary collaboration, and prioritizing patient-centred outcomes, researchers can contribute to developing more effective and personalized approaches to epilepsy management. By acknowledging the current limitations and focusing on these areas for future research, we aim to make substantial contributions to the field, advancing the application of GNNs in medical data analysis and improving the detection and understanding of epileptic seizures from EEG data.

We acknowledge the inherent subject dependency in EEG signals, characterized by non-linearity and variability. Although we did not implement Leave-One-Subject-Out Cross-Validation (LOSO CV), we employed a standard train-test split with stratified sampling to ensure class distribution and reduce bias. Feature standardization was applied to normalize the data, and a GNN was utilized to capture complex relationships within the EEG signals. We assessed model performance using various metrics, including accuracy, F1 score, and ROC AUC. Future work will incorporate the LOSO CV for a more robust evaluation of model generalization across subjects.

Limitation of the proposed approach

Our research demonstrates the efficacy of the proposed GNN model for EEG-based seizure detection, and it is important to acknowledge certain limitations that may affect the generalizability and applicability of our findings. Our model was trained and evaluated exclusively on the CHB-MIT scalp EEG dataset, which, despite being well-established, limits the generalizability of our findings to other datasets with different characteristics. Future studies should validate the model on diverse datasets to ensure broader applicability.

Our evaluation focused on accuracy, precision, recall, and F1-score but did not fully address clinical implications, such as the impact of false positives and false negatives. Incorporating additional metrics like AUC-ROC and analyzing clinical relevance will provide a more comprehensive evaluation. The complexity of GNNs also poses a challenge to model interpretability, which is crucial for clinical decision-making. Future work should incorporate explainability techniques to enhance transparency and improve trust in clinical applications. While Z-score normalization was effective, exploring a broader range of preprocessing techniques could further optimize performance. Furthermore, the substantial computational resources required for the GNN model may limit its applicability in resource-constrained settings, highlighting the need for optimizing computational efficiency through techniques like model compression, pruning, quantization, or leveraging lightweight GNN frameworks.