Epileptic Seizure Detection from EEG Signals using Autoencoder-based Graph Convolutional Neural Network

Electroencephalogram (EEG)-based computer-assisted technology for automatic seizure detection empowers neurologists to make accurate diagnoses. Despite advancements, the potential of deep neural networks in seizure detection from time-series brain data remains underutilized. The focus on grid-like data has been a significant limitation of previous approaches. Physiological recordings often have irregular and unordered structures, making them difficult to consider as a matrix. Since EEG signals are non-stationary and variable, collecting sufficient EEG data requires a lot of time and has a poor sample-to-feature ratio, which significantly lowers the effectiveness of seizure detection. In this study, a novel hybrid framework is proposed for detecting epileptic seizures using a graph convolutional neural network (GCNN) combined with an autoencoder (AE). The nodes and vertices of the GCNN are able to accurately detect seizures by combining temporal analysis with anatomical disruption without any loss of adjacent spatial information. In the proposed method autoencoder (AE) is applied in the place of the traditional filtering method in two ways, conventional autoencoder (CAE) for feature extraction and variational autoencoder (VAE) for data augmentation which provides improved accuracy over traditional classification using GCNN. Data augmentation (DA) based signal synthesis is employed to scale down the imbalance of acquired data and reduce biases. Furthermore, the application Chebyshev learning filter and graph coarsening produces 6-8% more improved results than the conventional CNN-based method with a much lower Gaussian error rate. Experiments on the NMT-Scalp dataset show that the model effectively detects seizures from raw EEG signals, achieving 98.89% accuracy without additional feature extraction.