Multi-scale kernel and electrode attention network for EEG-based epileptic seizure detection

Epilepsy is a common chronic neurological disorder, and automated detection of epileptic seizures using multi-channel electroencephalography (EEG) is of great significance for improving diagnostic efficiency. Existing deep learning methods either reduce the number of EEG channels to lower computational costs or perform modality transformations to enhance feature representation. However, these approaches often result in information loss or additional computational cost, potentially leading to suboptimal detection performance in resource-constrained settings. To address these issues, we propose the Multi-scale Kernel and Electrode Attention Network (MKEANet). This novel end-to-end feature extraction network operates directly on full-channel EEG signals without channel reduction or modality conversion, thereby avoiding these limitations. MKEANet employs multi-scale convolutional structures to capture diverse electrode channel combinations. This design enables the model to adapt to temporal variations in epileptic discharge waveforms while enhancing spatial modelling of cross-channel EEG representations. An Electrode Attention Module (EAM) is also introduced to adaptively assign weights to different channels, thereby enhancing discriminative capability. Experiments on two public datasets demonstrate that MKEANet achieves state-of-the-art performance, with 99.29% accuracy, 98.29% sensitivity, and 99.59% specificity on the CHB-MIT dataset and 99.42% accuracy, 98.69% sensitivity, and 99.54% specificity on the Siena Scalp dataset, demonstrating the effectiveness of MKEANet in epileptic seizure detection based on raw full-channel EEG signals.