Epileptic Seizure Detection Based on Convolutional Self-Attention Adaptive Dimensionality Expansion Network Using EEG Signals

Abstract

Abstract. Traditional epileptic seizure detection methods suffer from poor interpretability, excessive parameters, and fixed parameters in time-domain dimensionality expansion. To address these, this study proposes a convolutional self-attention adaptive dimensionality expansion network (CSADI-Net), integrating convolutional self-attention and adaptive dimensionality expansion. Convolutional self-attention uses convolutional layers to generate Q (query, representing task-related attention cues), K (key, representing inherent signal characteristics), and V (value, representing input data), reducing trainable parameters (TP). Adaptive dimensionality expansion combines with network training for parameter adjustment. Class activation heatmaps enable visual interpretability. Validated on children's hospital Boston and the Massachusetts institute of technology (CHB-MIT) (Accuracy:98.87%, F1:98.49%) and temple university hospital (TUH) (Accuracy:98.26%, F1:98.13%) datasets, it outperforms CNN, CNN-LSTM, and linear self-attention Transformer. With high accuracy, antinoise ability, and interpretability, it provides a new perspective for seizure detection.