Efficient Deep Spiking Neural Network for Complex EEG Signals

Spiking neural networks (SNNs) offer a biologically inspired, energy-efficient alternative to conventional artificial neural networks (ANNs). However, deep SNNs struggle to process complex EEG signals because their spike-based representations are sparse and highly time-specific, which often causes important information to be lost as it passes through multiple layers. In this work, we address these limitations by proposing an Efficient Deep Spiking Neural Network (EDSNN) for automatic sleep stage classification. The model is trained and evaluated on raw single-channel recordings from two large-scale sleep EEG datasets: Sleep-EDF-78 and SHHS. The proposed architecture incorporates multi-scale convolutional spiking layers with skip connections to preserve spiking information across deep layers. It also employs aggregated spike-based supervision, class-weighted label smoothing, and incremental data feeding to improve training stability and generalization. We evaluated the proposed EDSNN on two challenging and imbalanced sleep EEG datasets, SHHS and Sleep-EDF-78. The model achieved accuracies of 81.64% on Sleep-EDF-78 and, 80.55% on SHHS, demonstrating competitive performance compared to conventional deep neural networks, while offering faster inference and lower energy consumption due to its spiking architecture.