EEGSNet: A novel EEG cognitive recognition model using spiking neural network

• Published in: Biomedical signal processing and control Vol. 105; p. 107610
• Main Authors: Shi, Jiankai, Zhao, Yue, Wang, Chu, Zeng, Hong, Dai, Guojun
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2025
• Subjects: Adaptive thresholding; EEGSNet; Electroencephalography (EEG); Spiking neural network (SNN); 4D; EEGSNet; Electroencephalography (EEG); Spiking neural network (SNN); Adaptive thresholding
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.107610

Spiking Neural Networks (SNNs) are designed to closely mimic biological neurons, thereby enhancing temporal processing capabilities. While SNNs have shown promise in electroencephalography (EEG) cognitive recognition tasks, existing studies focus on the classification performance of SNN models, ignoring the unique characteristics of EEG signals with spatial–spectral–temporal (SST) multidimensional features and non-stationarity. Consequently, enhancing the robustness and generalization ability of SNN models in EEG cognitive recognition remains challenging. In this study, we propose EEGSNet, a novel SNN-based model designed to enhance the recognition capability for different EEG cognitive tasks. Initially, we transform EEG data into a 4D representation to explore SST-based global features, and introduce a dilated convolution structure with an alternating strategy to capture fine-grained features in the spatial domain. Additionally, we introduce a novel sample-based adaptive thresholding strategy, integrated with the parametric Leaky-Integrate-and-Fire (PLIF) module, to improve the generalization ability of the SNN. We conduct comprehensive evaluations comparing the inter- and cross-subject cognitive recognition performance of EEGSNet with current representative methods on three EEG datasets (1 public and 2 self-collected). EEGSNet achieves a maximum accuracy of 99.03% on the 5-fold cross-validation inter-subject task and demonstrates a maximum accuracy improvement of 5.49% on the leave-one-out cross-validation cross-subject task. Experimental results demonstrate that EEGSNet achieves state-of-the-art performance and generalization ability, which can further promote the development of brain-computer interfaces (BCI) applications. •EEGSNet, a novel EEG cognitive recognition model, enhances recognition for EEG tasks.•A dilated convolution with an alternating strategy extracts fine-grained features.•An adaptive thresholding strategy is introduced to improve generalization in EEG.•EEGSNet excels in performance and generalization across EEG datasets and tasks.