EEG-DBNet: A Dual-Branch Network for Temporal-Spectral Decoding in Motor-Imagery Brain-Computer Interfaces

Objective: Non-invasive brain-computer interfaces (BCIs) based on motor imagery electroencephalography (EEG) offer practical assistance for patients with impaired limb mobility. However, the inherently low signal-to-noise ratio and poor spatial resolution of EEG pose significant challenges to effective feature extraction. This study aims to develop an end-to-end deep learning model that captures EEG representations across multiple dimensions and scales to enhance decoding performance. Methods: We propose a novel dual-branch network, EEG-DBNet, which simultaneously models the temporal and spectral patterns of EEG signals. Each branch comprises local and global convolution modules: the local convolution block maps EEG from the time-spatial domain to the time-spectral domain to extract localized features, while the global convolution block leverages causal dilated convolutions and attention mechanisms to capture global dependencies. The outputs from both branches are concatenated and passed through fully connected layers for final classification. Results: EEG-DBNet achieves a classification accuracy of \(85.87\%\) on the BCI Competition IV-2a dataset, outperforming state-of-the-art methods. The source code is publicly available at: https://github.com/xicheng105/EEG-DBNet . Conclusion: Incorporating spectral feature modeling significantly enhances the decoding capability of EEG-based BCIs. Moreover, extracting global temporal-spectral dependencies compensates for the local receptive field limitations of conventional convolutional neural networks. Significance: The proposed EEG-DBNet provides a new architectural paradigm for motor imagery EEG decoding, offering improved performance and generalizability for non-invasive BCI applications.