SSM-Transformer-Based Dual-Scale Convolutional Neural Network for High-Performance EEG Decoding

Electroencephalogram (EEG) decoding remains a critical challenge in brain-computer interfaces (BCIs) due to the high-dimensional spatial complexity, task-dependent spectral variations, and non-stationary temporal dynamics of EEG signals. To address these challenges, we propose the SSM-Transformer-based Dual-Scale Convolutional Neural Network (STDCNN), a novel hybrid architecture that integrates a dual-scale convolutional module, a state-space model (SSM), and a Transformer module. The dual-scale convolutional module effectively captures spatial and spectral features at multiple scales, providing a rich representation of EEG signals. The SSM models dynamic temporal variations with linear-time complexity, offering efficient sequence processing. To compensate for the SSMs limited ability to capture long-range dependencies, the Transformer module applies selfattention to explicitly model global temporal relationships, thereby further enhancing the quality of temporal feature representations. Comprehensive evaluations across four EEG datasets - BCI-IV 2a and 2b (motor imagery tasks), SEED (emotion recognition task), and a collected dataset (motor intention task) - demonstrate STDCNN’s superior performance. STDCNN achieves state-of-the-art classification accuracy of 86.73% and 91.02% on the BCI-IV2a and 2b datasets, respectively, 98.80% on the SEED dataset, and 92.06% on the collected dataset, significantly outperforming existing models. These results highlight STDCNN’s potential robust and scalable solution for highperformance EEG decoding in diverse BCI applications.