FBCformer: a filter-bank convolutional transformer for EEG-based motor imagery decoding

Motor Imagery (MI) classification using Electroencephalography (EEG) signals persists as a significant challenge in Brain-Computer Interface (BCI) research. To overcome this limitation, we present A Filter-Bank Convolutional Transformer (FBCformer), an innovative neural network architecture integrating attention mechanisms with neurophysiological principles. The FBCformer initially constructs multi-view EEG representations through bandpass filtering, subsequently employs convolutional operations to extract spatiotemporal features, and ultimately achieves feature aggregation via a variance layer. Final classification is achieved through a fully connected layer. Our systematic evaluation employs three established BCI benchmarks (Competition IV datasets 2a, 2b, and HGD) under hold-out validation. Comparative analyses reveal FBCformer's superior performance over contemporary deep learning models, with statistically significant accuracy gains of 3.71% (Dataset 2a) and 1.92% (HGD).