Fusion of Multiscale Features via Centralized Sparse-attention Network for EEG Motor Imagery Classification

xiangrui cai, Shaocheng Ma, Lei Cao, Yilin Dong

Keywords: Brain Computer Interface;Electroencephalography;Motor Imagery;Sparse-attention;Multi-branch;Feature Fusion;

Abstract: Motor imagery (MI) is an important research direction in brain-computer interfaces (BCIs) and has shown broad application value in motor rehabilitation. In recent years, a number of approaches have leveraged multiscale temporal convolution modules to capture the temporal dynamics of MI data, followed by a unified spatial module to perform spatial feature modeling. However, this design implicitly assumes that all temporal scales share the same spatial structure, overlooking the inherent spatiotemporal heterogeneity of EEG signals. To address this limitation, we design a multi-branch parallel architecture, where each temporal scale is equipped with its own spatial feature extraction module. This design mitigates the risk of spatial information confusion or loss arising from shared weights, while enhancing the flexibility and discriminative capacity of feature representations. Furthermore, to tackle the challenge of multi-branch feature fusion, we introduce the Fusion of Multiscale Features via Centralized Sparse-attention Network (EEG-CSANet). Specifically, EEG-CSANet adopts a main–auxiliary collaborative fusion architecture: the main branch leverages multiscale multi-head self-attention to model core spatiotemporal patterns, while the auxiliary branch employs multiscale sparse cross-attention to achieve efficient local interactions with the main branch. Experimental results demonstrate that EEG-CSANet achieves state-of-the-art (SOTA) performance across three public MI datasets. In particular, it significantly outperforms all compared SOTA methods on the BCI Competition IV 2a and 2b datasets, and also achieves the best results in subject-independent experiments on the 2a dataset. The related code is publicly available at: https://anonymous.4open.science/r/test-tj654478-EB7B

Primary Area: applications to neuroscience & cognitive science

Submission Number: 9447