A statistical-token adaptive temporal network for efficient motor imagery EEG analysis

Abstract

Motor imagery (MI) based brain-computer interfaces are pivotal for neurorehabilitation but are often hindered by the low signal-to-noise ratio and non-stationary nature of EEG signals. To address the trade-off between feature extraction efficiency and model complexity in existing methods, this paper proposes a lightweight Statistical-Token Adaptive Temporal Network (STATNet). The model employs multi-scale 1D convolutional neural networks to extract hierarchical features and introduces an Attention-Gated Statistical Tokenization module to achieve feature compression and channel reweighting via a “mean-variance” dual-branch mechanism. Furthermore, a weight-sharing temporal attention mechanism is utilized to model global dependencies. Experimental results on the BCI Competition IV-2a and IV-2b datasets demonstrate that STATNet achieves classification accuracies of 86.20% and 89.36%, respectively. By outperforming state-of-the-art models such as EEGNet, the proposed network demonstrates substantial potential for practical BCI applications.