Feature reweighting for EEG-based motor imagery classification

Classifying motor imagery (MI) using non-invasive electroencephalographic (EEG) signals is crucial for predicting the intention of limb movements. Convolutional neural networks (CNNs) have been widely adopted for MI-EEG classification, but challenges such as low signal-to-noise ratio, non-stationarity, and non-linearity of EEG signals, along with the presence of irrelevant information in feature maps, can degrade performance. This work proposes a novel feature reweighting approach to address these issues by introducing a feature reweighting module that suppresses irrelevant temporal and channel features. The module generates relevance scores to reweight feature maps, thereby reducing the influence of noise and irrelevant data. Experimental results demonstrate significant improvements in MI-EEG classification on the Physionet motor imagery and BCI Competition IV-2a datasets, achieving performance gains of 9.34% and 3.82%, respectively, over state-of-the-art CNN-based methods. Furthermore, the proposed method showed competitive or superior performance on both speech imagery and motor movement tasks, highlighting its generalizability and robustness.