Graph convolution neural network channel selection with attention for motor imagery EEG decoding.
TL;DR
Experimental results demonstrate that the proposed approach achieves performance comparable to that obtained with all channels while using significantly fewer electrodes, and UniEEG-Net exhibits classification accuracy surpassing current state-of-the-art models.
Abstract
Accurate decoding of motor imagery electroencephalography (MI-EEG) signals is critical for practical brain-computer interface (BCI) systems. However, conventional approaches typically rely on dense multi-channel recordings, which not only introduce data redundancy but may also incorporate noise, thereby hindering real-world deployment. To address this challenge, we propose a graph neural network-based co-optimization framework that simultaneously performs channel selection and MI classification. The framework comprises two core components: one is the Key Channel Locator (KCL), which models EEG electrodes as graph nodes and identifies a subject-specific, fixed-size subset of informative channels through a dual-perspective evaluation that integrates graph convolutional topology with self-attention-derived feature importance, and the other is the UniEEG-Net, which efficiently decodes MI tasks from the selected channels using multi-scale temporal convolutions, depthwise separable spatial projection, and a self-attention mechanism. We extensively validate the proposed method on three datasets, including BCI Competition IV 2a, High Gamma, and a newly collected dataset. Experimental results demonstrate that our approach achieves performance comparable to that obtained with all channels while using significantly fewer electrodes. Moreover, UniEEG-Net exhibits classification accuracy surpassing current state-of-the-art models. The entire system is thus well-suited for real-world BCI applications, particularly in neurorehabilitation.