BCIwiki (bciwiki.com) — A Tianjin University team has developed a graph neural network-based co-optimization framework that automatically selects a small set of key EEG channels for motor imagery decoding while maintaining classification accuracy comparable to using all channels. The study was published in Chaos on August 24, 2026, by Li H, Dang W, Liu L, Du P, Cui X, and Hao Y.
The framework consists of two core components: 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 UniEEG-Net, which decodes motor imagery tasks from the selected channels using multi-scale temporal convolutions, depthwise separable spatial projection, and a self-attention mechanism. The method was validated on three datasets, including BCI Competition IV 2a, High Gamma, and a newly collected dataset. Results showed that the approach achieved performance comparable to using all channels while using significantly fewer electrodes, and UniEEG-Net's classification accuracy surpassed current state-of-the-art models.
The team said the entire system is well-suited for real-world BCI applications, particularly in neurorehabilitation.