Motor imagery (MI)-based brain–computer interfaces (BCIs) provide a promising non-invasive solution for motor rehabilitation and assistive control. However, the use of multichannel electroencephalography (EEG) often results in high-dimensional and noise-prone data, posing both practical and computational challenges for real-world implementation. To address these issues, we propose a few-shot channel selection framework that integrates wavelet scattering transforms (WSTs) for robust, shift-invariant feature extraction with a squeeze-and-excitation (SE) convolutional neural network (CNN) for end-to-end channel selection. Our approach identifies informative EEG channels using only a small number of labeled samples from a target individual, allowing subsequent model training with the reduced channel set. We evaluated the proposed method on two public upper-limb MI EEG datasets (SHU and Stroke; 75 participants) and an in-house lower-limb MI dataset comprising 12 healthy and 5 spinal cord injury participants. Across all datasets, the framework retained or improved classification performance using as few as three channels. Topographic analyses revealed consistent channel selection patterns in motor, parietal, and prefrontal cortical regions. These findings demonstrate the feasibility of efficient and scalable MI-BCI systems that require minimal setup, supporting broader applications across both healthy and impaired populations.
Sung et al. (Thu,) studied this question.