Motor imagery (MI) EEG classification, a core BCI task, faces challenges due to EEG's low signal-to-noise ratio and non-stationarity. Traditional supervised learning methods perform poorly in cross-subject and small-sample scenarios, limiting practical use. We propose CMHA-Net, a MI-EEG-optimized CNN integrating depthwise separable convolution, deep convolution and multi-head attention, combined with a Meta-SGD-based meta-transfer learning framework. Experiments on BCI-IV-2a and High Gamma datasets show 81.61% and 88.15% accuracy, outperforming existing models by 4-15% and excelling in small-sample cases, advancing clinical and real-world BCI applications.
Li et al. (2026) studied this question.