Classification of motor imagery EEG signals using CNN–Transformer hybrid models

Motor imagery-based brain–computer interfaces aim to translate electroencephalographic activity into control commands without requiring overt movement. However, strong inter-subject variability, exacerbated by the limited spatial resolution of compact electrode montages, makes subject-independent decoding particularly challenging. This work presents a compact hybrid convolutional neural network–Transformer architecture designed for stable binary left/right motor imagery decoding using a low-density eight-channel sensorimotor montage. The model integrates convolutional feature extraction with a parameter-efficient temporal self-attention mechanism to capture both localized spatiotemporal patterns and long-range temporal dependencies in the electroencephalographic signal. Under strict subject-wise five-fold cross-validation on the PhysioNet EEG Motor Movement/Imagery dataset, the proposed convolutional neural network–Transformer achieved a mean macro F1-score of 0.816. Moreover, the statistical analysis revealed no significant performance difference ( p > 0.05 ) relative to larger architectural variants, thus justifying the selection of the proposed compact model for its computational efficiency and scalability. With only 0.232 million parameters and approximately 0.034 giga-floating-point operations per forward pass, the proposed convolutional neural network–Transformer constitutes a compact and resource-efficient backbone that supports low-latency subject-independent motor imagery decoding under low-density EEG.