A weighted multi-scale attention-enhanced temporal convolutional network for motor imagery EEG decoding in brain-computer interfaces.

Accurate decoding of motor imagery electroencephalogram signals plays a critical role in brain-computer interfaces for neurorehabilitation and assistive technologies. However, existing multi-scale temporal methods often overlook scale-specific importance and fail to jointly capture transient and long-term neural dynamics, we propose a Weighted Multi-scale Attention-enhanced Temporal Convolutional Network (WMA-TCNet). The model employs parallel multi-scale temporal convolutions to capture neural patterns associated with distinct EEG rhythms. A global-aware scale attention mechanism adaptively weights each branch to emphasize task-relevant temporal information. A weighted Channel-Preserving Prior Path is introduced to maintain channel-wise dependencies and enhance spatial modeling stability across cortical regions. In addition, a temporal attention-guided TCN jointly captures local and long-range temporal dependencies. Experiments on BCI Competition IV 2a and 2b datasets show that WMA-TCNet achieves accuracies of 85.8% and 90.0% in subject-dependent settings, and 68.6% and 79.5% in cross-subject scenarios. These results demonstrate improved decoding performance and robustness, while providing a biologically meaningful framework for modeling multi-scale neural dynamics, with potential applications in brain-computer interfaces and neurorehabilitation.