A neural network with multi-type attention for enhancing motor imagery EEG decoding

Despite the widespread adoption of deep learning techniques in motor imagery (MI) electroencephalogram (EEG) decoding, the limited decoding performance persists due to the low signal-to-noise ratio of EEG signals and insufficient exploration of MI-related information from temporal, frequency and spatial domains. Therefore, this paper proposed a novel end-to-end neural network with multi-type attention (MTANet) to extract the spatiotemporal-frequency coupling features in MI EEG and enhance MI EEG decoding. In MTANet, the EEG inception attention module was established to efficiently extract and weight temporal-frequency-spatial features inherent in EEG signals and the parallel temporal sequence attention module was introduced to enhance feature representation through parallel temporal convolutional networks, each processing weighted segments of different feature sequences. The proposed MTANet model achieved average accuracies of 83.39% (±0.09), 89.47% (±0.08) and 77.06% (±0.12) on the BCI Competition IV datasets IIa and IIb, and the BCI Competition III dataset IIIa, respectively, outperforming seven state-of-art models. The experimental results demonstrated that the MTANet model, empowered by the EEG inception attention and parallel temporal sequence attention modules, effectively improved the accuracy and stability of MI decoding.