Multi-Scale convolutional neural networks integrated with self-attention for motor imagery EEG decoding

Brain-computer interface (BCI), as a cutting-edge technology with great application prospects, has received widespread attention in recent years. Motor imagery (MI) electroencephalography (EEG) classification is a key component of brain-computer interfaces, widely used in applications such as assisting people with disabilities, controlling devices, and interacting with environments. However, since convolutional neural networks (CNNs) extract only local temporal features, they may be unable to capture the long-term dependencies used for EEG decoding, which can have an impact on the decoding performance. In order to address this problem, this paper proposes a novel deep learning network that combines a multi-scale convolutional neural network with an attention mechanism to capture temporal information and global dependencies. First, a multi-scale structure is designed to extract spatial-temporal information at different scales and multimodal information from both the mean and variance perspectives. Second, a squeeze-excite-compress (SEC) module is used to enhance the feature response of each branch and reduce information redundancy. Finally, an encoder with a multi-head attention mechanism extracts more discriminative features and highlights the most valuable information in MI-EEG data. In addition, this paper uses a data augmentation method of signal reorganization to expand the dataset and further enhance the generalization ability of the network. Our method was evaluated by performing experiments on the BCI Competition IV-2a (BCI-IV-2a) and High Gamma Dataset (HGD) with classification accuracies of 85.26% and 95.86%, respectively. The experimental results show that our method achieves state-of-the-art performance and has great potential to be a new baseline for general EEG decoding.