Accurate decoding of motor imagery electroencephalogram (MIEEG) signals remains a challenge in neuroscience research and clinical applications. The non-stationary nature and low signal-to-noise ratio of MI-EEG signals often result in low decoding accuracy and poor classification outcomes. Although convolutional neural network-based methods excel at capturing local features, they struggle to capture long-term dependencies essential for effective EEG decoding. To address these challenges, this study proposes a novel end-toend neural network model, EEG-Times-ECANet. Firstly, we introduce TimesNet, which extracts the global correlation of EEG data based on its periodic changes. Subsequently, a convolutional neural network (CNN) is constructed to learn timefrequency features of EEG data. Then, the Efficient Channel Attention (ECA) mechanism module is seamlessly integrated into the model post the TimesNet and CNN combination. This integration enhances the internal relationship between channels, facilitating more comprehensive feature extraction. Finally, the extracted features are fed into the classifier to obtain the classification results. This study employs a within-subject cross-trial group training strategy to validate classification results. The experimental results demonstrate that the proposed model achieved average accuracy rates of 80.45% on the BCI Competition IV 2a dataset, 86.76% on the BCI Competition IV 2b dataset and 78.7% on the BCI Competition III 4a dataset, outperforming existing methods. The ablation experiments and feature visualization methods were performed to verify the modules’ role within the model. The results showed that the proposed EEGTimes-ECANet method effectively decodes EEG global features , extracts channel correlation features, and is suitable for EEG signal feature extraction.
Jiang et al. (Wed,) studied this question.