A Domain Adaptation Deep Learning Network for EEG-Based Motor Imagery Classification

The correlation between neighboring electroencephalography (EEG) channels reveals brain signal interconnectedness, and how to represent this correlation is being studied. Simultaneously, variations in EEG signals among individuals may present difficulties in the model’s ability to generalize across different individuals. A model may perform well on one person but not on others, limiting its reliability and generalizability in practical applications. We propose a domain adaptation-based deep learning network to address the issues above. Initially, the EEG data is transformed into a three-dimensional (3D) matrix to preserve the correlation between EEG channels, and subsequently, the spatial-temporal characteristics of the data are acquired by using the 3D convolution module. The spatial-feature map attention mechanism reinforces spatial features in the feature map, allowing the subsequent convolution module to learn spatial feature information. Finally, a domain adaptation strategy is employed for both single-source and multi-source domain scenarios. The objective of this strategy is to address the issue of variability in the EEG signal by minimizing the discrepancy between the source and target domains using a maximum mean discrepancy loss function. The proposed method was validated on two datasets, namely the BCIC IV 2a and OpenBMI datasets. We achieved an accuracy of 70.42% in an intra-subject OpenBMI experiment, which is 5.51% higher than the state-of-the-art approach. On the BCIC IV 2a dataset, we conducted intra-subject and inter-subject experiments, achieving accuracy results of 73.91% and 67.88%, respectively, which are 5.38% and 1.61% better than the state-of-the-art method.