Cross-Session EEG Motor Imagery Classification Method Based on Transfer Learning

Motor imagery (MI) electroencephalogram (EEG) signals exhibit cross-session variability, which hinders traditional EEG classification models from achieving satisfactory performance in cross-session EEG data classification. Therefore, further adjustments and optimizations are needed. To resolve this problem, the research develops a transfer learning-based approach for cross-session motor imagery EEG classification. To comprehensively analyze EEG data, this study introduces an attention-optimized separable CNN framework for multi-domain feature extraction (spectral, spatial, and temporal). In parallel, supervised learning is applied to reduce the classification error of source domain data. Meanwhile, transfer learning is leveraged to enhance the alignment of feature representations between source and target domains, enabling the training of a neural network capable of effectively processing cross-session EEG datasets. Finally, the proposed algorithm is compared with several selected advanced algorithms and evaluated on the publicly available 2008 BCI Competition IV 2a dataset. The results show that without any manual feature extraction, the proposed cross-session EEG motor imagery classification method based on transfer learning achieves an average accuracy of 73.1% and a Kappa value of 64.8% for the four-class classification dataset.