Motor imagery brain–computer interfaces (MI-BCIs) play a significant role in intelligent control systems. Given the advantages of non-invasive measures represented by electroencephalography (EEG) and magnetoencephalography (MEG), MI classification based on EEG and MEG data has become a focal point in the BCI field. However, the non-linear and non-stationary characteristics of EEG/MEG data result in distribution discrepancies among MI samples from different subjects, severely limiting the application of constructing online MI-BCIs.
To address this challenge, this study proposes a novel Transfer Riemannian Space-based Sparse Coding (TRSSC) method upon transfer learning theory to establish an efficient cross-subject MI classification process. The TRSSC method first computes the covariance centroid matrices of EEG/MEG samples and aligns each subject’s sample set to the identity matrix to reduce marginal distribution discrepancies. Subsequently, it extracts the aligned samples’ Riemannian tangent space (RTS) features for each subject and constructs domain-invariant sparse RTS features through minimizing maximum mean discrepancy (MMD) and sparse coding for cross-subject MI classification.
To validate the feasibility and effectiveness of the TRSSC method, comparative experiments were conducted on three representative MI-EEG/MEG datasets. The experimental results demonstrate that the TRSSC method surpasses state-of-the-art methods in classifying EEG/MEG data while maintaining a similar time complexity. Furthermore, through ablation experiments and parameters sensitivity analysis, the study confirms the completeness of the TRSSC method in domain-invariant sparse feature representation and its insensitivity to parameters.
The proposed TRSSC method offers a new option for online constructing MI-BCIs, showcasing high cross-subject classification accuracy and efficiency. Based on mobile EEG/MEG recording devices, TRSSC method has a potential in enhancing classification performance within a small computational burden.
