Manifold-based multi-branch transfer learning for MI-EEG decoding

We present a novel decoding approach for motor imagery electroencephalograms (MI-EEG) in Brain-Computer Interface (BCI) systems, aiming to advance neurorehabilitation and human–computer interaction. However, due to the non-stationary properties and low signal-to-noise ratio of MI-EEG, as well as significant individual variability, achieving accurate cross-subject decoding remains a challenge for real-world applications. To address this, we propose a manifold-based data processing method combined with a multi-branch network and enhanced by transfer learning to improve cross-subject performance. First, we identify the frequency bands relevant to motor imagery tasks and align multi-band data on the Symmetric Positive Definite (SPD) manifold using the Log-Euclidean Metric (LEM), including both full-band data and motor imagery-specific frequency bands. We then reconstruct these data from the manifold using a low-rank representation (LRP). Finally, we use a multi-branch network to extract and fuse deep features from the various frequency bands. We validated our approach on the BCI Competition IV-2a dataset and our JS-MI dataset, demonstrating that our method excels in cross-subject MI-EEG decoding tasks, with average classification accuracies of 74.55% and 71.94%, respectively. Our findings highlight the potential of this approach to improve BCI applications and facilitate more effective neurorehabilitation and human–computer interaction through enhanced MI-EEG decoding. • We propose a manifold-based MI-EEG data processing method • Our method includes a multi-branch network and transfer learning • Validated on the public BCI Competition IV-2a dataset and our JS-MI dataset • Improved MI-EEG decoding due to manifold alignment and multi-band feature interaction