A Band-Aware Riemannian Network with Domain Adaptation for Motor Imagery EEG Signal Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery electroencephalography (MI-EEG) decoding faces challenges like low signal-to-noise ratio (SNR) and domain shift across sessions and subjects.
• These issues hinder the practical application of brain-computer interfaces (BCIs).

Purpose of the Study:

• To propose an end-to-end MI-EEG decoding method, BARN-DA, addressing SNR and domain shift.
• To enhance feature extraction and reduce cross-domain discrepancies for improved BCI performance.

Main Methods:

• Developed BARN-DA featuring Band-Aware Channel Attention (BACA) and Multi-Scale Kernel Perception (MSKP) modules.
• Employed Riemannian manifold mapping with Riemannian Maximum Mean Discrepancy (R-MMD) loss for feature alignment.
• Utilized log-Euclidean metric for embedding Symmetric Positive Definite (SPD) matrices into Reproducing Kernel Hilbert Space (RKHS).

Main Results:

• BARN-DA achieved high average cross-session accuracies: 84.65% (BCIC IV 2a), 89.19% (BCIC IV 2b), 61.76% (SHU).
• Achieved high average cross-subject accuracies: 65.49% (BCIC IV 2a), 78.78% (BCIC IV 2b), 78.14% (BCIC III 4a).
• Demonstrated superior classification accuracy and generalization compared to state-of-the-art methods.

Conclusions:

• BARN-DA effectively mitigates low SNR and domain shift issues in MI-EEG decoding.
• Provides a robust technical solution for practical brain-computer interface systems.