RAP²G: Relation-Aware Progressive Pseudo-label Generation for Cross-subject MI-EEG Recognition

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) rely on motor imagery electroencephalography (MI-EEG) classification.
• Inter-subject variability poses a significant challenge for MI-EEG accuracy.
• Unsupervised domain adaptation (UDA) methods, using pseudo-labeling, attempt to address this but have limitations.

Purpose of the Study:

• To overcome limitations in existing pseudo-labeling techniques for cross-subject MI-EEG classification.
• To develop a novel UDA framework for robust MI-EEG classification across diverse individuals.
• To enhance the reliability and practicality of BCI systems.

Main Methods:

• Proposed the relation-aware progressive pseudo-label generation (RAP2G) method, a novel UDA framework.
• Combined Optimal Transport (OT) with structure-aware regularization and dynamic pseudo-label selection.
• Leveraged feature similarity and OT confidence for adaptive pseudo-label generation and selection.

Main Results:

• RAP2G demonstrated superior performance compared to state-of-the-art UDA techniques and baselines.
• Ablation studies validated the effectiveness of the structure-aware component.
• Visualizations revealed improved feature separability and attention maps consistent with motor cortex organization.

Conclusions:

• RAP2G offers an effective solution for robust cross-subject MI-EEG classification.
• The method enhances label-free adaptation, supporting more reliable BCI systems for biomedical applications.