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

Motor imagery electroencephalography (MI-EEG) classification is essential for brain-computer interfaces (BCIs), but achieving high accuracy across different individuals remains challenging due to significant inter-subject variability. Recently, unsupervised domain adaptation (UDA) methods have addressed this problem by adapting models without target labels, often using pseudo-labeling. However, existing pseudo label techniques evaluate each sample in isolation and employ a simple threshold-based strategy, overlooking the relationship among samples and often excluding useful data points. We aim to overcome these limitations for cross-subject MI-EEG classification.

We propose the relation-aware progressive pseudo-label generation (RAP2G) method, a novel UDA framework combining Optimal Transport (OT) with structure aware regularization and dynamic pseudo-label selection. RAP2G leverages the inherent structure within the target subject's data by incorporating feature similarity into the OT-based pseudo label generation process. It also adaptively selects pseudo-labeled samples using a progressive schedule based on OT confidence. We evaluated RAP2G on three public benchmarks, BCI Competition IV dataset 2a, BCI Competition IV dataset 2b, and the High Gamma Dataset, using leave-one-subject-out cross-validation.

RAP2G consistently outperforms existing state-of-the-art UDA techniques and baseline models. Ablation studies confirm the contribution of the structure-aware component. Visualizations show enhanced feature separability after adaptation, and the learned attention maps are qualitatively consistent with known motor-cortex organization.

RAP2G provides an effective approach for robust cross-subject MI-EEG classification.

By improving label-free adaptation across subjects, this work supports more reliable and practical BCI systems for biomedical applications.