New Method Corrects EEG Data Misalignment in Brain-Computer Interfaces.

Brain-Computer Interfaces (BCIs) often struggle with varying data distributions across users or sessions, a problem typically addressed by unsupervised domain adaptation (UDA). Existing methods, like Riemannian alignment, assume balanced data classes, which is often not the case in real-world BCI applications where label distributions can shift dynamically. This class imbalance leads to geometric misalignment of target data. A new framework, OSPDIM (Online SPD manifold information maximization), is proposed to tackle this issue. It's a source-free online UDA method specifically designed for label shifts on the Riemannian manifold. OSPDIM introduces a bias parameter into the tangent space mapping, which is optimized through information maximization to correct the geometric skew caused by imbalanced data streams. Unlike traditional offline methods that rely on global batch statistics, OSPDIM estimates and corrects this geometric bias on-the-fly. Evaluations on motor imagery datasets demonstrate that OSPDIM significantly outperforms standard Riemannian baselines, especially in challenging online adaptation scenarios with severe class imbalance, making it a robust solution for practical BCI systems.