Prototype-Debiased Latent Alignment for Class-Imbalanced EEG Decoding

Abstract

Subject-independent EEG decoding enables practical brain-computer interfaces by eliminating per-user calibration. Latent Alignment (LA) achieves state-of-the-art performance by standardizing latent features using unlabeled context sets from target subjects. However, LA's robustness to class-imbalanced context sets---common in real-world deployment---remains untested. We discover that LA's accuracy degrades substantially under imbalance: on PhysioNet Sleep, weighted accuracy drops 9.7 percentage points when context compositions follow realistic Dirichlet distributions. We identify the mechanism as \emph{prototype-mixture mean shift}: the context mean becomes biased toward majority-class prototypes, causing systematic misalignment. We propose Prototype-Debiased Latent Alignment (PD-LA), which corrects this bias using class prototypes and estimated class proportions. With oracle (ground-truth) priors, PD-LA recovers 76% of the accuracy gap (+7.4pp) and improves extreme-imbalance accuracy by +20.9pp. Correlation analysis confirms the mechanism: the shift-accuracy correlation drops from $r$=0.52 to $r$=-0.14 after correction. The practical variant shows limited effectiveness (+1.6pp) because LA's normalization makes predictions invariant to context composition, identifying prior estimation as the key bottleneck for future work.