BCIwiki (bciwiki.com) — Researchers have proposed SPDAlign, an interpretable framework that corrects distribution shifts in EEG-based brain-computer interfaces (BCIs) across subjects and sessions without labeled calibration data, in a preprint posted to arXiv on October 5, 2026 (arXiv:2610.06315). The paper has not been peer reviewed.
The non-stationary nature of EEG means a model trained on one user or session often meets a different data distribution on the next, limiting the practical use of BCIs in rehabilitation and communication. Using a classic generative model of EEG, the authors studied shifts introduced by the domain-specific forward process, which is associated with factors such as head geometry. They show theoretically that such shifts can be recovered solely through linear transformations on the Symmetric Positive Definite (SPD) manifold.
Building on that result, SPDAlign first aligns domain-specific means, then corrects global rotations across domains using an optimal transport technique called Wasserstein Procrustes. The authors evaluated the approach through simulations and report competitive performance on extensive public EEG datasets. The abstract does not give specific accuracy figures.
Because SPDAlign is globally linear, the authors say it is intrinsically interpretable: it can identify frequency ranges of interest, determine spatial patterns reflecting source-sensor relationships and address cross-subject variability.