BCIwiki (bciwiki.com) — EEG signals drift with wear time, electrode condition and the user's state, so a brain-computer interface model calibrated in the lab often degrades quickly in everyday use. A preprint uploaded to arXiv on October 5, 2026 benchmarks online adaptation: across 4 public datasets (3 motor-imagery and 1 movement-decoding), letting models update after the true label is revealed improved 13 of 14 model/dataset pairs on the 2 largest streams, with relative accuracy gains of up to about 18% over a frozen model. The preprint has not been peer reviewed.
The benchmark compares two widely used pipeline families, Common Spatial Patterns (CSP) and Riemannian covariance-based methods, under time-ordered prequential evaluation, in which a model predicts the current sample before updating on the revealed label. The authors also examine 3 questions: which pipelines benefit most from label-revealed updates, whether controlled forgetting of older data improves robustness, and how a minimal-calibration cold start compares with starting from a pretrained model. A Shapley-based data valuation across temporal blocks assigned the largest mean value to the most recent block in each of the 3 datasets analyzed, while older blocks retained positive value.