Per-session covariance whitening did not improve cross-day SEEG load decoding over per-session z-scoring

We tested whether unsupervised Euclidean alignment (EA), which whitens each session with its own shrinkage covariance, improves cross-day SEEG decoding over per-session z-scoring (ZS). We used archived log10 band-power features from the 4–6 s maintenance window of the verbal Sternberg task in OpenNeuro ds004752 v1.0.1: 10 subjects, 56 sessions, and 35 later-day targets. A Ledoit-Wolf shrinkage LDA decoded set size 8 versus 4. With all earlier days pooled for training, EA exceeded ZS by a mean 0.014 AUC, but the difference was inconsistent (6/10 subjects; exact two-sided sign-flip p=0.43). Both ZS and EA exceeded training-set z-scoring (NONE) after Holm correction (mean differences +0.031 and +0.045 AUC; p=0.039 for each). Using only the latest earlier session, EA minus ZS was −0.003 AUC (p=0.75). The session covariance required substantial shrinkage (median 0.38; range 0.15–0.67), with about 30–50 clean trials and 180–354 features per session. The test had low power: a true EA gain of about 0.01–0.02 AUC is not excluded. Findings apply to this task, time window, and band-power features; more trials or fewer features could change the result and were not tested.

SEEG · cross-day decoding · Euclidean alignment · covariance whitening · z-scoring · working memory · band power · shrinkage LDA · OpenNeuro ds004752