Few-trial recentering did not improve AUC in cross-day SEEG load decoding

We tested whether a working-memory load decoder trained on one day could be recentered with the first k unlabeled trials from a different day. We used late-maintenance (4–6 s) log band-power features from the verbal Sternberg task in OpenNeuro ds004752 v1.0.1 and an archived feature contribution, with 10 subjects and 203 cross-day session pairs. A Ledoit–Wolf shrinkage LDA distinguished set size 8 from set size 4. The primary metric was AUC on trials at position 16 or later. No recentering gave AUC 0.682; using the first 8 unlabeled trials to estimate both mean and SD gave 0.681 (difference −0.001, 95% subject-bootstrap CI −0.008 to 0.006; p = 0.84, Holm p = 1.0). Full-session normalization, an offline oracle, gave 0.690. Mean-only recentering leaves AUC unchanged by construction for a fixed linear decoder. Post-hoc balanced accuracy increased from 0.581 without recentering to 0.607 with 8 trials and 0.618 with 16 trials, but no gain remained significant after Holm correction. Estimating SD from 2 trials reduced AUC to 0.552. These offline results cover one task, feature window, decoder, and 10 subjects; they provide no evidence that few-trial recentering improves AUC in this setting.

SEEG · cross-day decoding · working memory · brain-computer interface · feature recentering · Ledoit-Wolf LDA · AUC · unlabeled calibration