Spectral SHAPE features do not improve cross-day SEEG working-memory load decoding

We tested whether spectral SHAPE features (log band power minus the per-channel mean across bands) transfer better across days than FULL band power in stereo-EEG working-memory decoding. The analysis used archived features from OpenNeuro ds004752 v1.0.1: 10 subjects, 35 target sessions on later days, six late-maintenance bands (4–6 s), and set size 8 versus 4. Shrinkage LDA was trained on all earlier-day sessions. Without new-day data, SHAPE minus FULL AUC was +0.010 (95% bootstrap CI −0.027 to 0.061; 5/10 subjects positive; exact sign-flip p=0.40), so the primary hypothesis was not supported. With label-free per-session z-scoring, SHAPE minus FULL was −0.017 (Holm-adjusted p=0.61); the cross-day loss difference was also not reliable (Holm-adjusted p=0.61). Between-day mean shift was large (RMS 1.6 within-session SD per feature), and 71% was a per-channel broadband level shift on average (range 0.37–0.82; isotropic reference 0.17). SHAPE removes this shift but also removes level information, and its net AUC effect was near zero. One subject (sub-01; shift 5.6 SD) drove most of the positive primary mean. Exploratorily, label-free per-session z-scoring of FULL gave the highest cross-day AUC (0.709 versus 0.675 without it; 8/10 subjects; uncorrected p=0.012). These results apply to one Sternberg task, one time window, and 10 subjects; effects smaller than 0.03 AUC cannot be excluded. The result favors per-session recentering of FULL features when unlabeled new-day trials are available, for this task and dataset.

stereo-EEG · working memory · cross-day decoding · brain-computer interfaces · band power · feature normalization · shrinkage LDA