Across 23 RepFR1 SEEG patients, band power weakly separated repeated from first word presentations at matched serial position, with no univariate suppression

A memory brain-computer interface (BCI) may use familiarity signals: the brain response to a repeated item differs from the response to a new item. In RepFR1 (OpenNeuro ds005411), repeats occur late in each list, so serial position alone predicts the label (AUC 0.85). We analyzed archived band-power features from sci 02694702 for 23 SEEG patients, one session each, lists 0–9, five bands, and 0–0.75 s after word onset. A shrinkage linear discriminant analysis with train-fold serial-position centering and five-fold cross-validation grouped by list produced held-out scores. We counted only repeat–first pairs at the same serial position (stratified AUC), so serial position cannot cause the result. The primary result was a mean stratified AUC of 0.542 (95% CI 0.515–0.570; one-sided sign-flip p = 0.0039; 18/23 patients above 0.5). A label-shuffle check gave 0.502. Secondary results were null after Holm correction: mean 70–150 Hz suppression d = −0.016 (p = 0.47), short-lag minus long-lag AUC = −0.008 (p = 0.66), and third-versus-second presentation AUC = 0.521 (p = 0.23). No band showed a univariate effect. The small, distributed repetition signal remains after exact serial-position control and differs from the conclusion of sci eebec63b (four patients), but is too weak for a single-trial BCI readout. The source report does not establish the exact OpenNeuro release used; the feature-file hashes and source URL are retained. The package includes code, per-patient results, input hashes, a figure, and reproduction steps.

RepFR1 · stereo-EEG · band power · repetition · serial position · stratified AUC · shrinkage LDA · memory · brain-computer interface