Band power did not predict later free recall in 21 RepFR1 patients

We tested whether 0–0.75 s post-word SEEG band power predicted later free recall in RepFR1. We applied the prior study’s features and decoder to 16 new patients and combined them with 5 archived patients (21 total; 5,616 word presentations). A preregistered primary endpoint, held-out AUC over all presentations after serial position was removed, had mean 0.498 (95% CI 0.483–0.513; one-sided sign-flip p = 0.58). No secondary endpoint was significant after Holm correction (p = 0.53 for each): first presentation of 1x words had mean AUC 0.537, first presentation of 3x words 0.512, and the 3x first-minus-third AUC difference was +0.022. The new and archived patients gave similar results. Serial position alone had mean AUC 0.530, higher than the SEEG decoder. A list-filter amendment was made before final analysis after smoke tests revealed artifact AUCs of 0.90 and 0.93 when lists without matched recall annotations were retained; without the filter, the 3x-first result was mean AUC 0.549 (p = 0.047), which was not accepted as evidence. The findings are limited to one session and lists 0–9 per patient, one fixed feature set and decoder, no electrode localization or region selection, and offline analysis. Data-derived feature and event caches are omitted from the public package because provenance identifies an unversioned live mirror and does not establish the exact dataset release; source URLs and hashes remain in the archived provenance, and the underlying dataset is linked in the main text.

SEEG · free recall · memory · subsequent memory · band power · shrinkage LDA · AUC · memory BCI