List-blocked folds retained above-chance word-recall decoding in 19 FR1 patients, with little change in total AUC

We tested whether random trial folds inflate SEEG word-recall decoding by allowing words from the same list into training and test sets. Using archived band-power features from 19 patients in OpenNeuro ds004789 FR1, session 0, we compared one fixed ridge decoder under random stratified word folds, contiguous list-blocked folds, and training on the first half of lists with testing on the second half. Mean AUCs were 0.646, 0.634, and 0.625. Random-fold AUC was not significantly higher than list-blocked or chronological AUC (differences +0.011 and +0.021; Holm p = 0.40). With random folds, between-list pairs ranked better than within-list pairs (+0.034 in 18/19 patients; Holm p = 0.0006), while this gap was absent with list-blocked folds (0.634 vs 0.632). A list recall-rate baseline did not predict recall (AUC 0.509). Within-list AUC with blocked folds was above chance in all 19 patients (mean 0.632). In this offline, single-session test, most decoder signal was item-level encoding within lists, while random folds allowed some list-level information but had a small effect on total AUC. The study used one linear decoder and penalty; the random-minus-blocked test had low power for effects below about 0.02 AUC. The chronological split used 50% less training data, so its lower AUC mixes drift and training-set size.

SEEG · word recall · memory decoding · fold design · cross-validation · ridge classifier · AUC · list-level state · item encoding