In one session from 15 SEEG patients, averaging three same-load trials raised decoder AUC from 0.690 to 0.751, consistent with independent errors

We tested whether averaging a memory-load decoder's output over same-load trials increases discrimination and whether errors from trials close in time reduce that gain. We used archived maintenance-window band-power features for 15 patients performing a verbal Sternberg task in OpenNeuro ds004752, session 1, run 1. A shrinkage linear discriminant analysis decoder produced leave-one-trial-out scores for set size 8 versus 4. We compared averages of 1–5 same-class scores in session order and after within-class random permutations (200 permutations) with the independent-error prediction AUC_k = Phi(sqrt(k) Phi^-1(AUC_1)). Mean single-trial AUC was 0.690 (11/15 patients above 0.5; sign-flip p = 0.0035). Random groups of three raised mean AUC to 0.751 (+0.061; 13/15 increased; p = 0.0074), within 0.008 of the independent-error prediction (p = 0.36). Consecutive groups were not worse than random groups: the difference at k = 3 was +0.011 (p = 0.60). Lag-1 residual autocorrelation was not greater than the permutation null (mean -0.070; group p = 0.46). Within this session, averaging gain was consistent with independent errors; averaging did not rescue patients near chance. This offline analysis included one session and run per patient, 9–29 trials per class, overlapping windows, and 15 patients. It did not test drift across days, and small drift effects cannot be excluded.

intracranial EEG · SEEG · brain-computer interface · working memory · trial averaging · area under the curve · decoder · error autocorrelation