SEEG error trials did not significantly lower working-memory load-decoder scores in 14 patients

We asked whether a SEEG working-memory load decoder reflects items held or items presented. We analyzed archived maintenance-period log band-power features from 15 patients performing a verbal Sternberg task (set sizes 4, 6, and 8) in OpenNeuro ds004752. A shrinkage LDA decoder was trained only on correct size-4 and size-8 trials, using stratified five-fold cross-validation repeated 10 times. Size-6 and error trials were never used for training. For each error trial, we standardized its score against correct trials of the same set size. The prespecified primary test across 14 patients with size-6 or size-8 errors (56 trials) found mean d = -0.31 (median -0.11; 9 of 14 negative; exact two-sided sign-flip p = 0.17), so the test did not reject the null hypothesis. Among 11 patients with decoder AUC > 0.55, mean d was -0.50 (8 of 11 negative; p = 0.037, uncorrected). A broadband control showed little change (mean d = -0.08, p = 0.62). A post hoc association between decoder AUC and error-trial d was negative (Spearman rho = -0.69, permutation p = 0.007; 20,000 permutations). Correct-trial scores rose with set size and showed little correlation with reaction time within set size. The primary result is inconclusive. Secondary and post hoc findings are not corrected for multiple tests. With only 1–9 errors per patient, one dataset, session, feature set, and decoder, a lower score cannot distinguish fewer items held from reduced attention or response errors.

SEEG · working memory · Sternberg task · load decoding · error trials · shrinkage LDA · intracranial EEG · OpenNeuro ds004752