Spreading eight SEEG pairs across shafts modestly improved word-recall decoding, but not within list position

We compared within-patient word-recall decoding from bipolar SEEG pairs sampled across many shafts with decoding from the same number of pairs concentrated on fewer shafts. The analysis used archived band-power features from OpenNeuro ds004789 v3.1.0 (FR1, 18 patients), an L2 logistic-regression decoder, five list-grouped folds, and 30 random subset draws per condition and patient. In the pre-registered primary comparison, one pair on each of eight shafts exceeded eight pairs on one shaft by a median AUC of 0.010 (95% CI 0.001 to 0.026; 14/18 patients; exact two-sided Wilcoxon p = 0.012). Three of four secondary tests supported higher decoding with broader shaft coverage; the difference at 16 pairs was not significant. The advantage was not significant when comparisons were restricted to words at the same list position (+0.007 AUC, p = 0.35), so the result does not establish improved item-memory decoding independent of list position. Random eight-pair subsets performed similarly to the eight-shaft condition (median AUC 0.607 vs 0.608). Pair-feature correlations were higher for non-contact-sharing pairs on the same shaft than for pairs on different shafts (median mean |r| 0.23 vs 0.12; 18/18 patients, p = 7.6e-6), a descriptive pattern consistent with redundancy but not a causal test. The effect was small and applies to this single offline dataset and within-patient design; shaft count and sampled brain regions were not separated.

SEEG · word recall · memory decoding · bipolar pairs · shaft sampling · AUC · intracranial EEG