In 10 SEEG patients, training ranked 8 channels matched all channels and beat random sets in leave-one-list-out recall decoding, but not across session halves

We tested whether patient-specific decoding of later word recall from SEEG band-power features needs all channels, and whether channel selection improves on random sets. We analyzed archived features for 10 patients in OpenNeuro ds004789, FR1 session 0, using shrinkage LDA and within-serial-position AUC. In leave-one-list-out tests, median AUC was 0.615 for the 8 training-ranked channels, 0.605 for all channels, and 0.572 for random 8-channel sets. The selected set was not significantly different from all channels (exact two-sided Wilcoxon p = 0.92) and exceeded random sets (8/10 patients; one-sided p = 0.0098). In the test trained on the first half of lists and evaluated on the second half, median AUC was 0.542 for selected channels, 0.565 for all channels, and 0.534 for random sets; selected channels did not outperform random sets (p = 0.81). The selected channels differed across patients. These findings suggest that a small patient-specific set can decode recall in leave-one-list-out tests, but do not show that a set selected early remains best later. The study includes 10 patients, one session and task, one feature set, and one classifier; no difference is not proof of equivalence, and the half-split estimates are noisy.

SEEG · word recall · channel selection · memory decoding · band power · shrinkage LDA · serial position