In 18 SEEG patients, two-frequency word-recall decoders were not non-inferior to an eight-frequency decoder

We tested whether two-frequency decoders could match an eight-frequency (3–140 Hz) decoder for later word recall using archived per-word features from 18 patients in OpenNeuro ds004789 v3.1.0, FR1, session 0. The preplanned analysis used L2 logistic regression and five-fold list-block cross-validation, with a non-inferiority margin of 0.02 AUC and Holm correction across three primary tests. None of the 80+140 Hz, 5+140 Hz, or nested-CV-selected two-frequency decoders was non-inferior to the eight-frequency decoder (median AUC differences −0.026, −0.041, and −0.023; all significantly worse by two-sided exact Wilcoxon tests, p ≤ 0.012). A five-frequency set (3, 5, 8, 80, 140 Hz) was non-inferior in an uncorrected secondary test, while greedy frequency selection was unstable. The findings apply to one offline session, fixed archived frequencies, and 18 patients; there was no closed-loop test, and the 0.02 AUC margin is not a clinical threshold.

SEEG · word recall · memory decoding · brain-computer interface · frequency selection · non-inferiority · cross-validation · AUC · OpenNeuro ds004789