SEEG band power detects spoken recall before speech onset in 9 patients, beyond recall timing

In nine patients from OpenNeuro ds004789 (FR1, session 0), we tested whether SEEG band power in the final second before an isolated spoken free recall distinguishes recall events from silent periods in the same recall period. A per-patient L2 logistic decoder with leave-one-list-out cross-validation reached median AUC 0.785 in the prespecified window from 1.0 to 0.25 s before speech (8/9 patients above chance; one-sided exact Wilcoxon p = 0.006). A time-since-recall-start model had median AUC 0.633, and the band-power decoder was higher in 8/9 patients (median difference +0.158; two-sided p = 0.012). Within time terciles, median AUC was 0.725 (8/9 above chance; p = 0.064). Earlier windows had median AUCs of 0.604 and 0.643. A no-fit 70–150 Hz channel mean was weak (median AUC 0.553). The result suggests that the signal is not only the early-recall timing effect, but the window is close to speech onset and speech-motor preparation may contribute. Only 9 of 19 planned patients were analyzed, in one session and offline; controls were not matched for every confound.

stereo-EEG · free recall · memory decoding · band power · brain-computer interfaces · cross-validation