A short-window SEEG band-power decoder did not exceed chance during or outside 50 Hz encoding stimulation

We tested whether a word-recall decoder trained on control-list SEEG band power could distinguish later-recalled from not-recalled words presented during 50 Hz encoding stimulation in RAM FR2 sessions from OpenNeuro ds005489. In 12 eligible sessions from 12 patients, the pre-specified decoder had a mean held-out control-word AUC of 0.46 (one-sided sign-flip p=0.82 for AUC > 0.5), so decoder validity failed before stimulation was considered. For stimulated words, mean AUC was 0.525 (8/12 sessions above 0.5; exact one-sided sign-flip p=0.143), and the stimulated-minus-control AUC difference was +0.061 (two-sided p=0.21). Using contacts at least 20 mm from the anode gave stimulated-word AUC 0.516 (p=0.24). An exploratory decoder trained on all non-stimulated words had control and stimulated AUCs of 0.509 (p=0.38) and 0.527 (p=0.15), respectively; their difference was +0.017 (p=0.69). After removing +/-3 Hz around 50 Hz harmonics, stimulated-word features shifted more at contacts less than 20 mm from the anode in all six bands, with significant near-versus-far differences in 3-7, 13-30, and 103-147 Hz (two-sided sign-flip tests over 15 sessions). At far contacts, shifts did not exceed those for non-stimulated words on the same stimulated lists (all p >= 0.12). These results do not measure a read-while-write cost: this per-session decoder did not read encoding success with 30-60 training words even without stimulation. The findings are limited to this dataset and feature set; a null result does not show that SEEG cannot decode encoding success during stimulation. Harmonic removal does not remove broadband artifacts or amplifier recovery effects, and near-anode sets contained 4-18 contacts per session.

SEEG · free recall · memory decoding · brain-computer interface · electrical stimulation · band power · 50 Hz stimulation