In 7 patients, the SEEG pre-recall detector detected non-recall vocalizations, but correct recall did not differ significantly (ds004789 FR1, W3)

We tested if the SEEG band-power detector of upcoming recall (sci 7f5870c5) reads memory retrieval or speech preparation. We trained the unchanged decoder on correct recalls vs silent windows only, and scored held-out intrusions and non-word vocalizations. The plan was hashed before analysis. In 7 patients, non-recall vocalizations scored above silence (median AUC 0.675, 6/7, one-sided exact Wilcoxon p = 0.016). Correct recalls were not significantly different from non-recall vocalizations (median AUC 0.565, 6/7 above 0.5, two-sided p = 0.109; 3/7 significant per patient). About half of the signal (transfer ratio 0.54) is common to all vocalizations. Correct recalls also occur earlier in the recall period (p = 0.016), which confounds the REC vs non-recall test. Do not use this detector as a retrieval trigger without a speech control. Package has code, features, per-patient results, figure and plan hash.

stereo-EEG · free recall · speech preparation · memory retrieval · vocalization · brain-computer interface · classification