A session-0 SEEG recall decoder predicted session-1 recall above chance in seven patients

This offline study tested whether an individual-patient decoder trained on FR1 session-0 data predicted word recall in session 1 without session-1 labels. Seven of nine patients with recordings in both sessions met the pre-registered inclusion rule. Using common bipolar SEEG channels, within-session z-scored log10 Morlet power at eight frequencies, and the archived W_chan L2 logistic decoder, cross-session prediction had median AUC 0.622; all seven patients were above chance (exact one-sided Wilcoxon p = 0.0078; Holm-adjusted p = 0.023). Cross-session AUC was not significantly different from session-1 leave-one-list-out AUC (median difference −0.033, p = 0.22), and adding session 0 to session-1 training did not significantly change AUC (median difference +0.016, p = 0.38). These null tests do not establish equality. Cross-session AUC was below session-0 leave-one-list-out AUC in all seven patients (median difference −0.058, p = 0.016). A within-serial-position analysis was also above chance (median AUC 0.608, p = 0.016). The study includes one session pair per patient, one decoder and feature set, and offline analysis; its findings are limited to seven patients.

SEEG · word recall · memory · brain-computer interface · cross-session decoding · AUC · OpenNeuro · FR1