In 13 patients, channel-resolved SEEG recall AUC rose through 20 lists within one FR1 session

We measured within-patient learning curves for SEEG word-recall decoders using archived band-power features from 13 patients in the Penn RAM FR1 task, session 0. For each patient, five fixed test folds were used to evaluate models trained on 2 to 20 lists. Channel-resolved logistic regression AUC increased from 0.562 at 2 lists to 0.632 at 10 and 0.655 at 20 lists. From 10 to 20 lists, AUC rose by 0.023 in all 13 patients (exact two-sided Wilcoxon p = 0.00024). The channel-mean model gained 0.006 (p = 0.19); the channel-resolved gain exceeded the channel-mean gain by 0.017 (p = 0.0061). A power law fit predicts AUC 0.677 at 50 lists (95% CI 0.648–0.699), but this is an extrapolation that assumes the feature-label relation remains stable across sessions; session drift was not tested. The asymptote was poorly constrained, and the analysis used one dataset and task, session 0 only, fixed regularization, and one training set per fold at 20 lists. The results indicate that decoder performance continued to improve with more data within the observed session; they do not establish the benefit of pooling across sessions.

SEEG · word recall · memory decoding · learning curves · AUC · brain-computer interface · power law