Within one offline SEEG session, decoder accuracy did not decline with training-list lag despite feature drift (18 patients)

We tested if a SEEG word-recall decoder loses accuracy during one session. Data: OpenNeuro ds004789 FR1, session 0, 18 patients, per-word spectral power. Each decoder used exactly 4 training lists. The metric was within-list AUC. The plan was fixed before the fits. Decoders from the 4 nearest lists and the 4 farthest lists had the same AUC (median difference -0.008, 95% CI -0.054 to +0.039, 8/18 positive, exact Wilcoxon p = 0.80). Recalibration on the 4 most recent lists did not do better than a decoder fixed on lists 1-4 (median +0.000, CI -0.044 to +0.017, p = 0.50). AUC did not change with lag (median rho +0.01, p = 0.55). But the feature mean drifted with lag (median rho +0.34, 16/18 patients, p = 0.00033). Limits: offline, one session, small 48-word decoders with AUC near 0.6, and within-list AUC does not measure threshold drift.

SEEG · free recall · memory decoding · brain-computer interface · neural feature drift · within-list AUC · logistic regression