SEEG band-power patterns drift in 19 of 19 patients, but did not predict next-word recall beyond lag in 9 patients

We tested two temporal-context predictions using archived SEEG band-power features from OpenNeuro ds004789 (FR1 delayed free recall, session 0). Across all 19 patients, similarity between word-encoding patterns declined with serial-position lag: the median per-patient Spearman rho was -0.142 (95% CI -0.184 to -0.102; 9/9 pre-registered patients negative; exact p = 0.0020), and it was negative in all 19 patients in a post hoc analysis (p = 1.9e-6). Post hoc, similarity also declined with distance between lists in 18/19 patients, indicating that some drift was session-level. The primary recall-cue test included 9 patients and 376 transitions. Among words recalled later in the same period, lag-residualized similarity to the last recalled word did not significantly predict the next word (median percentile 0.511, 95% CI 0.488 to 0.577; p = 0.125). Raw similarity and the secondary lag-residualized test over all candidates were positive, but can reflect temporal lag or encoding success. Behavioral temporal contiguity was strong (median percentile 0.673; 9/9; p = 0.0020). The negative primary result has low power and does not exclude a small effect. The drift may reflect neural context, arousal, attention, or recording changes; these causes cannot be separated here. Features were band power only, with no region selection, and only session 0 was studied.

SEEG · temporal context · word-list learning · episodic memory · band power · recall · neural drift · brain-computer interface