In 21 SEEG patients, theta and HFA memory-effect maps weakly repeated across sessions

Question: do the electrode-pair maps of the subsequent memory effect (SME) in human SEEG repeat across sessions? Also, does a recall classifier trained on one session operate on a different day? Data: OpenNeuro ds004789 (RAM FR1, delayed free recall), bipolar SEEG. Session-0 features come from archived package d3aeb47e. Session-1 features were extracted with the same code. Features were log band power in the 0–1.6 s after word onset for theta, alpha, beta, low gamma, and high-frequency activity (HFA, 70–150 Hz). We removed list and serial-position means and z-scored each pair. The SME was the standardized difference between recalled and not-recalled words. For each subject, we correlated the SME maps of the two sessions across common pairs. We used a sign-flip test on Fisher z with Holm correction for H1 (HFA) and H2 (theta). The plan was written and hashed before session-1 data were read. Results: 21 of 23 fetched subjects were included. Theta map mean r = 0.18 (17/21 positive, Holm p = 0.002). HFA map mean r = 0.10 (16/21 positive, Holm p = 0.014). An L2 logistic classifier trained on session 0 gave session-1 AUC = 0.552 (95% CI 0.531–0.573, 18/21 above 0.5). For the top 20% of pairs, the sign repeated in 58–62% of pairs. The within-session split-half map r was not higher than the cross-session r. This result is consistent with trial noise, rather than change between days, setting most of the limit on map reliability, but each split half had only half the words. Effects were small; there was one session pair per subject; pairs were matched by label; and one secondary benchmark (within-session CV AUC) was not computed because of CPU time. Session-1 files were fetched from an unversioned public S3 path; recorded hashes identify the files used, but do not establish a match to a versioned dataset snapshot.

SEEG · subsequent memory effect · free recall · cross-session reliability · electrophysiology · memory decoding