SEEG band power predicts later free recall in ds004865, including after serial position is controlled

We decoded later free recall from SEEG band power during word encoding in OpenNeuro ds004865 (pyFR). The data are 19 subjects, one session each, 4,341 words and 973 recalled. Only channels labeled SEEG were used (bipolar). The features were Welch log power in five bands (3-150 Hz), 0-1.6 s after word onset. The decoder was L2 logistic regression with C = 0.01 fixed and leave-one-list-out cross-validation. Median held-out AUC was 0.580 (15 of 19 above 0.5; sign-flip p = 0.0002). 8 of 19 subjects had a permutation p < 0.05 (binomial p = 1.8e-6). A permutation inside each serial position kept the effect: median within-position AUC 0.578, p = 0.013, 7 of 19 subjects with p < 0.05. Serial position alone gave AUC 0.555. A 70-150 Hz-only decoder was weaker (median 0.530). Recalled words had more 70-150 Hz power (15 of 19 subjects, p = 0.043) and less 8-12 Hz power (16 of 19, p = 0.021). All sessions showed a word-locked ERP, so the event alignment is plausible. The effect is smaller than the ds004789 FR1 results (median AUC 0.63-0.69). Limits: clinical placement, one session, fixed C, Welch features, offline only. sub-TJ069 was excluded because of an event-to-EDF mismatch. The package has the code, the per-word features, provenance hashes, results and a figure.

free recall · SEEG · memory decoding · subsequent-memory effect · serial position · Welch power · AUC · pyFR