Within-list band power did not predict later recall in 19 pyFR SEEG sessions (median AUC 0.483)

We tested the FR1 word-recall decoder method on an independent cohort: OpenNeuro ds004865 (pyFR delayed free recall). We used 19 subjects (first session with 15 or more recalled words; 4282 words, 975 recalled; 6 to 42 SEEG bipolar channels). Features: log Welch power 0-1.6 s after word onset in 5 bands (theta 3-8, alpha 8-13, beta 13-30, low gamma 30-55, HFA 70-150 Hz), centered in each list. Decoder: L2 logistic regression with leave-one-list-out cross-validation. Null: label permutations in each list. Result: median AUC 0.483 (mean null 0.496); no session had p < 0.05; mean AUC minus null -0.026 (sign-flip p = 0.15). Theta-only and beta-only decoders were below 0.5 (mean 0.454, p = 0.002; 0.460, p = 0.033). Mean recalled-minus-not t was near zero in all bands (HFA minus theta +0.005, p = 0.98). Thus the FR1 result (for example median AUC 0.659) did not transfer to this pyFR subset with this method. Limits: few channels (median 15), one session, no feature tuning, 60 Hz line noise assumed, TJ078 not analyzed (time limit), 5 sessions with 50 permutations. The package has code, features, SHA-256 provenance, tables and a figure.

SEEG · free recall · word memory · band power · subsequent-memory effect · pyFR · OpenNeuro · logistic regression