In six pyFR SEEG patients, an FR1 word-recall decoder did not exceed within-list permutation nulls

We tested whether a within-patient word-recall decoder developed on RAM FR1 transfers to an independent free-recall cohort, pyFR (OpenNeuro ds004865), using SEEG-typed bipolar channels only. For the first eligible session in six of 19 planned patients, we used 0–1.6 s word epochs, six log-power bands from 3–150 Hz, L2 logistic regression (C = 7.2 × 10⁻⁴), and leave-one-list-out cross-validation. Pooled AUC had a median of 0.518 (range 0.233–0.590); same-list AUC had a median of 0.573 (0.245–0.607). Four of six patients were above chance on each metric, but none had within-list permutation p < 0.05; group sign-flip p values were 0.73 (pooled) and 0.61 (same-list). Four same-list AUCs were 0.56–0.61, near the FR1 one-shaft reference of 0.576, but below each patient’s 95% permutation limit. Two patients had unexplained below-chance AUCs. High-frequency (62–150 Hz) power was higher for recalled words in five of six patients (median channel-mean t = 0.31, p = 0.12); the 3–8 Hz effect was lower in two of six. A training-free high-minus-low frequency tilt index had median AUC 0.507. This interim test does not show transfer, and it has low power: only six patients were read, channels were few, no raw-signal artifact checks or anatomical channel selection were done, and 13 planned patients remain unread. The source release for the analyzed live-mirror files was not recorded. Two unusually low AUCs may reflect a recall relation that changes sign across lists, an artifact, or an event-to-EDF time offset; their cause is unknown.

free recall · word memory · SEEG · intracranial EEG · memory decoding · leave-one-list-out · AUC · permutation test · pyFR · FR1