Channel-frequency signs predict free recall better than frequency-level signs, while low-frequency power decreases in 18 SEEG patients

We tested whether the sign structure reported for SEEG decoding of working memory also occurs when encoding activity predicts later free recall. In archived features from 18 patients in OpenNeuro ds004789 FR1 (session 0; eight frequencies; offline), a decoder using one learned sign per channel-frequency had mean AUC 0.638, compared with 0.573 for one sign per frequency (difference +0.065, 95% CI +0.020 to +0.110; positive in 14/18 patients; exact sign-flip p = 0.0015). This replicates the decoder result from ds004752 (+0.071, 14/15 patients). The band-level result differed: mean channel t was negative at 3–22 Hz in 14–17 of 18 patients (Holm-adjusted p = 0.00018–0.016), while the predicted high-gamma increase was small and not significant at 80 Hz (+0.08, p = 0.72) or 140 Hz (+0.15, p = 0.33). At low frequencies, significant channel effects were mostly negative; at 80 Hz both signs exceeded the permutation null, although per-patient evidence for mixed signs was weak. The sign-per-channel-frequency decoder outperformed shrinkage LDA on the same features (AUC 0.638 vs 0.603, p = 0.010). Signs derived from the other 17 patients gave AUC 0.582, similar to fold-learned frequency signs and lower than channel-frequency signs. These offline findings come from one task, one session, one feature window and montage, and an 18-patient subset; they do not test closed-loop stimulation. Different features and cross-validation prevent direct comparison of exact AUC values with ds004752.

SEEG · free recall · memory decoding · neural signal processing · sign decoder · intracranial EEG