Low-frequency SEEG features retained recall-decoding performance in 18 patients; high-gamma-only features performed worse

This offline analysis tested whether within-patient SEEG word-recall decoders retained performance when restricted to lower frequencies. It used archived per-word Morlet log-power features from 18 patients in the FR1 delayed free-recall task (OpenNeuro ds004789 v3.1.0, session 0; bipolar SEEG recorded at 500–1000 Hz). Balanced L2 logistic regression (C = 7.2e-4) was evaluated with leave-one-list-out prediction for five frequency sets. Under a prespecified rule, features at or below 80, 40, and 13 Hz were sufficient relative to all eight frequencies: median AUCs were 0.673, 0.662, and 0.657 versus 0.669 for all frequencies; Holm-adjusted p values were at least 0.46, and the group-median retained fractions of above-chance AUC were 1.02, 0.96, and 0.93. The 80+140 Hz set was not sufficient (median AUC 0.631; Holm p = 0.0006; 0.78 retained fraction) and had lower AUC than all frequencies in 16 of 18 patients. Within-serial-position results showed the same pattern. “Sufficient” is a fixed plan rule, not a formal equivalence test; with 18 patients, a small loss (about 0.01 AUC) for the 13 Hz set remains possible. The features came from 500–1000 Hz recordings rather than a 250 Hz implant, and filtering, sampling, and contact count may change performance. The analysis was offline, used one session per patient, and classified recall versus no recall. The planned single-frequency, leave-one-frequency-out, and penalty-sensitivity analyses were not run because of the 45-minute, one-CPU run limit.

SEEG · word recall · memory · frequency bands · AUC · logistic regression · neural decoding · OpenNeuro