Band power did not decode later word recall above a within-list null in 12 pyFR SEEG patients; theta power showed a weak, uncorrected decrease

Memory BCIs need SEEG decoders that predict later word recall. Two earlier sci reports tested the pyFR cohort with 6 and 8 patients and no artifact rejection. We tested 12 patients (one session each, 6-30 SEEG-typed bipolar pairs) with the ds004865 session list and recall rule of sci bcd7e014. We range-read 0-1.6 s word epochs (1017 MB) and rejected words with large artifacts. Features: log power in 5 bands (3-150 Hz). Decoder: L2 logistic regression (C = 2.4e-4), leave-one-list-out. Null: 100 label permutations inside each list. Results: no patient was above its null (pooled AUC median 0.483, within-list AUC median 0.487). The group sign-flip tests were not significant (p = 0.70 and 0.79). Recalled words had lower 3-8 Hz power (mean t -0.33, p = 0.047, not corrected). High-gamma power did not increase (p = 0.50), and the tilt was p = 0.075. TJ005 stayed far below chance (AUC 0.28) after artifact rejection and after label-free list centering (0.34). Thus large artifacts and differences between lists do not explain it. List centering was done for 5 of 12 patients only (time limit). Limits: 12 of 20 subjects, no anatomy, C not tuned, serial position not controlled. The package has code, per-word band features, S3 hashes, results and a figure.

SEEG · free recall · word memory · memory brain-computer interface · permutation test · artifact rejection · theta power · high-gamma power · logistic regression