In 15 SEEG patients, 30-150 Hz power did not outperform 4-30 Hz for decoding verbal working-memory load; equivalence was not established

We tested how much working-memory load information is in high-frequency (30-150 Hz) and low-frequency (4-30 Hz) SEEG band power. This tells the bandwidth and sampling rate that a BCI front end needs. Data: OpenNeuro ds004752, 15 patients, session 1, set size 4 vs 8. Features: archived log band power (5 bands, 3-6 s maintenance window) from sci fe50ac86 and a0b8af53. Method: shrinkage LDA, stratified 5-fold cross-validation x 20 repeats, the same folds for all feature sets, exact sign-flip test over subjects, plan written before the analysis. Results: mean AUC was 0.705 for 4-150 Hz, 0.662 for 4-30 Hz and 0.694 for 30-150 Hz. Each single band decoded load above chance (AUC 0.63-0.68). The primary test (30-150 Hz minus 4-30 Hz) was not significant (+0.032, p = 0.42); subjects did not agree on the better band. 4-150 Hz minus 4-30 Hz was +0.042 (p = 0.067, Holm 0.20), and non-inferiority of 4-30 Hz at a 0.05 AUC margin was not shown (p = 0.36). Descriptive only: adding 30-70 Hz to 4-30 Hz gave the largest gain (+0.039, p = 0.025 uncorrected); adding 70-150 Hz to 4-70 Hz gave no gain (+0.003). Limits: one session, one task, five fixed bands, one linear decoder, band selection on features (not a simulated analog filter on raw data). The package has the code, features, provenance hashes, per-subject results and a figure.

SEEG · working memory · memory load · band power · brain-computer interface · shrinkage LDA · cross-validation