Removing features above 70 Hz changed AUC little in two offline SEEG tasks; non-inferiority held only in ds004789

Question: which minimum ADC sampling rate does an implanted SEEG band-power BCI decoder need? Method: archived band-power features of two OpenNeuro datasets (ds004752 Sternberg set size 8 vs 4, 15 patients, 5 bands 4-150 Hz; ds004789 FR1 word recall, 19 patients, 8 Morlet frequencies 3-140 Hz). Shrinkage LDA, stratified 5-fold CV, AUC for each patient. Feature sets: cumulative low-pass sets, leave-one-band-out and single-band sets. Fs_min = 2.5 x the upper band edge. Paired sign-flip tests across patients, bootstrap CIs. Hypotheses were pre-registered in PLAN.md. Results: removal of all features above 70 Hz (Fs_min 375-490 Hz to 140-175 Hz) changed AUC by -0.005 (95% CI -0.034 to +0.023, p = 0.77) in ds004752 and by -0.003 (-0.017 to +0.013, p = 0.70) in ds004789. Non-inferiority at a 0.02 margin was shown in ds004789 (p = 0.019) but not in ds004752 (p = 0.16). High-gamma features alone had information (AUC 0.628 and 0.568) but were redundant with the lower bands. The 70-150 Hz band was not the most necessary band (H3 not supported); in ds004752 the 30-70 Hz band was (-0.025, p = 0.043). Loss grew below about 30 Hz (ds004752 <= 8 Hz: -0.083, p = 0.03). Limits: offline; low-pass simulated by band removal, not resampling; one linear decoder; two cognitive tasks, not motor or speech; ds004789 used 1 CV repeat, not the 3 in the plan (CPU time). Package: code, logs, per-patient JSON results, figure, tables, feature checksums and reproduction commands.

SEEG · brain-computer interface · sampling rate · band power · shrinkage LDA · AUC · high gamma · working memory · free recall