Separate power features above 150 Hz add no AUC beyond 70–150 Hz power in eight SEEG subjects

In eight subjects performing a key-press task, shrinkage LDA classified move versus rest from 0.5 s SEEG windows with mean AUC 0.758 using 70–150 Hz power. Adding separate 150–250 Hz and 250–400 Hz features gave AUC 0.759 (+0.002; exact two-sided sign-flip p = 0.73), so those features did not improve this decoder. Power above 150 Hz alone carried information (H1+H2 AUC 0.661, p = 0.008 vs 0.5) but was weaker than 70–150 Hz power. In a secondary test, one 70–400 Hz broadband power value per channel gave AUC 0.784 (+0.026; 7/8 subjects; p = 0.016); this result did not survive Bonferroni correction for 10 secondary tests. Rest power at 250–400 Hz was 16.3–16.8 dB above the EDF quantization floor, while the nearly flat spectrum above about 250 Hz suggests instrument noise may contribute substantially; amplifier noise was not measured. These offline results from eight epilepsy patients, one task, 0.5 s windows, and one decoder suggest that about 400 Hz sampling is sufficient for this decoder when separate band features are used. Broadband pooling to 400 Hz may give a small gain that needs independent confirmation. Shorter windows, other tasks, and other decoders may produce different results.

SEEG · brain-computer interface · high-frequency power · key-press detection · move-rest decoding · shrinkage LDA · sampling rate · OpenNeuro ds004473