Four fixed temporal scales increased mean 20–500 ms centered bipolar SEEG forecast error relative to one scale in eight patient sessions with five calibration pulses

The method with four scales increased mean error by 0.48% relative to one scale for the whole waveform. We tested five early calibration pulses against five later individual pulses in public OpenNeuro ds004696. The fixed cohort had eight patients, 86 sites, and 886 pairs of sites and readouts. The window was 20–500 ms after a pulse. The descriptive 95% patient bootstrap interval for this increase was 0.22% to 0.75%. 7 of eight patients had higher mean error. The exact two-sided sign-flip value was p = 0.0234375. The four-scale method still reduced mean error from a zero forecast by 23.35%. Four fixed cosine groups used scales from calibration pulses. All eight source files passed hash and reconstruction checks. The package preserves exact band products, individual errors, source metadata, code, figures, and a retrieval recipe. This is an offline voltage forecast result within one session per patient. Online BCI control and functional effects were not tested.

SEEG · voltage forecasting · electrical stimulation · discrete cosine transform · brain-computer interface