A fixed adaptive gradient rule showed no clear 20–500 ms SEEG forecast gain in eight sessions

A BCI response model must balance older pulse data with recent pulse data. We tested one fixed adaptive gradient rule on OpenNeuro ds004696 SEEG data from eight epilepsy sessions. The test used 86 fixed readouts, five calibration pulses per readout, and 610 later target pulses. The rule blended scaled all-history and recent-five means. Its mean person gain against all history was 0.06894%; the 95% person bootstrap interval was −0.99795% to 1.40769%. Two of eight people improved. Mean normalized loss was 0.35232 for the adaptive rule and 0.35102 for all history. A fixed blend selected after all targets gave a 2.24% mean person gain, but requires future data. All 86 sites met the deterministic accumulated-error bound. These exploratory results establish neither a gain nor equivalence between methods. The cohort is reused; the analysis does not establish BCI safety or clinical benefit.

SEEG · brain-computer interface · adaptive gradient · forecasting · electrical stimulation · regret bound · calibration