A five-pulse geometric median had 7.06% higher later 20–500 ms centered bipolar SEEG error than a noise-scaled mean in eight patient sessions

A short brain-computer interface calibration can use five early electrical pulse responses to forecast later brain voltage. We tested a geometric median on public OpenNeuro ds004696, version 1.0.1. The fixed sample has eight epilepsy patient sessions, 86 stimulation sites, and 886 stimulation-to-recording pairs. The primary score uses five separate later pulses in the strict 20–500 ms window of time-centered bipolar SEEG. The median gave a 0.11% mean patient error reduction from the ordinary mean. Its descriptive 95% patient bootstrap interval was −1.03% to 1.08%; the two-sided patient sign-flip p-value was 0.852. The median had 7.06% higher error than the noise-scaled mean, with an interval of 5.70% to 8.49%. A separate simulation multiplied one early voltage waveform by 2, 4, or 8. The median limited this scale error compared with the ordinary mean. It had no clear advantage over noise scaling at gain 8. These are offline voltage results, not tests of a percept or an online controller. The package preserves exact trial products, fitted coefficients, pulse errors, three figures, code, numerical checks, and retrieval instructions. Fresh source rebuilds matched all 48 arrays and 16 tables; 44 metadata files also matched. The study adds a fixed-input method comparison and simulation to the cited parent.

stereoelectroencephalography · brain-computer interface · geometric median · pulse response · forecast error · calibration