A five-pulse geometric median did not show lower normalized SEEG forecast error than noise scaling under one simulated fivefold gain fault in eight patient sessions

We tested five-pulse response calibration with public SEEG data from OpenNeuro ds004696 version 1.0.1. Eight patient sessions included 86 stimulation sites and 886 stimulation-to-recording pairs. One early pulse waveform had a simulated +5 voltage gain. Each forecast was tested on five separate later pulses. The primary loss normalized each pair by its unchanged calibration energy and gave each patient equal weight. The geometric median had a mean patient error increase of 2.52% relative to a noise-scaled mean. The 95% patient bootstrap interval was -1.51% to 6.62%; the exact two-sided sign-flip p value was 0.3047. The median reduced error by 46.15% relative to the ordinary mean under the same fault. With raw voltage loss, it reduced error by 18.42% relative to noise scaling. The loss definition changed the comparison. An added proof shows a finite amplitude bound for noise scaling when one pulse is replaced by any waveform. The package has exact trial products, all fits and losses, five figures, code, source hashes, and rebuild instructions. A source rebuild for one patient and a full analysis rerun matched the archived inputs and results. This study did not test an online BCI or a new session.

SEEG · signal processing · geometric median · brain-computer interface · forecast error · robust estimation · electrical stimulation