A scale update had lower geometric mean error than half-size pulse forecasts and higher error than zero in eight SEEG sessions at 20–500 ms

We tested one new calibration pulse after a recorded contact order change. Five old pulses supplied the response template. Four later pulses supplied the measured target. We used public STReEF SEEG data from eight people, 96 fixed conditions, and 6,803 stimulation-to-readout instances. At 20 to 500 ms, the update gave 26.5% lower geometric mean error than a fixed half-size new-pulse forecast. The error ratio was 0.735, with a descriptive 95% person bootstrap interval of 0.713 to 0.756. All eight person ratios were below one. The update error was 3.9% above zero-forecast error on this score. Its ratio to zero was 1.039, with an interval of 0.924 to 1.144. An arithmetic energy score gave a different result. The gain over the half-size pulse was larger before stimulation. A projection proof and a numerical control show how noise reduction can give a comparative gain without useful response prediction. This exploratory offline test does not establish neural artifact removal or a live BCI gain. The package preserves source code, figures, every instance error, version pins, hashes, and reproduction instructions. A full raw rebuild verified all 960 ranges and matched all input arrays exactly. A fresh analysis repeated the results and figures exactly.

SEEG · stereoelectroencephalography · intracranial EEG · brain-computer interface · waveform forecasting · template update · electrical stimulation · STReEF · OpenNeuro · forecast error