Growing means gave no clear gain above a random order reference for 20–500 ms SEEG forecasts in eight patients

This study asks if past pulse responses improve SEEG forecasts beyond the gain from a larger mean alone. It uses public OpenNeuro ds004696 version 1.0.1 and the same eight patients as prior sci studies. There are 86 included stimulation sites and 886 site/readout combinations. Each forecast uses only completed past retained responses. For a fixed set of ten waveforms, the study proves an exact mean error reference under random pulse order. The proof does not require independent pulse noise. In the primary 20–500 ms window, the growing mean gain above this reference was 0.0092 in normalized error units. The descriptive 95% patient interval was -0.0038 to 0.0250. A joint order test gave p=0.144 under stated patient independence and pulse exchangeability assumptions. The full-window test gave no clear evidence of extra growing gain from the recorded order. This does not prove fixed responses or pulse independence. Secondary early-window and recent-mean results were more favorable. The recent gain was sensitive to the scoring scale. The package preserves a proof, compact voltage data, event maps, five figures, source hashes, and retrieval instructions. GPL analysis and extraction scripts are omitted, so the exact archived analysis implementation is unavailable for public rerun. It preserves exact analysis replay checks and two fresh public source reconstructions. This is an exploratory cohort reanalysis. It does not test a clinical benefit or an online controller.

SEEG · brain-computer interface · voltage forecasting · randomization test · pulse order · forecast calibration