A bounded shift model reduced the mean relative prediction gap to two free waveforms in seven selected SEEG patients under fixed source correction

We tested whether one shared waveform with bounded contact shifts reduces a fixed derivative model's prediction gap to two free waveforms. We used public OpenNeuro ds004696 version 1.0.1, 80 selected stimulation groups from seven patients, and all 252 five-of-ten pulse splits. The response window was 100 to 500 ms. The primary shift bound was 39.0625 ms per contact. With fixed source correction and equal patient weights, shifts reduced the mean relative gap by 52.50 percent. The remaining gap was 3.52 percent of one-waveform test loss. The exact two-sided patient sign test gave p = 0.078125. It assumes independent patients and symmetric null differences. A different loss denominator slightly favored shifts. The cohort was used earlier, and source correction uses all exported pulses. A larger iteration limit gave a similar primary mean. All seven pinned source objects were rebuilt. All compact arrays and a second full analysis matched exactly. The SCI package preserves voltage, source versions, scores, plans, figures, logs, and license records. GPL analysis code is omitted, limiting public computational reproducibility. It supplies a bounded timing benchmark. It does not establish a neural timing cause.

stereoelectroencephalography · electrical stimulation · waveform modeling · time shifts · prediction · signal processing