A last-pulse correction increased SEEG forecast error over the past mean in three of five sessions at 20–100 ms

We tested a last-response correction for the next SEEG pulse response at a known site. The public source was OpenNeuro ds004457 v1.0.2. The test used five patients, 183 sites, and 732 scored pulses. Each forecast and fitted weight used only complete earlier waveforms. At 20 to 100 ms, mean patient error reduction against all earlier responses was +5.93 percent. The descriptive 95 percent patient interval was −6.80 to +26.34 percent. Three sessions had more error; two had less error. Removing each patient’s largest-reference-error site changed the mean reduction to −5.00 percent. Separate recording contacts gave −2.66 percent. One site supplied 90.4 percent of patient 5’s benefit. Thus, the positive mean was sensitive to the measured signal set. We checked 110 raw source files and rebuilt every input exactly. The package preserves code, bins, input times, fitted weights, forecasts, losses, three figures, and repeat instructions. This offline result uses retrospective labels and reused patients. It does not establish a neural cause or live BCI benefit.

stereoelectroencephalography · electrical stimulation · voltage forecast · brain-computer interface · forecast error