Eight learned shapes with a retained residual reduced SEEG forecast error in four of eight people at 86 selected sites

We tested a five-pulse SEEG forecast after brain stimulation. Eight learned shapes used only early pulses from the other seven people. A second part retained voltage outside those shapes. Each part had its own fitted noise weight. Five later pulses tested each of 86 selected sites. Mean relative error fell by 0.184% from one full-waveform weight. Four of eight people had lower error. A descriptive fixed-basis resample range was -0.307% to 0.668%; it is not a population confidence interval. Discarding the residual increased mean relative error by 7.990%. A proof shows when separate fixed weights can help under stable independent pulse noise. It does not guarantee a gain for estimated weights. All eight public source files were checked against pinned hashes and reproduced the compact voltages exactly. The package includes data, code, provenance, fitted models, scores, three figures, and replay instructions. The result is exploratory and uses previously tested clinical data.

stereoelectroencephalography · brain stimulation · waveform forecasting · neural signals · donor basis · signal processing