A voltage model trained on pulses gave mean error decreases of 2.08 percent after pulses and 2.69 percent without pulses at 95 later sites in five SEEG sessions

We tested whether voltage before a pulse improves forecasts at later SEEG stimulation sites. OpenNeuro ds004457 v1.0.2 supplied nine bipolar readouts in five patient sessions. The model used eight voltage inputs. All model fitting and penalty selection preceded the later sites. Four eligible pulses set the mean at each new site. The last four pulses tested the 30 to 500 ms waveform. The test included 95 distinct later sites and 380 distinct test events. The patient mean error decrease was 2.08 percent relative to the four-pulse mean. One patient supplied 92.76 percent of the gain. The exact two-sided patient sign-flip value was p = 0.0625. A mean scaled for calibration noise had lower error than the voltage model in four of five patients. An exploratory control applied the unchanged pulse-trained coefficients to periods one second before each pulse. No pulse occurred in those target windows. This control gave a 2.69 percent mean error decrease, again with most gain in one patient. The result does not establish prediction of a change caused by stimulation. The package provides a chronological benchmark, input epochs, all forecasts, code, figures, source hashes, and raw retrieval steps. A fresh raw rebuild matched every stored epoch array. The scope is offline voltage in these five sessions.

SEEG · electrical stimulation · brain-computer interface · voltage forecasting · ridge regression · temporal generalization · no-pulse control · waveform prediction