A fixed SEEG model increased shape error by a mean of 15.3 percent in five sessions at 20 to 100 ms

We tested a fixed voltage model in five patient sessions from OpenNeuro ds004457 version 1.0.2. Each of 134 remote stimulation sites supplied four fit pulses and four test pulses. The input covered minus 100 to minus 20 ms. The target covered 20 to 100 ms in 5 ms bins. The reference was each site's mean fit response. The primary model removed each window mean. It used a ridge penalty of 0.1 times the fit row count. Relative shape error increased by a mean of 15.3 percent across patients. A secondary raw model reduced total voltage error by 90.9 percent. The same raw predictions increased shape error by 1.7 percent. An exact error audit assigned 97.5 percent of baseline error to window means. Large raw gains also occurred before stimulation. The raw score did not establish a shape forecast gain. The five clinical patients, reused cohort, and offline site split limit the result. The package preserves code, voltage bins, predictions, tables, figures, source hashes, and reproduction instructions. Fresh source extraction matched all archived arrays. An independent solver and a complete replay also passed.

stereoelectroencephalography · electrical stimulation · voltage forecasting · shape error · ridge regression · neural interface · error decomposition · forecast calibration