A prior from past sessions reduced simulated 50 ms gap forecast error in 215 depth SEEG records from seven people with 2 s calibration and ridge 0.1

A prior from earlier sessions improved a short voltage forecast fit in public OpenNeuro ds007095 version 1.0.0. The test used 215 later depth SEEG records from seven people. Each record had two bipolar channels at 200 Hz. The current fit used 2 s of calibration. Both primary models used ridge 0.1. Only the penalty center differed. The primary prior used 20 s from each earlier selected record of the same person. Mean normalized error across simulated 50 ms gaps fell from 0.78451 to 0.76556. This was a 2.42% reduction. Error fell in six people. The descriptive 95% whole-person bootstrap interval for the reduction was [0.00774, 0.03216]. The exact two-sided sign-flip reference was p = 0.03125. Its interpretation requires independent people and symmetric null effects. The past fit alone had lower error than the short update in every person. A current 20 s fit had the lowest mean among eight fixed methods. At the 50 ms endpoint, the primary prior was worse than the constant mean. The small ridge 0.01 check had an interval that included zero. This is an exploratory study on detector-selected clinical records and previously tested intervals. It does not test real stimulation artifacts or preserved neural information. The archive includes 222 native EDF records, source hashes, a full selection ledger, code, all primary forecasts and coefficients, four figures, and reproduction instructions. Independent numerical checks passed. A separate full run reproduced all 41 analysis files exactly.

depth intracranial EEG · forecasting · ridge regression · short calibration · transfer learning · simulated signal gaps · responsive neurostimulation