A two-second update decreased mean SEEG gamma forecast error between files at a 50 ms gap; 20 of 36 patients had higher error

We tested a past mean update for fixed gamma power forecasts in 36 patients. The public SEEG data came from OpenNeuro ds004100 version 1.1.3. Each patient supplied four contact pairs and two 60-second files. At a nominal 50 ms target gap, the update decreased mean normalized error by 6.58% compared with a frozen ridge forecast. The 95% paired patient bootstrap interval was 1.49% to 11.81%. Error increased in 20 of 36 patients. The median patient error increased by 0.74%. A baseline-only rolling mean had slightly lower error than the updated ridge forecast. In a later part of the calibration file, the update increased mean error by 3.50%. We also proved a constant-shift rule for the update when all test log powers, including the target, have the same shift. The package contains 72 native voltage excerpts, pinned source checks, code, models, patient scores, and three figures. This exploratory study uses reused clinical patients and short files. It does not test task decoding or brain stimulation.

SEEG · intracranial EEG · gamma power · forecasting · ridge regression · signal drift · brain-computer interface