Early linear calibration with up to three extra SEEG pairs gave a 0.30% mean decrease in future theta forecast error in seven rest records

We tested whether extra SEEG voltage inputs improve a future theta voltage forecast under short early calibration. Seven public rest records supplied four fixed pairs each. We selected linear models in early data, refit them on 56 seconds, and tested a separate later period. The model could retain the own-pair rule or add the other three histories. Mean squared error decreased by 0.30% relative to own inputs; the descriptive 95% person bootstrap interval was −0.28% to 0.87%, and four of seven person means decreased. Own-pair and joint-selection skill against a fixed calibration mean were 20.62% and 20.88%. The aligned result exceeded all 19 circular shift controls. Restricting extra inputs to different shaft labels gave a 0.15% mean decrease, with an interval across zero. The target contains only future samples but has about 1050 ms of support after the cutoff; it is not an instantaneous 50 ms voltage forecast. Input timing checks start at stored samples, not the original acquisition stream. This exploratory analysis reuses epilepsy records and does not show a clear mean benefit across people or establish stimulation benefit.

SEEG · theta · linear forecasting · ridge regression · future voltage · stereoelectroencephalography · brain interfaces · forecast error · calibration