Copulas reduced mean HFO forecast loss on later dates at contacts 1–2 in seven SEEG patients with matched shaft probabilities

Public SEEG marks came from OpenNeuro ds003498 version 1.1.1. We fit each patient model on the first acquisition date. We tested it on later dates in seven patients. The main test used contacts 1–2 and 2,646 target windows. A Gaussian copula joined each shaft's activity probability into a joint forecast. The comparison assumed independent activity with the same shaft probabilities. Mean forecast loss fell by 0.056448 nats per ten second window. Six patients improved and one became worse. The exact patient bootstrap 95 percent interval was 0.011638 to 0.100331 nats. The two-sided patient sign-flip test gave p=0.078125. All four planned sensitivity checks are reported. This is a secondary, exploratory study of retrospective detector marks. The results do not test clinical benefit. The package preserves source inputs and hashes, code, forecasts, figures, and reproduction instructions. A new public download reproduced all 16 analysis files with identical bytes.

high-frequency oscillations · stereoelectroencephalography · Gaussian copula · joint forecasting · negative binomial model · event counts · forecast evaluation