A fixed variance mixture gave no clear probability score gain for 4–30 Hz power at 100–500 ms after a pulse in eight SEEG sessions

We tested forecasts of power after electrical stimulation in public SEEG data. The data came from OpenNeuro ds004696 version 1.0.1 and eight people with epilepsy. The fixed cohort had 86 stimulation sites and 886 bipolar readouts. Each forecast used five local pulses. The primary target was natural log 4–30 Hz power at 100–500 ms after a pulse. We compared local variance with an equal mixture of local variance and variance from earlier sites. Both forecasts had the same mean and Student t distribution form. The test used 611 distinct later pulses. The mean negative log density change was −0.0196 nats. The descriptive 95% person bootstrap interval was −0.0399 to +0.0055 nats. The mixture increased 90% interval coverage from 91.0% to 94.5% and increased mean width. A secondary pre-pulse test showed a larger score gain. The package contains figures, all compact inputs, every forecast, code, source hashes, and a new full source check for one person. These reused clinical data do not establish a live BCI benefit or a stimulation policy.

SEEG · electrical stimulation · probability forecast · variance transfer · brain-computer interface · negative log likelihood · calibration