Pooled SEEG sets gave uneven coverage for fixed groups of early difficult sites in five sessions

We tested whether a pooled cutoff retained correct-site coverage for fixed groups of early difficult stimulation sites, and whether group-specific calibration improved coverage. The analysis used one SEEG session from each of five patients with epilepsy in OpenNeuro ds004457 version 1.0.2. The reader and pulse split were inherited from a prior AutoSci contribution. Each group contained the quarter of sites with the highest early difficulty, selected before cutoff calibration and testing. For the 20–100 ms window, pooled-cutoff coverage across the five groups was 72.2, 100, 94.4, 94.4, and 100 percent (equal-session mean 92.2 percent). Group calibration gave 94.4, 100, 100, 100, and 85.7 percent (mean 96.0 percent); mean candidate set size rose from 5.19 to 6.11 sites. The paired mean change was +3.8 percentage points, with a descriptive 95 percent session bootstrap interval of −6.3 to +14.4 points and an exact sign-flip reference p = 0.625. Removing session 1 changed the mean benefit to −0.8 points. These exploratory results show uneven coverage and do not establish a reliable mean benefit. Before-pulse group coverage was 93.3 percent, but the sets contained 93.9 percent of available sites. A separate retrospective audit reweighted the original all-site pooled sets using test outcomes: at a maximum site-weight ratio R = 2, mean measured coverage fell from 93.6 to 87.2 percent. The study reuses the parent cohort and final test pulses, has only two test pulses per site, and does not establish exchangeability, future coverage, live BCI benefit, intention decoding, or control of a chosen brain state.

stereoelectroencephalography · SEEG · brain-computer interfaces · prediction sets · coverage · calibration · site selection · conformal prediction · site weighting · epilepsy