Eight SEEG tests did not establish 50 percent precision under binomial bounds that assume independent trials, stable class distributions, and a 1 percent task rate

We tested whether a reader with no calibration false calls can establish at least 50 percent precision at an assumed 1 percent task-window rate. We used 2,014 retained trials from eight public SEEG records in OpenNeuro ds004473 version 1.0.2. Separate time blocks fitted the model, set the threshold, and tested the calls. A threshold at the largest calibration rest score reduced mean test false-call rate from 28.53 to 0.93 percent. It also reduced mean task detection from 69.49 to 14.13 percent. At the assumed task rate, precision for an equal-person class mixture was 13.25 percent. No record reached a 50 percent lower precision bound. Under the stated exact binomial rule, even perfect observed test decisions need 368 independent windows per class to pass that goal. Actual tests had 32 to 75 per class. The bounds assume independent trials and stable class distributions. This is an exploratory study in a reused clinical cohort. Its response-aligned windows do not label command intent. The package includes three figures, unchanged source features, new models and predictions, source versions and hashes, a sample-size calculator, and reproduction code. A fresh check rebuilt 240 windows from 48 source trials with no feature difference.

stereoelectroencephalography · brain-computer interface · precision · false-call rate · calibration · binomial confidence bounds · sample size