Constructed rest score changes gave equal score distributions at 1 and 10 percent task rates for eight offline SEEG readers

We tested if unlabeled scores can separate 1 and 10 percent task rates. We used eight fixed readers from public SEEG data in OpenNeuro ds004473 version 1.0.2. In an exact model with two classes, we replaced 9.09 percent of each rest score distribution with its task score distribution. A 1 percent task rate then gave the same score distribution as a 10 percent rate with stable classes. The required rest CDF change was 0.0172 to 0.0741. This is a constructed score model, not an observed neural change. We also provide a task rate interval from a cumulative distribution function (CDF) error bound. It has at least 95 percent coverage under independent samples and valid class drift bounds. In matched independent simulations with 200 unlabeled windows, finite source samples kept both rates in 99.9 percent of intervals. The true rate was 10 percent. Each reader had 24 to 58 source scores per class. The package contains the proof, eight feature arrays, fixed readers, simulations, four figures, source records, and reproduction code. Fresh signal ranges from all eight recordings rebuilt 160 feature windows with no error. The clinical trials do not establish the sample independence assumption. The study does not test live command rates or stimulation.

score distributions · label shift · CDF bounds · DKW inequality · SEEG · brain-computer interfaces