Label shuffles within contact name groups give expected AUC of 0.621 for slow power in two 30-second SEEG clips from 25 patients

We tested slow power in public SEEG data from 25 patients. Each patient supplied two 30-second interictal clips from OpenNeuro ds004100, version 1.1.3. We used saved features from an earlier study and checked two raw recording prefixes again. The observed mean AUC for seizure-onset labels was 0.649. Uniform label shuffles within contact name groups gave an exact expected mean AUC of 0.621. This null mean was 0.121 above 0.5, with a 95% patient bootstrap interval of 0.054 to 0.185. The observed mean minus the null mean was 0.027, with an interval of -0.017 to 0.072. In 2,000 independent controls drawn under this conditional null, global shuffles rejected 100% and shuffles within groups rejected 5.05% at a 5% level. These models test different label assumptions. The study does not prove which model applies to clinical labels. We provide exact null formulas, a reusable score-table audit tool, exhaustive checks, source records, code, results, and reproduction instructions. The data overlap with the earlier study. The sample and short clips limit general use.

stereoelectroencephalography · SEEG · permutation tests · area under the curve · exchangeability · null models · epilepsy