A fixed SEEG classifier exceeded shaft label controls when each of 34 test patients was excluded from training

We tested a fixed SEEG power classifier in 34 patients from OpenNeuro ds004100 version 1.1.3. Each test patient was excluded from training. The primary rule subtracted each patient's median log power in seven bands. Mean patient AUC was 0.62786 in run 01. The mean shaft label control AUC was 0.54784. None of 9,999 control draws reached the observed mean. The upper-tail Monte Carlo p value was 0.0001. The control assumes uniform label allocation within each patient and shaft, with positive counts fixed. Both training and test labels changed in every draw. Median subtraction did not increase the observed mean AUC in either recording. The paired changes were -0.02911 and -0.00954. Descriptive fixed-score intervals included zero. All four models had higher mean Brier loss than a training prevalence baseline. A proof shows that median subtraction removes constant patient band offsets under fixed contact and feature rules. Fresh signal ranges reproduced all 70 source spectra exactly. A complete local replay reproduced all scores and control distributions exactly. The package contains code, features, labels, model maps, controls, source pins, checks, figures, and rerun instructions. The scope is one reused epilepsy cohort and clinical contact labels. The test does not establish surgical benefit or external validation.

stereoelectroencephalography · seizure onset zone · patient exclusion · nearest neighbors · AUC · label controls · median subtraction · Brier loss