The mean spectral AUC gain stayed positive under up to one adjacent label exchange per patient. The test used fixed scores and onset counts in 31 SEEG patients.

We tested how changes in evaluation labels affect a fixed SEEG ranking result. The public source was OpenNeuro ds004100 version 1.1.3. We used archived scores for 31 patients, 2,879 contacts, and 271 onset labels. Each model excluded the test patient from training. The primary contrast compared contact number plus six CAR spectral features with contact number alone. The mean patient AUC gain was 0.074032. Under up to one adjacent positive-negative label exchange per patient, its exact range was 0.052324 to 0.079182. Adjacent contacts had consecutive final numbers in the same parsed name prefix. Scores and onset counts stayed fixed. The lower mean stayed positive over all allowed disjoint adjacent exchanges. When any two contacts in one prefix could exchange labels, 31 exchanges removed the mean gain. The exact lower mean was positive at a total budget of 30 and negative at 31. The latter changes moved 31 of 271 original onset labels. We provide integer and rational bound code, proofs, contact changes, frozen inputs, source recipes, and independent checks. These are exact bounds for hypothetical evaluation labels. They do not measure clinical label error or surgical benefit.

SEEG · seizure onset labels · AUC · label sensitivity · exact bounds · spectral features · ranking