With agreed labels fixed, source AUC was higher than mean reference AUC for all disputed N1 or P1 labels at 20 to 90 ms in UMCU21 SEEG

We tested how expert label disagreements affect a stimulation-response reader comparison. The public UMCU21 record from OpenNeuro ds004774 version 1.0.0 supplied 330 pulses and six SEEG contacts. SEEG means stereoelectroencephalography. Two experts had labels for 188 matched cells at 33 selected stimulation sites. The primary endpoint was an N1 or P1 response. There were 38 disputed binary labels. We kept agreed labels fixed and allowed each disputed label either value. A rank identity gave exact AUC difference limits without enumerating all label choices. AUC is the area under the receiver operating characteristic curve. At 10 to 90 ms, mean minus source AUC ranged from -0.0968 to +0.0087. At 20 to 90 ms, it ranged from -0.1466 to -0.0187. Thus, source AUC was higher for every permitted primary label choice in the later window. The result did not hold for N1-only labels. Two disputed-label changes reversed expert SB's N1 comparison while keeping eight positives. The package adds an exact audit method with a proof, source code, voltage data, label assignments, and three figures. Small exhaustive tests, all 816 saved bound endpoints, a clean numerical replay, and fresh checks of all 330 raw ranges passed. The limits apply to fixed scores and this reused record. They are not confidence intervals or proof of correct neural labels. No live interface or new stimulation was tested.

AUC · label uncertainty · SEEG · N1 · P1 · stimulation response · robustness analysis · brain-computer interface