Fitting gamma score direction raised AUC across two runs with shared random pair labels in 35 public SEEG maps

We tested how a fitted score direction changes measured seizure-onset localization. We used two 20-second interictal SEEG runs from each of 35 patients in OpenNeuro ds004100, version 1.1.3. The fixed score was relative gamma power from adjacent contact differences. Choosing direction with the current patient's source labels gave mean run 02 AUC 0.651. Choosing it with the other patients' labels gave 0.511. In 10,000 random-label draws with fixed positive counts, one pair label map shared between runs gave mean AUC 0.597. A fresh random test map gave 0.500. A split between runs can thus retain a direction selection gain under this model. We give a proof for identical ranks, complete simulation draws, source code, metadata, raw example crops, and a full retrieval recipe. Independent signal checks matched all 7,170 pair-run scores from 70 source crops. This is a secondary analysis of the same contact sites. The model does not preserve clinical label clusters. The result does not validate gamma power for clinical use or estimate localization in patients with unknown onset labels.

stereoelectroencephalography · seizure-onset zone · gamma power · area under the curve · direction selection · label leakage · random-label model · patient calibration · validation design