Direction from the same patient's labels gave higher mean run 02 relative gamma AUC than direction from other patients in 35 SEEG patients

We tested direction choice for relative gamma power in public SEEG data from OpenNeuro ds004100, version 1.1.3. The test used 35 patients and two 20 s samples per patient. Direction set with the same patient's run 01 onset labels gave mean run 02 AUC of 0.651. Direction set with labels from the other 34 patients gave 0.511. The paired difference was 0.140. Its 95 percent patient bootstrap interval was 0.050 to 0.341. Shared random labels gave same-patient mean AUC of 0.597. Independent random labels across runs gave 0.501. An exact rank example shows why a new recording can retain an effect of label use. A shaft label-shift model retained most labels and gave a range that contained the observed mean. A planned secondary test of absolute gamma power gave fixed high-direction mean AUC of 0.743. Thus, the main effect is specific to the tested relative score. The labels require complete source onset marks. This is a validation benchmark, not a clinical tool for an unknown patient. The package contains checked code, scores, spectra, patient results, control draws, source hashes, and retrieval instructions.

stereoelectroencephalography · epilepsy · gamma power · patient-level validation · AUC · label leakage · recording holdout