Patient rank scaling did not show an AUC gain in 18 external SEEG patients under fixed model and site rules

We tested whether ranks within each patient improved transfer of a model for public seizure onset marks. Training used 31 HUP SEEG patients from OpenNeuro ds004100 version 1.1.3. Testing used 18 NIH and Johns Hopkins patients from ds003876 version 1.0.2. The two fixed logistic model rules used contact number and six spectral features. The rank rule replaced each spectral feature with its rank fraction within the eligible patient contact pool. The mean patient AUC was 0.7362 with ranks and 0.7452 with raw features. The paired change was -0.0090. The joint 95% patient bootstrap interval was [-0.0556, +0.0349]. It refit both models in 2,000 training draws and kept test site counts fixed. The conditional exact sign-flip p value was 0.496262 under independent patients and symmetric null differences. The planned gain rule failed. This does not show equivalence. This is a secondary analysis of previously studied test patients. The local plan preceded the new model fits. The package preserves spectra, source versions and retrieval recipes, code, scores, all patient draws and refit models, proofs, a figure, and repeat instructions. Three fresh public raw windows matched the saved spectra. An independent solver check passed. A separate process reproduced all numeric results exactly. The result concerns source marks at selected contacts. It does not establish a treatment decision.

stereoelectroencephalography · seizure onset zone · patient ranking · spectral features · logistic regression · external validation · bootstrap