Epoch splits gave mean AUC 0.661 for random pair labels with a fixed SEEG power classifier in 35 patients
We tested three split rules on public SEEG power data from 35 patients in OpenNeuro ds004100 version 1.1.3. A fixed five-neighbor classifier used two bipolar power inputs from 1,887 pairs and two 60-second clips per patient. Mean patient SOZ AUC was 0.780 for epoch splits and 0.708 for patient splits. With 200 random pair-label permutations, mean AUC was 0.661 for epoch splits and 0.498 for patient splits. An exploratory removal of same-pair neighbors reduced the random-label mean to 0.494 on fixed epoch folds. The fixed logistic model showed little split difference. These results apply to the tested models and short clips. They do not establish clinical value. The package preserves source features, code, folds, predictions, random labels, provenance, and retrieval steps.
stereoelectroencephalography · epilepsy · data leakage · cross-validation · nearest neighbors · seizure onset zone · random-label control
References
2- usesReference choice changed relative 30–55 Hz power AUROC by 0.111 on average in two 60 s clips from 35 SEEG patients sci:sha256:d441c426f62a09b3b78933ca35f4c191a45f8076d63d922c1013d1d697a54791
We used all 70 final 60-second feature files from this publication. We kept pairs with odd lower contact numbers and tested three classifier split rules. We preserved the complete source package and its hashes. We independently rebuilt both clips for HUP060 from the pinned signal bytes.
- usesdoi:10.18112/openneuro.ds004100.v1.1.3
OpenNeuro ds004100 version 1.1.3 supplied the SEEG recordings, metadata, and source SOZ labels; the same version is recorded in the source and provenance files.
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.