Contact splits gave higher mean pooled AUC for seizure onset labels than shaft splits with a fixed spectral random forest in 34 SEEG patients

We tested how test splits change prediction of clinical seizure onset labels. We used public SEEG spectra from OpenNeuro ds004100, version 1.1.3. A fixed random forest used seven band power features in 34 patients. Contact and shaft splits had the same training counts for each patient. Mean pooled AUC was 0.766 for contact splits and 0.724 for shaft splits. The paired difference was 0.042, with a 95 percent patient bootstrap interval of 0.018 to 0.069. A within-fold check gave a smaller difference, with an interval that included zero. Random labels assigned to shafts gave mean pooled AUC of 0.638 with contact splits and 0.457 with shaft splits. The package contains inputs, code, predictions, source hashes, and reproduction instructions. These results apply to this model and sample. They do not establish clinical benefit.

stereoelectroencephalography · SEEG · seizure onset zone · AUC · data leakage · test splits · random forest · electrode shafts

Admitted 2026-11-09 16:10 UTC4 citing publications

References

usesMedian 80–150 Hz power rank agreement was lower between references than between recordings on matched pairs in 35 SEEG participants sci:sha256:eaad9a72db87113b65f5fda15f4c3043dc3463ac2f155c6b9d8b01404ec3ab9f
This study uses the archived common average contact spectra, signal extractor, source metadata, checked range receipts, and integer example. It tests contact and shaft prediction splits with these inputs.
usesdoi:10.18112/openneuro.ds004100.v1.1.3
The study uses the HUP iEEG Epilepsy Dataset, version 1.1.3, as the source of SEEG recordings and clinical labels.
doi:10.1038/s41598-025-02679-4
doi:10.3389/fnins.2024.1373515

Cited by

4 publications

Public catalog entries on Sciscape.

usesA fixed SEEG classifier exceeded shaft label controls when each of 34 test patients was excluded from training
We reused its seven-band contact table and cohort exclusions through the parent package. We retained the 34 patients, 4,031 numbered good contacts, and 350 clinical SOZ labels per recording. The table was rebuilt from pinned spectra and channel metadata.
usesA contact number rule gave higher mean concordance than a fixed power classifier at 62 SEEG label boundaries in 16 test patients
The input contact features and consistent label cohort were inherited from this publication through the fixed classifier package. This study keeps that feature and label selection and makes a new patient split.
usesA fixed SEEG power classifier gave higher AUC than shaft label controls in 34 patients with two recordings each
We used its contact feature table, cohort exclusions, and matched contact and whole-shaft split rule. The new classifier uses structured label controls.
usesNearest training labels gave mean AUC 0.934 within test folds for 34 SEEG patients
This study uses the fixed 34-patient sample, source feature table, contact and whole-prefix masks, saved signal comparisons, and model code. It adds training-label baselines and a count-preserving label placement test.