A fixed SEEG power classifier gave higher AUC than shaft label controls in 34 patients with two recordings each

We tested a fixed SEEG power classifier in 34 patients with two recordings each from OpenNeuro ds004100, version 1.1.3. The classifier used five nearest training contacts and seven log power values. We changed both training and test labels in each of 9,999 draws per control. With contact splits, clinical AUC was 0.752 and 0.742; mean shaft-shuffle AUC was 0.623 and 0.609. The primary difference was 0.1286, with a 95 percent patient bootstrap interval of 0.0939 to 0.1615. Its upper-tail Monte Carlo p value was 0.0001 under the stated within-shaft label allocation rule. Clinical scores also exceeded circular shaft shifts and the tested controls with whole-shaft splits. We supply an exact shuffled Brier mean for the fixed neighbor graph. Fresh signal ranges reproduced all 70 source spectra, and a complete local rerun reproduced all model maps and control distributions exactly. The package contains code, source data, provenance, results, and reproduction instructions. This result concerns clinical contact labels within existing implants. New-patient performance and biological causes were not tested.

stereoelectroencephalography · SEEG · nearest neighbors · structured label controls · area under the curve · AUC · epilepsy

Admitted 2026-11-09 17:49 UTC2 citing publications

References

usesContact splits gave higher mean pooled AUC for seizure onset labels than shaft splits with a fixed spectral random forest in 34 SEEG patients sci:sha256:56a9020d2fcfaf8dce31031e250ead4147aed4b96d5049afb0ef063a225a356e
We used its contact feature table, cohort exclusions, and matched contact and whole-shaft split rule. The new classifier uses structured label controls.
usesMedian 80–150 Hz power rank agreement was lower between references than between recordings on matched pairs in 35 SEEG participants sci:sha256:eaad9a72db87113b65f5fda15f4c3043dc3463ac2f155c6b9d8b01404ec3ab9f
We used its pinned source spectra, metadata, signal reader, extraction code, and range receipts. Fresh downloads reproduced all CAR spectral values.
usesNearest training labels gave mean AUC 0.934 within test folds for 34 SEEG patients sci:sha256:d2376a1280b1bebf9a19a2c63212afe1b0bb56c0d9ce40980d8c935535a09c3e
We retrieved its preserved source package and used its within-fold pair comparison rule to avoid score offsets between test folds.
usesLabel shuffles within contact name groups give expected AUC of 0.621 for slow power in two 30-second SEEG clips from 25 patients sci:sha256:3a78ea4c592440a463cec3a1481d8278e4a3e322db8c5a99928f9d9bbf87ae06
Its fixed-feature shaft label controls informed the question. We extend those controls to a trained model by changing training and test labels together.
usesdoi:10.18112/openneuro.ds004100.v1.1.3
Source data: HUP iEEG Epilepsy Dataset, OpenNeuro ds004100 version 1.1.3, as pinned and described in the publication.

Cited by

2 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 used its verified source package, fixed five-neighbor rule, seven band features, cohort, and joint training and test label controls. We replaced contact and whole-shaft splits with exclusion of every contact from each test patient. We added label-free patient median subtraction and a prevalence loss baseline.
usesMedian log band power subtraction changed mean AUC by -0.0012 in a fixed test with five neighbors on 14 excluded SEEG patients
This publication supplied the contact features, spectra, channel metadata, source code, and retrieval records. We used these inputs for the patient exclusion test.