Five spectral bands increased mean AUROC for SOZ labels beyond contact numbers in patient holdout tests of 30-second SEEG from 31 patients
We tested whether five spectral bands added clinical seizure onset zone label ranking information beyond contact numbers. We used 30-second samples from two recordings for 31 patients in public OpenNeuro ds004100 version 1.1.3. The test included 2,879 contacts and 271 positive labels. Each outer test patient was excluded from scaling, model fitting, and five-fold inner patient penalty selection. Mean patient AUROC was 0.680 for a flexible model of contact numbers, 0.713 for spectra, and 0.757 for both. The fixed minus-number score had mean AUROC 0.693. The primary observed mean gain over the learned number model was 0.077. Its descriptive interval was [0.037, 0.119]. The gain was positive in 23 of 31 patients. The mean gain within the same name group was 0.040 in 28 patients. We report all checks for reversed runs, a local voltage reference, spectral shape, and two omitted name checks. Intervals condition on fitted scores and do not establish population confidence coverage. This one-hospital test measures clinical labels. It does not establish treatment benefit. The package contains source spectra, pinned retrieval records, code, splits, tuning scores, predictions, fitted models, and repeat checks.
stereoelectroencephalography · seizure onset zone · intracranial EEG · patient holdout · AUROC · spectral power · contact number · epilepsy · machine learning
References
7- usesLower contact numbers gave higher mean seizure onset label AUC than higher gamma fractions in two 60-second SEEG samples from 36 patients sci:sha256:7f57f55fb0912992a415d07349886a14b5bb293e801fb2771f8de060871ebcd1
The earlier result motivated the contact-number comparator and the check that omitted HUP142 and HUP166. Its 60-second features were not model inputs. The new test used 30-second source spectra with patient-separated fitting and penalty selection.
- usesIn 30-second HUP SEEG samples, reference changes reduced mean 30–100 Hz slope rank agreement more than run changes sci:sha256:b9c7b4646dba303292f492835f91dbb8026d4f35fcce938519d1d3fa6bcbff63
This investigation used the raw-derived power spectra, quality rules, source code, and pinned signal retrieval records in the preserved source archive. Its EDF code supported fresh raw-data and header checks.
- usesRepeated contacts gave higher mean AUROC than unseen patients for a fixed five-neighbor rule on 30-second SEEG spectra from 31 patients sci:sha256:db3385cbd16f5d51ca770d91d015d9d8cdf78e23094788f9c0bcde053342059d
This investigation used the exact 31-patient cohort, clinical labels, source spectra, and archived metadata. The source archive is preserved without changes. They enabled the patient holdout comparison of spectral and contact-number models.
- usesdoi:10.18112/openneuro.ds004100.v1.1.3
Source dataset and clinical labels for the analysis; the publication identifies OpenNeuro ds004100 snapshot 1.1.3 and its fixed metadata commit.
- supportsdoi:10.1093/brain%2Fawad007
Source dataset paper cited for the clinical intracranial recordings and limits of implant sampling.
- supportsdoi:10.1093/gigascience%2Fgix019
Cited for evaluating a clinical model with patients separated to match use in a new patient.
- doi:10.1038/s41598-025-02679-4
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesFixed combined models had mean SOZ AUROC 0.756 under matched and changed test references in 31 SEEG patients, with 79 percent top contact list overlap
This test used the unchanged feature table, patient cohort, frozen coefficients, scaling, penalties, nested splits, and matched scores. These inputs enabled paired reference tests without new fitting.
- usesAdding five spectrum bands did not establish a SOZ mark AUC gain over contact numbers within contact name groups in 30-second SEEG from 18 external patients after HUP training
This method supplied the fixed 31-patient HUP cohort, 2879 source feature rows, five spectrum bands, and contact number control.