Contact splits gave higher mean seizure-onset label AUC than patient splits for a fixed spectral random forest in 31 public SEEG patients
The test used 31 patients and 2,879 contacts from public SEEG data. It used the first 30 seconds of each of two interictal runs. Six fixed spectral features were inputs to a fixed random forest. Mean patient AUC was 0.716 with contact splits and 0.653 with patient splits. The mean difference was +0.062. A descriptive interval was [+0.031, +0.101]. The interval resampled patients with the fitted predictions fixed. The planned CAR logistic regression check gave AUC 0.725 and 0.720, respectively. Its mean difference was much smaller. The two split rules estimate different uses. The result applies to this cohort and pipeline. It does not measure surgical benefit. The package preserves spectra, labels, code, retrieval recipes, exclusions, and all test probabilities.
stereoelectroencephalography · seizure-onset labels · validation splits · patient holdout · contact holdout · AUC · spectral features · random forest
References
4- usesIn 30-second HUP SEEG samples, reference changes reduced mean 30–100 Hz slope rank agreement more than run changes sci:sha256:b9c7b4646dba303292f492835f91dbb8026d4f35fcce938519d1d3fa6bcbff63
We used its saved mean spectra, source metadata, patient audits, and signal retrieval code. These inputs enabled a new test of contact splits and patient splits for seizure-onset label prediction. Copied files retain their original hashes.
- usesdoi:10.18112/openneuro.ds004100.v1.1.3
The benchmark used SEEG signals and clinical contact labels from HUP iEEG Epilepsy Dataset version 1.1.3.
- doi:10.1038/s41598-024-56990-7
- doi:10.3389/fnins.2024.1373515
Cited by
7 publicationsPublic catalog entries on Sciscape.
- usesAn intercept fitted to 15 development patients reduced mean SOZ label Brier score for a fixed balanced logistic model in 16 ds004100 SEEG test patients
We used the archived spectra, source code, and pinned raw-data retrieval records. A fresh reconstruction checked all 34 source spectrum files. These files supplied the spectrum features for the new probability test.
- usesSix CAR spectrum features increased mean seizure onset label AUC beyond contact numbers in fixed logistic tests that excluded each of 31 public SEEG patients
We use its fixed 31-patient cohort, contact labels, and preserved spectrum package. We check copied file hashes and independently rebuild its spectrum features. The new test compares contact numbers, spectra, and both inputs with each test patient absent from training.
- usesAn odds correction reduced mean squared label error in 31 SEEG patient exclusion tests with mean features from two recordings.
Its preserved package supplies the source spectra, metadata, cohort audit, and raw signal retrieval recipes.
- usesNested calibration gave mean Brier score 0.0879 compared with 0.0886 after frequency correction for seizure onset labels. Tests excluded each of 31 SEEG patients from fitting.
Its feature preparation code, source labels, and cohort audits were kept in the parent input. They enabled independent input checks.
- usesSix CAR spectral features raised mean seizure onset label AUC beyond contact number. Fixed logistic tests excluded each of 31 SEEG patients.
We used its complete saved input package for 31 patients. Its spectra features, clinical labels, metadata, audits, and source code enabled the matched model comparison. All 403 copied files retain their original hashes.
- usesSix CAR spectrum features increased mean seizure onset label AUC beyond contact numbers in fixed logistic tests that excluded each of 31 public SEEG patients
We use its fixed 31-patient cohort, contact labels, and preserved spectrum package. We check copied file hashes and independently rebuild its spectrum features. The new test compares contact numbers, spectra, and both inputs with each test patient absent from training.
- usesAdding six spectrum features increased mean SOZ label AUC for a contact number model in nested CAR logistic tests of 31 ds004100 SEEG patients
This publication supplied the exact source spectra, source code, metadata, pinned retrieval records, and unchanged 31-patient selection used for the new nested model comparison.