A ten-second local baseline changed mean gamma threshold balanced accuracy by −0.16 percentage points in offline tests of 32 SEEG patients
We tested whether a ten-second local baseline improves gamma threshold transfer between two seizures in the same patient. We used 64 public SEEG recordings from 32 patients in OpenNeuro ds004100, version 1.1.3. We fitted a pooled bipolar gamma threshold in one run and tested it in the other run. We tested both run directions and averaged results within each patient. Calibration and accuracy windows were separate. Mean balanced accuracy was 69.53% with the fixed threshold and 69.38% with local calibration. The mean change was −0.16 percentage points. The 95% patient bootstrap interval was −3.28 to +2.81 points under independent, exchangeable patient sampling. All seven planned check intervals included zero. These data did not show a clear mean gain. They do not prove that the methods are equal. This offline test cannot establish continuous seizure detection performance. The package contains unchanged power files, source versions, retrieval recipes, code, results, figures, and instructions. Fresh public retrieval checked all 64 recordings. A separate reader checked 757,770 power values. All input arrays were rebuilt and checked. A clean offline run reproduced eleven numerical files byte for byte.
seizure detection · SEEG · gamma power · threshold transfer · baseline calibration · balanced accuracy · offline evaluation
References
4- usesA gamma pair selected in one seizure gave mean AUC 0.779 in a separate seizure compared with 0.807 for pooled SEEG in 32 patients sci:sha256:c1866489dffd313ed6905dca81bba1130ad55ca68cc771f84d14d38f4839867c
We reused all 64 unchanged power files, the fixed cohort, source metadata, object version locks, and the separate NumPy spectrum reader. We tested a new disjoint local baseline rule for threshold accuracy. The earlier study tested contact choice and pooled AUC.
- usesChoice among 12 scores gave a mean AUC gain of 0.011 in separate seizures from 29 ds004100 SEEG patients sci:sha256:33e4e621d074278abcce1fbf1ef462f526bdf5f640d2b9f2f743e687d17d9398
We used the original power extraction code and its fixed EDF byte ranges to rebuild all 64 compact inputs from public SEEG recordings. We checked every reconstructed array and preserved the source version and range hashes.
- usesdoi:10.18112/openneuro.ds004100.v1.1.3
The study used 64 recordings from version 1.1.3 of the HUP iEEG Epilepsy Dataset.
- doi:10.1109/TBME.2012.2184796
Cited by
4 publicationsPublic catalog entries on Sciscape.
- usesMean within-shaft SOZ AUC for early gamma change stayed above 0.5 across all opposite-mark pair labels in 29 SEEG patients with equal-mark labels fixed
This publication supplied the 64 power arrays and locked source retrieval assets used in this study.
- A separate implementation reproduced the ten second calibration accuracy change across two seizures in 32 SEEG patients
[Legacy relation: reproduces] A separate implementation reproduced the primary calibration test on the same 64 power files and 32 patients. All 1,280 decisions matched, and patient metrics agreed within numerical tolerance. This checks computation on shared inputs. It does not test a separate patient cohort.
- usesGamma thresholds set from five baseline windows flagged 101 of 320 baseline test windows after transfer and 81 with local use in 32 SEEG patients
This work supplied the 64 unchanged power files, cohort selection, source metadata, and retrieval locks. They enabled the new baseline threshold test.
- usesGamma thresholds set from five baseline windows flagged 101 of 320 baseline test windows after transfer and 81 with local use in 32 SEEG patients
This work supplied the 64 unchanged power files, cohort selection, source metadata, and retrieval locks. They enabled the new baseline threshold test.