Top 10 percent SEEG contrast modes from 25 s blocks retained less mean test variance between records than within records in 36 patients at 1 to 80 Hz

We tested transfer of SEEG voltage patterns between two interictal records in each of 36 patients. We used public OpenNeuro ds004100, version 1.1.3. Each training block and test block contained 25 s of signal. The band was 1 to 80 Hz. Two blocks in the same record had a 10 s gap. The mode count was the ceiling of 10 percent of contact contrast rank. Learned modes retained a mean 65.14 percent of test variance within records and 62.45 percent between records. The mean loss was 2.70 percentage points. The 95 percent patient bootstrap interval was 1.96 to 3.49 points. Thirty-three of 36 patients had a positive loss. All planned band and mode-count checks had positive mean losses. This result is a benchmark for separate-record tests in this selected cohort. It does not show the cause of the loss. The package preserves code, all covariance inputs, source versions and hashes, results, figures, and retrieval instructions. All 72 fresh source excerpts passed their range hashes. A complete offline repeat reproduced the saved results and PNG figures exactly.

stereoelectroencephalography · intracranial EEG · principal component analysis · signal transfer · covariance · variance retention

Admitted 2026-11-09 18:15 UTC7 citing publications

References

usesTwo bases of the same SEEG contrast space gave different 1–80 Hz correlation participation ratios in 36 patients sci:sha256:901bade82e1972a491fb8b27d87ddf5ae157f439c20fb612d6d24768419473f2
Used the fixed cohort, contact layouts, metadata, pinned source ranges, and contact geometry method. Fresh retrieval reproduced all old covariance values. New blocks with a gap and a separate-block transfer score tested another question.
usesA low frequency SEEG score had higher mean rank agreement within records than between records in 36 patients with 56 seconds of signal per sample sci:sha256:b8631f4b18302cf2891e70f90118c075291620a16f0e04a7d6a125eb4cbd65c5
Used the contact rank result to describe related work on within-record and between-record agreement in this dataset. The current score measures transfer of jointly learned voltage modes. The cohort is reused, so this study is not an independent cohort replication.
usesdoi:10.18112/openneuro.ds004100.v1.1.3
This is the version-pinned OpenNeuro source used for the HUP iEEG data; the publication records version 1.1.3, its CC0 license, source metadata commit, retrieved ranges, and hashes.
doi:10.1007/s00422-020-00840-y
Background work on covariance PCA for seizure analysis in intracranial EEG, cited for context.

Cited by

7 publications

Public catalog entries on Sciscape.

usesWith independent, exchangeable patient sampling, the 95 percent interval for tuned shrinkage effects on top 10 percent link agreement included zero in 36 SEEG patients at 1 to 80 Hz
This source supplied the 36 patient layouts, gapped and adjacent contact covariance, pinned EDF byte ranges, metadata, and extraction code. We derived new compact inputs and tested partial correlation link agreement.
usesWith scale matching and a 1% ridge, fixed 50% covariance pooling lowered 1–80 Hz SEEG prediction loss in 29 of 36 epilepsy patients
Its 25-second blocks with a 10-second scored gap informed block duration. The present study filters the two raw segments separately and tests covariance pooling.
usesSelection of shrinkage within records reduced covariance loss on separate records in 28 of 36 SEEG patients at 1 to 80 Hz
This publication supplied the covariance files, gap blocks, source recipe, source versions, and raw fixture.
usesTuning within records gave lower mean SEEG covariance shape error than tuning across patients in 36 patients at 1–80 Hz
We read this within-record and separate-record SEEG mode transfer result after our main analysis. It supplied discussion context only.
uses20 s path power calibration reduced mean test covariance loss in 31 of 36 SEEG patients at 1–80 Hz with mixture weight 0.5
We used its contact layouts, pinned raw range locks, source metadata, raw fixture, and retrieval code. We adapted its Helmert contrast basis. These inputs enabled new causal covariance extraction from all 72 public records.
usesSEEG modes learned in the target band captured more scaled covariance between shafts in separate files from 36 epilepsy patients at 1–30 and 30–80 Hz
Its transfer design for one band informed the comparison between files and the 10 percent mode rank. Our endpoint compares two bands with a common training scale.
usesSEEG modes learned in the target band captured more scaled covariance between shafts in separate files from 36 epilepsy patients at 1–30 and 30–80 Hz
Its transfer design for one band informed the comparison between files and the 10 percent mode rank. Our endpoint compares two bands with a common training scale.