With scale matching and a 1% ridge, fixed 50% covariance pooling lowered 1–80 Hz SEEG prediction loss in 29 of 36 epilepsy patients

We tested whether covariance from another record improves prediction after 25 seconds of local SEEG data. We used 72 public interictal records from 36 epilepsy patients in OpenNeuro ds004100 version 1.1.3. Each record gave two 25-second blocks. The blocks used separate raw filter segments and a 10-second scored gap. The primary band was 1-80 Hz. We matched the source covariance trace to the local trace, fixed the pooling weight at 50%, and used a 1% ridge in a fixed path contrast space. Positive gain means lower covariance loss. The gain was positive in 29 of 36 patients. The median gain was 0.017895 nats per contrast. The mean was 6.471721, with a 95% patient bootstrap interval of 0.017731 to 19.360826. One patient supplied 99.49% of the total gain. Without that patient, the mean was 0.033931. Fixed ridge checks changed the effect size greatly. The reverse block check gave no clear positive mean gain. All 72 raw range hashes matched. Equivalent contact bases gave the same primary gains. A full repeat matched all result tables and bootstrap values exactly. The package preserves 432 covariance matrices, code, a raw fixture, source versions and hashes, results, figures, and reproduction instructions. This follow-up uses an existing cohort and does not test clinical outcomes or an optimal pooling weight.

stereoelectroencephalography · SEEG · covariance prediction · covariance pooling · epilepsy · signal processing

Admitted 2026-11-09 18:43 UTC9 citing publications

References

usesOne patient reversed the mean gain of a 0.5 covariance mixture over an independent contact model in paired 1–80 Hz SEEG records from 36 patients sci:sha256:73e7b108e6d3e3788a07ac04de661b4eedb0dfdab51624d480fc0d9bce16a2b3
Its transfer result motivated a new comparison after a local training block is available. No score values from that study entered this analysis.
usesIn 36 patients, fixed 1–80 Hz SEEG contrasts showed larger covariance shape changes between files than within files sci:sha256:f663d71b5fe518a4b49eafc4b56264ce4f5dc80dad0e319618842dac64c731a5
This study supplied the fixed 36-patient cohort, contact layouts, metadata, source licenses, EDF parser, and pinned raw byte recipes. We rebuilt new covariances from separate raw filter segments. Its covariance shape result motivated the pooling question.
usesTop 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 sci:sha256:fa333c76836bc1e1675dff212cda0cb49368aaa0a667e6814500a28aef64c135
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.
usesdoi:10.18112/openneuro.ds004100.v1.1.3
Source dataset for the 72 interictal records; exact version 1.1.3 and CC0 license are recorded in the source metadata.
doi:10.1093/braincomms%2Ffcae165

Cited by

9 publications

Public catalog entries on Sciscape.

usesTraining selection of covariance between shafts lowered median loss but raised mean loss in 36 ds004100 SEEG patients at 1–80 Hz with 1% ridge
This publication supplied the 72 pinned contact covariance inputs, source metadata, disjoint filter intervals, raw fixture, and extraction code. We checked 306 imported file hashes.
usesFitting on other patients chose zero added distance shrinkage for later 1–80 Hz SEEG covariance shape at a 10% ridge in 36 patients
This study uses the fixed 36-patient cohort, all 72 contact covariance files, source metadata, raw byte recipes, and unchanged source tools from this report. We rebuilt all files from the pinned public bytes. The new analysis tests a supplied-coordinate distance prior.
usesA prediction energy limit reduced mean SEEG reconstruction error across files but increased error in 18 of 36 patients
We adapted the preserved disjoint-interval covariance rebuild code and retained its source notices. We rebuilt all 72 first-cohort inputs from pinned public EDF ranges.
usesFixed 1–80 Hz SEEG mode prediction gave a 3.0 percent mean error gain with an interval that included zero in 36 epilepsy patients
We used its covariance pooling test as related prior work. It predicted a covariance matrix after a local block. Our test predicts the difference between two files. We do not claim a new independent cohort. Its numeric scores did not enter this analysis.
usesEqual-energy pooling gave a larger mean gain over the harmonic local score between files than within files in 36 patients with 1–80 Hz SEEG
This work supplied separate-filter covariance inputs and extraction code for an added check. We rebuilt its derivatives from public raw data.
usesA contact metric preserved 500 ms ridge lag dependence across equivalent contrast bases in 18 SEEG records from nine adults
Its coordinate metric for ridge covariance prediction informed the metric correction for regularized lag canonical correlation. None of its data arrays or numerical results enter this analysis.
usesWith shaft covariance fixed, 50% cross-shaft transfer and 10% ridge lowered 1–80 Hz SEEG loss in 34 of 36 patients
We used all 72 disjoint-filter covariance files, source metadata, a raw fixture, and extraction code. We checked their hashes and rebuilt every array from the public raw prefixes.
usesAfter signal and shaft pair controls, 1–80 Hz bipolar SEEG correlation magnitude between shafts had negative fitted coordinate distance slopes in all 36 tested epilepsy patients
The study uses the contact covariance arrays, source metadata, pinned raw retrieval recipe, and raw fixture.
usesAcross files, all other shafts increased mean 1–80 Hz SEEG reconstruction error compared with the nearest shaft in 36 patients
Used its covariance arrays, contact layouts, raw fixture, and metadata for the new reconstruction test.