Block variance shrinkage reduced mean squared error for 1–80 Hz bipolar SEEG correlations between shafts in separate files from 36 patients

We tested a training-block rule for SEEG correlation maps in another public file. We used two 25 second blocks per file from 36 epilepsy patients in OpenNeuro ds004100 version 1.1.3. The rule fits one scalar from block variation. It uses no target map values. The primary maps contain signed bipolar correlations between named electrode shafts at 1 to 80 Hz. Mean squared error fell by 6.87 percent relative to the raw map. The patient bootstrap 95 percent interval was 3.48 to 10.11 percent. The rule reduced error in 28 patients. The rule beat a zero map in 34 patients. The gain remained when no contact was reused between signals. A fixed weight of 0.5 increased observed mean error. The high-band interval included no improvement. The targets are noisy file estimates. The acquisition gap and any overlap are unknown. The result does not establish true neural coupling or clinical benefit. The package preserves code, input moments, public retrieval pins, all scores, maps, bootstrap samples, and reproduction instructions.

stereoelectroencephalography · SEEG · correlation maps · shrinkage · mean squared error · electrophysiology

Admitted 2026-11-09 19:49 UTC6 citing publications

References

usesBipolar 1–80 Hz SEEG correlation magnitude maps had positive correlations between file excerpts after signal, shaft pair, and log distance controls in 36 epilepsy patients sci:sha256:36170aff18fbecc7f57d9bf3c205be69a8ce1d45617c7e9600a74df38e629542
This publication made the disjoint-filter covariance inputs and range pins. We retained its 36-patient cohort and processing ranges.
usesCentered late high gamma map shrinkage under a common average reference gave no lower later mean map error than zero maps in nine SEEG patients sci:sha256:ace6f61a218bf1f037799231c321d34a9bc93400bac7e4d2a71b1e2721ed56a8
Its result for a different SEEG task motivated the explicit zero-map comparator. We used no data or scores from that report.
usesIn 36 patients, fixed 1–80 Hz SEEG contrasts showed larger covariance shape changes between files than within files sci:sha256:f663d71b5fe518a4b49eafc4b56264ce4f5dc80dad0e319618842dac64c731a5
We used the inherited EDF header parser, calibrated voltage retrieval code, and public cohort selection. The imported code keeps its license notices.
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 sci:sha256:c7d0d78b39c231d770b52d4a9725d271150234e9665d80a63666a928ca72967c
This publication supplied the 72 contact covariance archives, metadata, layouts, and retrieval recipe used for the new signed correlation map error test.
usesdoi:10.18112/openneuro.ds004100.v1.1.3
Source data: HUP iEEG Epilepsy Dataset, OpenNeuro ds004100 version 1.1.3, used for all 36 patients and the covariance inputs.
doi:10.1093/braincomms%2Ffcae165

Cited by

6 publications

Public catalog entries on Sciscape.

usesA variance rule reduced mean squared error against full copying for bipolar gamma residual maps between two seizure files in each of 28 SEEG patients
This publication supplied method context for block variance shrinkage of interictal correlation maps. We read it after our scores were calculated. We used none of its data or weights. The present test uses early seizure power residuals and a five-window jackknife.
uses30–80 Hz bipolar SEEG maps retained positive mean correlation across file excerpts in 36 epilepsy patients after linear 1–30 Hz map and geometry controls
Supplied the 72 contact covariance archives, public metadata, and source manifests used as the analysis inputs. All input and metadata hashes were checked.
usesOne scalar weight predicted separate-file SEEG correlation maps better than repaired link-specific weights in 36 patients
This source supplied all 72 disjoint-filter covariance inputs for 36 patients. Its scalar block variance rule informed the comparator. This study tested a separate weight for each link with matrix repair. The cohort was reused.
usesBlock products changed a mean 37 percent of top 20 percent SEEG links at 30 to 80 Hz in 36 patients
We read its scalar shrinkage study to define the new selection question. A positive scalar keeps ranks fixed. The new comparison tests a block product rule that changes ranks. This citation records design context and reuse of the same cohort; it does not claim independent replication.
usesMaps of 30–80 minus 1–30 Hz correlation strength reduced mean error after shaft name and distance controls in separate SEEG files from 36 epilepsy patients
This work supplied all 72 contact covariance inputs, their public raw retrieval recipe, the selected cohort, and the training weight rule. We applied that rule to new band contrast residuals.
usesAdding 1 to 30 Hz maps to a fixed ridge model reduced 30 to 80 Hz SEEG residual error by 3.8 percent from run 1 to run 2 in 36 epilepsy patients
We used its 72 unshrunk contact covariance archives, layouts, source metadata, and pinned EDF range recipe. We rebuilt every input from public ranges and checked exact array agreement.