Complete test block normalization raised mean fixed linear 500 ms SEEG forecast skill by 0.174 percentage points in nine adults

We tested fixed univariate linear 500 ms forecasts on 513 SEEG contact pairs from nine adults. The public data are from OpenNeuro ds003688 version 1.0.7. Complete test block normalization raised mean voltage forecast skill by 0.174 percentage points. The adult bootstrap 95% interval was 0.097 to 0.263 points. An exact proof showed that the transform changes only the raw forecast intercept when coefficients stay fixed. An earlier ten-second calibration block reduced mean skill by 1.193 points. The same-record skill advantage remained near two points with frozen or complete test normalization. All eighteen compact inputs matched fresh native source reconstructions. A second analysis reproduced all numeric values exactly. The package preserves the proof, code, inputs, source hashes, results, and rebuild instructions. The sample is reused from two cited studies. The result concerns short processed SEEG records. It does not show clinical prediction value or real-time forecast performance.

SEEG · signal forecasting · block normalization · forecast skill · intercept · intracranial EEG · normalization · time series

Admitted 2026-11-09 17:50 UTC2 citing publications

References

usesFixed ECG models had negative mean test skill in nine adults' rest and film SEEG at 1 to 20 Hz sci:sha256:5b2695e903fadfbbd714f4198bb6739bcc8022b682617c1515dff94ebb286b00
This publication supplied the paired 124-second SEEG inputs for nine adults, the inherited participant and contact selection, immutable raw range provenance, and the source rebuild code. This audit uses only SEEG and excludes ECG from every model.
Fixed linear models gave lower mean normalized 500 ms SEEG prediction error within records than between paired rest and film records in nine adults sci:sha256:158bc9a242c61384bf400795feb8251c594f7805a112626ccbcc847fe76ae4b4
[Legacy relation: reproduces] The complete test normalization policy reproduces the earlier primary mean and each adult same-record minus other-record skill gap to numerical precision. This is a code and score check on reused data, not an independent biological sample.
usesFixed linear models gave lower mean normalized 500 ms SEEG prediction error within records than between paired rest and film records in nine adults sci:sha256:158bc9a242c61384bf400795feb8251c594f7805a112626ccbcc847fe76ae4b4
This publication supplied the fixed univariate temporal model, train and test blocks, and penalty settings. We held its fitted coefficients fixed while testing normalization policies and an exact intercept loss decomposition.
usesdoi:10.18112/openneuro.ds003688.v1.0.7
The analysis used paired rest and film SEEG records from OpenNeuro ds003688 version 1.0.7; the archived provenance records pinned source objects and byte ranges.
doi:10.1038/s41597-022-01173-0

Cited by

2 publications

Public catalog entries on Sciscape.

usesMean theta log power autocorrelation at 0.5 to 2 seconds exceeded phase controls in nine adults' rest SEEG
We reused 18 compact SEEG input arrays, source metadata, selection rules, provenance, and raw reconstruction code. These files enabled a new theta power correlation analysis. We did not reuse forecast results.
usesA fixed word timing model gave -0.071% mean high gamma prediction gain in later film blocks from nine selected adults with SEEG
This publication gave the nine adult and good SEEG contact lists, pinned native object URLs, and source metadata. The present study retrieves new native film data and calculates new high gamma power. It does not reuse the earlier low-rate voltage arrays or forecast results.