Real zero-sum EEG filters fitted at 10 Hz on the first 100 accepted seconds gave lower mean imaginary coherency on the last 100 accepted seconds in 88 ds004504 v1.0.9 records

We tested whether fitted EEG spatial filters retained a 10 Hz imaginary coherency score in later data. We used all 88 processed records in ds004504 v1.0.9. Real filters had zero-sum weights within fixed anterior and posterior sensor groups. We fitted the first 100 accepted seconds and scored the last 100 with frozen weights and sign. The mean fitted score was 0.431. The mean signed test score was 0.164. The mean paired decrease was 0.267 (95% person bootstrap interval, 0.238 to 0.295). The absolute test score fell in 87 of 88 records. Seven fixed secondary tests also gave lower mean frozen scores. A stationary zero-lag Gaussian control gave a mean fitted score of 0.303 and a signed test mean near zero. These results show limited transfer under the tested settings. They do not separate sampling noise, time changes, and artifacts. The package includes the full Fourier input, source hashes, filters, control, code, checks, and reproduction instructions.

imaginary coherency · EEG · spatial filters · filter transfer · held-out evaluation · signal processing

Admitted 2026-11-09 17:55 UTC9 citing publications

References

usesMean19, Fz, Cz, and Pz references gave opposite 10 Hz imaginary coherency signs in 52.1% of tested pairs in 88 ds004504 v1.0.9 records sci:sha256:094b90e4b977e5284645863ffa4e5a4798181fe7bf966f267482939912866202
This publication supplied the full complex Fourier archive for 88 processed ds004504 v1.0.9 records, extraction code, pinned source manifest, metadata, and one source record. We used those files for a new test of fitted filter transfer and checked three source records again.
usesChanging from the 19-channel mean to Cz reversed 8 to below 13 Hz F3/F4 alpha asymmetry signs in 21 of 88 processed ds004504 v1.0.9 records sci:sha256:b7482b02bde6c5afdad997d5679df96ebfb7aa2e28494ae337324618fa74632e
Its pinned source manifest and retrieval method were inherited through the Fourier archive publication. They enabled source identity checks and the full-cohort retrieval recipe.
usesdoi:10.18112/openneuro.ds004504.v1.0.9
Processed eyes-closed EEG records from this exact version supplied the study data.
doi:10.3390/data8060095
Dataset descriptor for the OpenNeuro EEG cohort used here.
supportsdoi:10.1016/j.neuroimage.2011.11.084
Introduces the maximized imaginary coherency method used for filter fitting.
doi:10.1109/ACCESS.2023.3294618
The dataset source authors identify this as the first study on the dataset and request citation.
doi:10.1007/s11571-026-10464-w
The dataset source authors request this complementary benchmark citation.

Cited by

9 publications

Public catalog entries on Sciscape.

usesThe mean 10 Hz cross-block score was 40.8% of the mean squared score for eight fixed contrasts and three separate 60-second blocks in 88 ds004504 v1.0.9 EEG records
We reused its full Fourier input for all 88 records, extraction audit, source metadata, and extraction and retrieval code. Its fixed anterior and posterior channel groups supplied the amended contrast groups. We preserved its license notice. We did not reuse its fitted filters.
usesTemporal halves had greater mean squared disagreement in fixed alpha EEG lag matrices than balanced random halves for the first 200 accepted seconds of 88 processed ds004504 v1.0.9 records
Its early and late filter comparison motivated a time-order audit of a fixed whole matrix in the same source group.
usesTransport of ridge penalties kept MIC fits and frozen-filter imaginary coherency equal across two equivalent zero-sum bases in 88 ds004504 EEG records
This publication supplied the Fourier archive for all 88 records, fixed source metadata and hashes, extraction scripts, and the two-group fit and separate-window scoring design. This audit applies stated ridge penalties to those same fixed contrast spaces. It adds no independent cohort.
usesThe mean 10 Hz cross-block score was 40.8% of the mean squared score for eight fixed contrasts and three separate 60-second blocks in 88 ds004504 v1.0.9 EEG records
We reused its full Fourier input for all 88 records, extraction audit, source metadata, and extraction and retrieval code. Its fixed anterior and posterior channel groups supplied the amended contrast groups. We preserved its license notice. We did not reuse its fitted filters.
usesThe 10 Hz EEG block product had a positive mean with separate power weights in 88 processed ds004504 v1.0.9 records
Its decrease in imaginary coherency after fixed spatial filters were transferred to later data motivated the separate-power signed block product. The new score uses all entries of fixed sensor contrast spaces instead of an optimized component.
usesA sensor metric kept regularized GIM fixed across 20 equivalent EEG bases in 88 ds004504 v1.0.9 records
This publication supplied the original 88-person Fourier archive through the cited triangle-study package. Its fixed Fourier extraction made the new basis audit possible. This is the same cohort, not new independent cohort evidence.
usesRaw triangle sums did not change after common reference subtraction in 88 ds004504 EEG records
This study uses its unchanged Fourier archive for all 88 records, accepted window starts, source metadata, pinned source manifest, extraction receipts, and MIT extraction code. These inputs permit the fixed triangle and normalization audit.
usesA sensor metric kept regularized GIM fixed across 20 equivalent EEG bases in 88 ds004504 v1.0.9 records
This publication supplied the original 88-person Fourier archive through the cited triangle-study package. Its fixed Fourier extraction made the new basis audit possible. This is the same cohort, not new independent cohort evidence.
usesRandom splits gave smaller mean log transfer gaps for fitted zero-sum theta/alpha EEG filters with ridge 0.1 in 88 ds004504 v1.0.9 records
The early-to-late transfer test of fitted imaginary-coherency filters informed the need to test fitted filters on separate time blocks. This study uses a different spectral score and adds a random split comparison with all fitting repeated.