Sharing people raised mean accuracy for balanced arbitrary labels by 25 percentage points across one month in 60 ds004148 v1.0.1 adults with a fixed posterior EEG model

This study tested whether a one-month EEG session gap removes person information. It used all 60 adults in OpenNeuro ds004148 v1.0.1 and a fixed posterior power model. Each fit used 48 training people. Three arbitrary classes had 20 people each. Across 4,999 balanced assignments, mean accuracy was 57.1% with shared people and 32.2% with separate people. The mean increase was 24.9 percentage points. The closest record came from the test person in 22 of 60 shared tests. The exact expected increase was 24.86 points under uniform balanced person labels. Independent test labels removed the excess mean score. All planned secondary settings and split repeats are included. The package preserves spectra, source hashes, code, fits, labels, results, and commands. All 99 analysis files matched in an offline repeat. The result applies to this fixed model and cohort. It does not test a clinical outcome or identify the neural source of person information.

electroencephalography · data leakage · person-level label permutation · nearest neighbor · repeated sessions · reproducibility

Admitted 2026-11-09 15:17 UTC2 citing publications

References

usesA fixed EEG model predicts shuffled labels when training and test sets share the 88 ds004504 subjects sci:sha256:7ba0bee1f809898d24fefdeb839b4d53689a26bcea0ff0429014af5f5ef0c2a9
The prior within-record arbitrary-label test informed the balanced person-label control and the question about separate visits. The present study used a different dataset and new model code.
usessci:sha256:eb6da8fdd843c3b66dd19edef3df5eb85f888448f106108e611352a1c6388159 not known to Sciscape
This study reused all 360 preserved resting EEG spectra, source metadata, raw hashes, and unchanged extraction code. They enabled the separate-visit person-label test and source replay.
usesdoi:10.18112/openneuro.ds004148.v1.0.1
This version-pinned dataset is the source of the 60 adults’ resting EEG records analyzed in this study.
doi:10.1038/s41597-022-01607-9
supportsdoi:10.3389/fnins.2024.1373515
This paper provides evidence about data leakage in EEG model tests, the specific background claim cited in the article.

Cited by

2 publications

Public catalog entries on Sciscape.

usesWindow splits reduced exact squared risk from 0.362 to 0.091 for fixed five-neighbor closed-eye EEG predictions in 49 ds003490 people with 25 arbitrary positive labels
This publication supplied prior method context for balanced arbitrary person-label controls with fixed EEG neighbor geometry. Its data and numerical results were not used in this benchmark. The new analysis uses squared loss and an exact fixed-count risk identity on ds003490.
supportsRandom labels balanced within folds give exact mean accuracy of 50% for separate people and 72.27% for shared people with a fixed EEG model across one month in 60 ds004148 adults
Both tests use five folds and 48 training people and find excess random person label accuracy across one month. The present test adds independent label balance within folds, equal training class counts, and a proof with separate probabilities for wrong-person matches within and outside the test fold. It uses the same cohort. This report was read after the primary analysis.