A fixed EEG model predicts shuffled labels when training and test sets share the 88 ds004504 subjects
This study tested whether shared subjects permit high EEG test scores for shuffled labels. The data were 88 recordings from OpenNeuro ds004504, version 1.0.9. The fixed model used three nearest neighbors and 95 relative band power features. Each person supplied 20 non-overlapping epochs. Each shuffled label stayed constant across one person's epochs. Across 999 assignments, mean participant balanced accuracy was 93.4% when training and test sets shared people. It was 33.1% when the sets kept people separate. Training on early epochs and testing on late epochs from the same people gave 69.6%. The median gap between these blocks was 570 seconds in the supplied derivative. For diagnosis labels, scores were 96.2% with shared people and 52.0% with separate people. The package preserves features, source hashes, source metadata, code, predictions, and reproduction instructions. A fresh source replay matched stored features. A second analysis matched all seven output files byte for byte. The result applies to this cohort, feature rule, and fixed model. One recording per person prevents separation of person traits from session and preprocessing effects. The study measures validation behavior. It does not establish clinical diagnostic accuracy.
electroencephalography · EEG · subject leakage · data leakage · validation · label permutation · nearest neighbors · balanced accuracy · OpenNeuro · ds004504
References
6- usesdoi:10.18112/openneuro.ds004504.v1.0.9
The study used the supplied derivatives and group labels from OpenNeuro ds004504 version 1.0.9; its pinned snapshot commit and source file hashes are preserved.
- supportsdoi:10.3390/data8060095
This paper describes the source dataset and preprocessing already present in the supplied derivatives.
- doi:10.1109/ACCESS.2023.3294618
- doi:10.1007/s11571-026-10464-w
- doi:10.3389/fnins.2024.1373515
- doi:10.1038/s41746-019-0178-x
Cited by
60 publicationsPublic catalog entries on Sciscape.
- usesTesting one score after selection from 171 EEG scores gave 19.5% false detections at p <= 0.05 under random person labels in 59 ds004504 v1.0.9 records
The source retrieval and MATLAB loading code inherits work from this publication. The package preserves the source versions, hashes, and license notice.
- usesA fixed EEG model had mean random-label balanced accuracy near 33% across rest and light stimulation in 72 shared people
This study uses the archived rest EEG features, feature code, source hashes, and extraction audit. It adds paired light-stimulation records and matched training rules.
- usesShared people raised random EEG label accuracy across one month in 60 ds004148 v1.0.1 adults using 48 training records
Its same-record EEG random person label test supplied the question and control idea. We tested the route across separate sessions one month apart in another cohort. No data or code from ds004504 were used.
- 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 inherited retrieval code adapted download, MATLAB loading, and boundary handling from this publication. Its license notice is preserved. This study does not use its diagnosis classifier.
- usesEqual Fp1 noise powers change posterior theta-to-alpha ranks after mean reference in 88 ds004504 v1.0.9 records under zero cross-power
The source replay uses its inherited download, MATLAB read, and boundary rules through publication 1c67ee52-47af-4f72-8bcd-63907cc1745b. Both upstream license notices are preserved.
- usesA change from average reference to Cz gives a median alpha Riemannian distance of 11.29 in 88 ds004504 v1.0.9 records with a ridge fraction of 0.001 in all 19 channels
The inherited source retrieval and EEGLAB boundary code was adapted from this publication. Its license and publication record are retained. Its prediction endpoint was not used.
- usesOmission of up to three other reference channels reversed F3/F4 alpha signs in 53 of 88 fixed ds004504 v1.0.9 records
The preserved extraction code uses retrieval and EEGLAB boundary handling from this work through the matrix parent. The package retains the source code and license notices.
- usesIndependent choices of 100 selected EEG seconds gave posterior theta-to-alpha AUC limits of 0.637 and 0.964 for 36 Alzheimer and 29 control ds004504 v1.0.9 records with common average reference
The reused source retrieval and boundary code adapts this publication. Its license notice is preserved. The present study did not use its classifier results.
- usesScalp theta/alpha added to age, gender, and dementia type increased MMSE test error in 59 ds004504 v1.0.9 records
We reused the complete EEG feature table, participant metadata, source manifest, and extraction code. We verified all 223 upstream file hashes. These inputs enabled a new person-level MMSE prediction test and a source extraction replay.
- usesSharing 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
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.