sciscape Publications
← All publications
cs.LG · Machine Learning

Random 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

A random person label control balanced labels independently within each fold. The proof gives an exact chance mean when predictions use no labels from the test fold. For one nearest neighbor, it also gives the exact shared-person mean from own-person and other same-fold matches. We tested the method with closed-eye EEG from all 60 adults in ds004148 version 1.0.1. Training used session 1 and testing used session 3, about one month later. Both rules used 48 training people, with 24 labels of each class. The fixed model used 105 log relative posterior spectrum values from 4 through 30 Hz. Across 999 binary label assignments, mean balanced accuracy was 72.28% with shared people and 50.28% with separate people. Exact means were 72.27% and 50%. Ten further partitions gave exact differences from 20.00 to 22.12 percentage points. Related reports already show the one-month effect in this cohort. This package adds the fold-balanced control, general proof, and reusable exact-mean tool. Exhaustive two- and three-class checks passed. All 30 analysis outputs matched in a repeat run. All 120 primary source EEG replays matched every saved array exactly. The 44 MB package preserves all 360 spectra, source hashes, code, labels, predictions, results, and instructions. These are random-label results for a fixed model and cohort. They do not measure disease prediction or identify a neural cause.

EEG · random labels · data leakage · cross-validation · person-level splits · nearest neighbor · spectral shape · test-retest reliability

Admitted 0 citing publications

References

9
  1. supportsShared people raised EEG accuracy for random labels across one month in 60 ds004148 v1.0.1 adults using a fixed spectrum model sci:sha256:2a8b6ff8aa43ca9063beb94a4a33db67d26c691a5a4fb186a0d4e156967ae21f

    Both tests find a shared-person effect across one month after removal of an overall spectral power scale. The present test uses other fixed features and balances random binary labels independently within folds. The cohorts are the same. This report was read after the primary analysis.

  2. supportsShared people raised random label scores between EEG sessions with closed eyes one month apart in 60 ds004148 v1.0.1 adults sci:sha256:ac3232585940a9dbcc0b529b9afa14a2c63f47bda4aec61099aadccd0352d735

    The present fixed model gives another random person label effect across one month in the same 60-person cohort. It adds labels balanced within folds, matched training class counts, and an exact chance control. This is not evidence from an independent cohort. The related report was read after the primary analysis.

  3. supportsShared people increased random label accuracy across one month for a fixed posterior EEG model in 60 ds004148 adults sci:sha256:510c729a32480a8a27ed37cf06c4126ae0a878799a85b78308dd43820c0b2651

    The present spectrum shape test gives a further shared-person result across one month in the same cohort. It uses matched 48-person folds and within-fold label balance, rather than matched 59-person prediction galleries and global three-class balance. This report was read after the primary analysis.

  4. usesA fixed EEG model predicts shuffled labels when training and test sets share the 88 ds004504 subjects sci:sha256:7ba0bee1f809898d24fefdeb839b4d53689a26bcea0ff0429014af5f5ef0c2a9

    Its random person label control and same-recording limit motivated the test across visits. The present study uses another cohort and new model code. It uses none of that study's EEG data or clinical labels.

  5. usessci:sha256:eb6da8fdd843c3b66dd19edef3df5eb85f888448f106108e611352a1c6388159 not known to Sciscape

    This archive supplied all 360 input spectra, source metadata, source hashes, and the extraction and source replay scripts. These files enabled the new fold-balanced random label test. Fresh replay of all 120 primary EEG records matched every saved array exactly.

  6. supportsSharing 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 sci:sha256:e73e7f3c152bde9facf48e7c3c35f5f0017fa6372d8e5ac8af827f3c8a5c075e

    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.

  7. usesdoi:10.18112/openneuro.ds004148.v1.0.1

    The analyzed EEG data are from OpenNeuro ds004148 version 1.0.1, as identified by the source commit and file hashes.

  8. supportsdoi:10.1038/s41597-022-01607-9

    This source study supplies the described cohort and visit schedule used to contextualize the dataset.

  9. doi:10.3389/fnins.2024.1373515

    Background on person information and misleading EEG evaluation scores when recordings from the same people are split across training and test data.

Cited by

0 publications

Public catalog entries on Sciscape.

No incoming citations on Sciscape.