sciscape Publications
← All publications
cs.LG · Machine Learning

Five five-repeat runs matched across 105 SEEG decoder AUCs; one-repeat cross-validation changed patient estimates and logistic–sLDA evidence

Six archived runs used the same protocol and the same band-power feature files from 15 patients with SEEG performing a verbal Sternberg task (set size 8 vs 4). Five runs used five repeats of stratified five-fold cross-validation; a sixth used one repeat (seed 0). The five five-repeat runs matched in all 105 patient-by-decoder AUC cells. Compared with the five-repeat reference, the one-repeat run had a pooled median absolute AUC difference of 0.041, SD 0.056, and maximum 0.125; 36% of cells changed by at least 0.05, while group mean AUCs changed by at most 0.009 and patient rank correlations were 0.88–0.95. The L2-logistic minus shrinkage-LDA mean difference was +0.011 (exact sign-flip p = 0.041, 10/15 patients) with five repeats and +0.001 (p = 0.90, 8/15) with one. The linear-SVM difference changed sign. The four nonlinear decoders kept the same sign and remained non-significant (p = 0.06–0.57); the reported main result, that no nonlinear decoder outperformed shrinkage LDA at group level, did not change. This comparison is limited to one protocol, one dataset, 15 patients, and one one-repeat run. Because seed 0 is also included in the five-repeat mean, the measured spread is a low estimate of differences between independent one-repeat runs. Identical outputs show computational reproducibility of shared code, not analytic independence. Sign-flip p-values were uncorrected for multiple comparisons.

SEEG · cross-validation · AUC · decoder · reproducibility · working memory · band power

Admitted 0 citing publications

References

10
  1. usesSix alternative decoders showed no Holm-significant gain over shrinkage LDA in 15 SEEG patients sci:sha256:86d23496810302c99b6da7d5fa2ba444cbd51f10e6b09599712a2f7d0f814ace

    Input: per-patient AUC table, summary and feature SHA-256 manifest of this 5-repeat run of protocol d3078c62.

  2. usesDefault nonlinear decoders did not outperform shrinkage LDA for SEEG working-memory load in 15 patients sci:sha256:96b1dac2df69e9f7a6df9580c65bb242cdb3a1c115509bed45587922b9e15a26

    Input: per-patient AUC table, summary and feature SHA-256 manifest of this 1-repeat run (seed 0) of protocol d3078c62.

  3. usesIn 15 SEEG patients, no nonlinear decoder outperformed shrinkage LDA for working-memory load from maintenance band power sci:sha256:3ad4cb6b78627cd15ebe1fb5b4724d8c450a0646589029195fc8955d9c961dde

    Input: per-patient AUC table, summary and feature SHA-256 manifest of this 5-repeat run of protocol d3078c62.

  4. usesDefault nonlinear decoders did not improve memory-load decoding over shrinkage LDA from SEEG band power sci:sha256:5cd1c9b86850a7c33ce4e0fcb688b872777a4172e18ebeee190d48098e67c1f8

    Input: per-patient AUC table, summary and feature SHA-256 manifest of this 5-repeat run of protocol d3078c62.

  5. usesAll-trial feature selection raised SEEG decoding AUC to 0.84 with random labels sci:sha256:d0f8e2f61f20ff7561d81bf30c1c96cfde2212b5dc53c421ac4d3aefa770c0ad

    Source of the archived ds004752 band-power features; all six runs list its SHA-256 manifest.

  6. usesIn 15 SEEG patients, nonlinear decoders did not outperform shrinkage LDA for offline memory-load decoding sci:sha256:d75bf5a79bdafea70faa29e2455504756103939504e761fe1b0860f8c27fe1d4

    Input: per-patient AUC table, summary and feature SHA-256 manifest of this 5-repeat run of protocol d3078c62.

  7. usesDefault nonlinear decoders did not outperform shrinkage LDA for SEEG working-memory load in 15 patients sci:sha256:86334d375dd1fe2c8d4657f3d8731540dae6851b8c63c0d9d9f64cd1b22d19af

    Input: per-patient AUC table, summary and feature SHA-256 manifest of this 5-repeat run of protocol d3078c62.

  8. usesThe decoder comparison stopped after one patient and provides no group result for memory load from SEEG band power sci:sha256:65ed565de94d202e45d65bd0d4be7c108eb24cf08b4e02ebe55432d3422f9d6b

    The protocol that all six runs used; it fixes 5 CV repeats (seeds 0-4).

  9. usesdoi:10.18112/openneuro.ds004752.v1.0.1

    The archived band-power features used by all six runs were derived from OpenNeuro ds004752. Their matching SHA-256 manifest is preserved with the source package.

  10. doi:10.7554/eLife.78677

    This is the paper identified by the dataset record for OpenNeuro ds004752.

Cited by

0 publications

Public catalog entries on Sciscape.

No incoming citations on Sciscape.