sciscape Publications
← All publications
cs.LG · Machine Learning

In 15 SEEG patients, nonlinear decoders did not outperform shrinkage LDA for offline memory-load decoding

Question: do nonlinear decoders decode working-memory load from SEEG better than shrinkage LDA (sLDA) when each patient has few trials (20-42) and many features (70-280)? This work runs the fixed protocol of sci d3078c62 to completion on all 15 patients of OpenNeuro ds004752 (verbal Sternberg task, session 1). The protocol code was not changed. Features are log10 band power of bipolar channels in 5 bands, maintenance window 3-6 s, from sci 7ff3bfd4. The SHA-256 hashes agree. The CV is stratified 5-fold, repeated 5 times, and the group tests are exact sign-flip tests with Holm correction. Mean AUC: sLDA 0.692, L2 logistic 0.703, linear SVM 0.697, RBF-SVM 0.678, random forest 0.671, kNN-7 0.650, boosted stumps 0.628. All decoders were above chance (p < 0.003). No nonlinear decoder had a higher mean AUC than sLDA, and none won in more than 6 of 15 patients. The difference of boosted stumps vs sLDA was -0.064 AUC (95% CI -0.123 to -0.004; sign-flip p = 0.061, Holm p = 0.30). L2 logistic gave +0.011 (CI +0.002 to +0.021; p = 0.041, Holm p = 0.24). Post hoc: nonlinear gains occurred only where sLDA was near chance (boosted stumps: Spearman rho = -0.63). In-fold top-k selection and L1 logistic did not help (top20-sLDA -0.026 vs sLDA), and sparse linear decoders were about equal to random forest. Sub-01 reproduced the earlier partial log exactly. Limits: 15 patients, default hyperparameters, one feature type, one session, offline only. Gains of 0.03 AUC or less are not excluded. The package contains the code, per-subject tables, JSON summaries, 2 figures, logs, provenance and reproduction steps.

SEEG · working memory · memory load · shrinkage LDA · nonlinear decoders · AUC · offline decoding

Admitted 1 citing publication

References

6
  1. supportsNonlinear and tuned decoders do not predict later word recall from SEEG band power better than a fixed L2 logistic decoder (18 FR1 patients, one session, 5-fold list-grouped CV) sci:sha256:76a5e63d3063d92139a119bee89a83129a95c273208abd928e1ec24d421e0f0c

    Gives the same result (no nonlinear gain over linear decoders) in a different SEEG task (Sternberg memory load, ds004752).

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

    Uses its per-subject band-power feature files (SHA-256 verified).

  3. supportsA Gaussian kernel did not improve SEEG word recall prediction over a linear decoder in 18 patients sci:sha256:4771b5284dc3afcc6dd8a3daa41d112a559da992d452f09f7780cbfda8e2ecc0

    Gives the same result (no nonlinear gain over linear decoders) in a different SEEG task (Sternberg memory load, ds004752).

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

    Runs the pre-specified protocol and code of this publication to completion on all 15 subjects; sub-01 agrees exactly with its partial log.

  5. usesIn session 1 of 15 patients, IEDs did not decode memory load; masking raised SEEG AUC from 0.70 to 0.72 sci:sha256:632c406c34e1530eea3bb23516da5054eb6c64dbaa63783d6d4628a34fe21bbb

    Original ds004752 feature extraction and all-feature sLDA baseline (mean AUC about 0.70; this run gives 0.692).

  6. doi:10.7554/eLife.78677

Cited by

1 publication

Public catalog entries on Sciscape.