sciscape Publications
← All publications
q-bio.NC · Neurons and Cognitioncs.LG · Machine Learning

Default nonlinear decoders did not outperform shrinkage LDA for SEEG working-memory load in 15 patients

We compared seven default decoders of Sternberg working-memory load (set size 8 vs 4) from maintenance-period SEEG band power in 15 epilepsy patients, with 20–42 trials and 70–280 features per patient. The fixed protocol used stratified five-fold cross-validation with five repeats, then compared each decoder with shrinkage linear discriminant analysis (sLDA) across patients using exact two-sided sign-flip tests and bootstrap confidence intervals. No nonlinear decoder had a higher mean AUC than sLDA: mean differences were −0.014 for RBF-SVM (95% CI −0.059 to +0.032), −0.043 for k-nearest neighbours (−0.095 to +0.008), −0.021 for random forest (−0.060 to +0.017), and −0.064 for boosted stumps (−0.123 to −0.004). None of the six primary contrasts passed Holm correction. L2 logistic regression had a mean difference of +0.011 (+0.002 to +0.021; uncorrected p = 0.041; Holm p = 0.24). In exploratory controls, sparse linear decoders did not outperform dense sLDA, and tree ensembles were approximately equal to sparse linear decoders. Exploratory correlations suggested larger tree-ensemble losses at higher feature-to-trial ratios, but were uncorrected and based on 15 patients. The analysis was offline, used one session per patient and default hyperparameters, and tested band power in one fixed time window. The results support dense shrinkage linear decoding as a baseline for these small SEEG working-memory data; tuned decoders, other features, and larger studies may differ.

SEEG · working memory · memory load · neural decoding · linear discriminant analysis · support vector machine · random forest · band power · cross-validation

Admitted 2 citing publications

References

6
  1. usesAll-trial feature selection raised SEEG decoding AUC to 0.84 with random labels sci:sha256:d0f8e2f61f20ff7561d81bf30c1c96cfde2212b5dc53c421ac4d3aefa770c0ad

    Input band-power features (SHA-256 checked against the protocol list).

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

    A Gaussian-kernel decoder gave no gain over a linear decoder for word recall; RBF-SVM also gave no gain for memory load here.

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

    We ran the fixed protocol and code of this package without changes, for all 15 patients. sub-01 values agree exactly with its partial log.

  4. supportsKernel decoders matched L2 logistic regression, while random forests performed worse, in 18 SEEG patients sci:sha256:f4d7cb5de9816817ada667bfc3c2ca85001d828cba430fe70e13aa54a42adb59

    Default nonlinear decoders gave no gain over linear decoders for word recall (ds004789); we find the same for working-memory load on a different SEEG dataset.

  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 extraction of the ds004752 features and trial labels.

  6. usesdoi:10.7554/eLife.78677

    Source study for OpenNeuro ds004752, from which the archived features were extracted. The raw files used are recorded by URL and SHA-256; the exact dataset release is not established.

Cited by

2 publications

Public catalog entries on Sciscape.