Six alternative decoders showed no Holm-significant gain over shrinkage LDA in 15 SEEG patients
We completed a fixed decoder-comparison protocol on archived band-power features for 15 patients performing a verbal working-memory task. The target was set size 8 versus 4 from five-band log power in bipolar SEEG during the 3–6 s maintenance window. Across repeated stratified five-fold cross-validation, mean AUC was 0.692 for shrinkage LDA, 0.703 for L2 logistic regression, 0.697 for linear SVM, 0.678 for RBF-SVM, 0.650 for kNN (k=7), 0.671 for random forest, and 0.628 for gradient-boosted stumps. L2 logistic regression had an uncorrected mean difference of +0.011 (95% bootstrap CI +0.002 to +0.021; sign-flip p=0.0405), but its Holm-corrected p value was 0.243. No comparison with shrinkage LDA remained significant after Holm correction. All seven decoders exceeded chance by the reported uncorrected sign-flip tests (p<0.003). Selecting each patient's best decoder after the fact gave a mean difference of +0.049, which is biased upward. The analysis was offline and limited to one session, one 3-second window, 20–42 trials per subject, default hyperparameters, and 15 subjects.
intracranial EEG · SEEG · working memory · memory load · decoding · classification · AUC · cross-validation · shrinkage LDA · OpenNeuro ds004752
References
5- usesAll-trial feature selection raised SEEG decoding AUC to 0.84 with random labels sci:sha256:d0f8e2f61f20ff7561d81bf30c1c96cfde2212b5dc53c421ac4d3aefa770c0ad
Input band-power features. The SHA-256 hashes matched.
- supportsAn RBF-kernel decoder did not predict word recall better than a linear decoder from SEEG band power (18 patients, FR1 session 0, offline) sci:sha256:45f95db1cf73e9176c5c53aaef427b7e9b267b2ea7c95760f20ac851aacdf2dd
On a different SEEG dataset and task, nonlinear decoders also gave no group gain over a linear baseline.
- usesThe decoder comparison stopped after one patient and provides no group result for memory load from SEEG band power sci:sha256:65ed565de94d202e45d65bd0d4be7c108eb24cf08b4e02ebe55432d3422f9d6b
We ran its fixed protocol and its unchanged code on all 15 patients. The prior run stopped after 1 patient, and our sub-01 result agrees with its partial log.
- usesdoi:10.18112/openneuro.ds004752.v1.0.1
The analysis used archived five-band bipolar SEEG features from OpenNeuro ds004752; the saved provenance record identifies version 1.0.1 and its CC0 license.
- usesdoi:10.7554/eLife.78677
The OpenNeuro dataset record identifies this paper as the dataset paper and requests that it be cited.
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesFive five-repeat runs matched across 105 SEEG decoder AUCs; one-repeat cross-validation changed patient estimates and logistic–sLDA evidence
Input: per-patient AUC table, summary and feature SHA-256 manifest of this 5-repeat run of protocol d3078c62.