In 15 SEEG patients, no nonlinear decoder outperformed shrinkage LDA for working-memory load from maintenance band power

We completed a fixed decoding protocol for 15 patients performing a Sternberg verbal working-memory task. We compared seven default decoders for set size 8 versus 4 from log10 Welch power in five bipolar SEEG frequency bands during the 3–6 s maintenance window. Each patient had 20–42 trials and 70–280 features. Stratified five-fold cross-validation was repeated five times, with the subject as the unit of analysis. Shrinkage LDA had mean AUC 0.692. No nonlinear decoder had a higher mean AUC: RBF-SVM was 0.014 lower, k-nearest neighbors 0.043 lower, random forest 0.021 lower, and boosted stumps 0.064 lower. The sign-flip p value for boosted stumps was 0.061; after Holm correction across six comparisons, no decoder difference was significant. L2 logistic regression was 0.011 higher (p = 0.041 before correction). A trial-index control identified order imbalance in two patients, but excluding them did not change the result. These findings support regularized linear decoding as a default for small-calibration SEEG read-out under the tested offline conditions. Limits include one session and run per patient, few trials, default hyperparameters without nested tuning, one feature set, and low power for differences smaller than about 0.03 AUC.

SEEG · working memory · decoding · shrinkage LDA · nonlinear classifiers · AUC · intracranial EEG · cross-validation