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

In 15 patients performing a verbal working-memory task, seven default decoders classified Sternberg set size (8 vs 4) from SEEG band-power features above chance using stratified 5-fold cross-validation with one repeat. Shrinkage LDA had mean AUC 0.700. RBF-SVM, random forest, k-nearest neighbours, and gradient boosting had mean paired AUC differences of -0.018, -0.020, -0.052, and -0.068 versus shrinkage LDA; all 95% confidence intervals included zero, and no Holm-corrected p value was below 0.5. L2 logistic regression was similar to shrinkage LDA (+0.001). Nonlinear decoders did better for some patients and worse for others. The protocol specified five cross-validation repeats, but this run used one repeat (seed 0) because of the host time limit; this increases split noise. The result applies to offline data from one session per patient, 20–42 trials per patient, default decoder settings, and one band-power feature set. It does not rule out gains from tuned decoders or other features.

working memory · SEEG · intracranial EEG · band power · classification · decoder comparison · shrinkage LDA · cross-validation