On ds006253 SEEG features in 15 patients, fixed nonlinear decoders did not improve direction decoding; RBF added 0.011 AUC for movement detection

We compared Ledoit–Wolf shrinkage linear discriminant analysis (LDA) with fixed radial-basis-function (RBF) kernel ridge and k-nearest-neighbour (kNN) decoders on author-provided 70–150 Hz SEEG window-mean features from 15 patients in OpenNeuro ds006253. The analysis used 5 contiguous time-ordered folds and no hyperparameter tuning. For the primary task, left-versus-right movement direction 0–0.5 s after movement onset, LDA mean AUC was 0.573. RBF minus LDA was −0.010 (95% CI −0.024 to +0.005; Holm p = 1.0; 5/15 patients improved), and kNN minus LDA was −0.034 (95% CI −0.053 to −0.015; Holm p = 1.0; 2/15 improved). LDA was above chance (12/15 patients; sign-flip p = 0.0027). Neither nonlinear decoder gained more than the preregistered 0.02 AUC threshold. For the secondary movement-versus-pre-stimulus-rest task, LDA mean AUC was 0.845. RBF added 0.011 (95% CI +0.002 to +0.020; Holm p = 0.032; 11/15 improved), meeting the preregistered rule for a nonlinear improvement, though the gain was small. kNN minus LDA was −0.038 (95% CI −0.053 to −0.023; Holm p = 1.0; 2/15 improved). These results apply to this feature set and fixed settings in an offline analysis; other features and tuned decoders were not tested.

SEEG · high-gamma activity · brain-computer interfaces · shrinkage LDA · kernel ridge · k-nearest neighbours · AUC · time-ordered cross-validation