Across two SEEG datasets, pooled results excluded a 0.02 AUC RBF-SVM gain over shrinkage LDA; ds006253 alone did not

We compared default nonlinear decoders with Ledoit–Wolf shrinkage linear discriminant analysis (LDA) for MOVE versus REST classification from 70–150 Hz high-gamma activity in one 0.5 s window per bipolar channel. In 15 patients from OpenNeuro ds006253, LDA reproduced the archived AUCs (mean 0.833; maximum absolute difference 4.7×10⁻⁵). The RBF-SVM minus LDA mean AUC was +0.009 (95% t-CI −0.004 to +0.023; p=0.17), so this dataset alone did not exclude a gain of 0.02. In a pooled analysis with eight archived patients from ds004473, the SVM difference was +0.005 (95% t-CI −0.005 to +0.014), below the pre-registered 0.02 margin. k-nearest neighbors and L2 logistic regression performed worse than LDA in each dataset and in the pooled analysis. The study used default hyperparameters and one time-ordered 60/40 split per patient. Its conclusions apply to this feature and task; the pooled test combines different tasks and assumes independent, exchangeable patients.

stereo-EEG · high-gamma activity · movement decoding · brain-computer interface · shrinkage LDA · RBF-SVM · decoder comparison · AUC