RBF kernel ridge decoding showed no detectable recall gain over a linear kernel in 18 patients with offline SEEG band power

We tested whether a nonlinear radial basis function (RBF) kernel ridge decoder predicted later word recall better than a linear kernel ridge decoder from archived SEEG band-power features. This independent reproduction of an earlier kernel logistic analysis used squared loss and different code. The data were features from OpenNeuro ds004789 v3.1.0, FR1 session 0: log10 power at 3, 5, 8, 13, 22, 40, 80, and 140 Hz, averaged over 0–1.6 s after word onset across bipolar channels. Eighteen of 19 patients met the inclusion rule. With nested list-grouped tuning and outer 5-fold cross-validation over whole lists, median AUC was 0.648 for the linear kernel and 0.654 for the RBF kernel. The primary median RBF-minus-linear difference was +0.004 (10/18 patients positive; exact two-sided Wilcoxon p = 0.28), within the pre-registered ±0.02 band. Within serial position, the median difference was +0.002 (p = 0.68). The RBF tuning often selected nonlinear widths, but did not increase AUC. An earlier logistic-loss analysis found a difference of −0.005 (Holm p = 0.24), so two estimators gave similar null results on these features. Both kernel ridge decoders had AUC about 0.02 below an archived fixed-C logistic decoder, but this comparison is confounded because that decoder trained on 24 of 25 lists per fold and these decoders trained on 20. The study is limited to one dataset, task, session, feature window, and 18 patients; it reports offline AUC, uses small tuning grids, and has no confidence interval. Small gains of about +0.01 cannot be excluded.

SEEG · word recall · memory decoding · kernel ridge regression · RBF kernel · linear decoder · AUC · cross-validation