An RBF-kernel decoder did not predict word recall better than a linear decoder from SEEG band power (18 patients, FR1 session 0, offline)

In an offline analysis of archived per-word SEEG band-power features from FR1 session 0, we compared a linear L2 logistic decoder with an RBF-kernel logistic decoder using the same nested, list-grouped tuning in 18 patients. Median outer-fold AUC was 0.646 for linear and 0.631 for RBF. The median RBF-minus-linear difference was -0.005 (bootstrap 95% CI [-0.024, +0.001]); 6 of 18 patient differences were positive and the exact Wilcoxon p value was 0.14 (Holm-adjusted p = 0.24). RBF AUC exceeded chance in all 18 patients (one-sided exact Wilcoxon p = 3.8e-6), while the median correlation between the decoders’ word scores was 0.85. Within serial position, the median difference was -0.013 (95% CI [-0.028, +0.010], 7 of 18 positive, Holm p = 0.24). A secondary comparison found that fixed-C linear decoding outperformed nested-C selection by a median 0.019 AUC (3 of 18 positive for nested minus fixed; p = 0.024, Holm p = 0.071); the fixed C was not selected inside the outer folds and may have been chosen with knowledge of the same dataset. The results do not show an RBF gain in this setting; a non-significant test does not prove equality. The study used one session, a single 1.6 s mean-band-power window, a reduced hyperparameter grid, and 18 patients, so other features, grids, models, or sessions may give different results. The analysis plan was written before model fitting; computational constraints led to documented changes before results were known.

SEEG · word recall · memory · logistic regression · RBF kernel · nested cross-validation · band power