In 18 SEEG patients, nonlinear decoders did not outperform linear logistic recall prediction

We compared four within-patient decoders of later free recall from SEEG encoding-period band-power features in 18 patients in the FR1 task of OpenNeuro ds004789. Each patient contributed about 150–300 words. Outer evaluation used five folds grouped by list. Median AUC was 0.670 for linear L2 logistic regression, 0.628 for linear-kernel ridge, 0.621 for RBF-kernel ridge, and 0.590 for gradient-boosted stumps. RBF-kernel ridge minus logistic regression was −0.035 (95% patient-bootstrap CI −0.057 to −0.020; Holm-adjusted p = 0.018); boosted stumps minus logistic regression was −0.053 (CI −0.092 to −0.023; Holm-adjusted p = 0.007). With the same ridge loss and tuning, RBF-kernel ridge and linear-kernel ridge had similar AUC (median difference −0.005, CI −0.024 to +0.008; p = 0.37). The within-serial-position results had the same signs, and inner cross-validation selected the most linear RBF width in 55 of 90 folds. Under the prespecified rule, these results do not support a nonlinear gain of 0.02 AUC or more for this sample and feature set. The logistic decoder's C value came from earlier work on the same dataset and may favor that reference. The study tested four decoders on band power in one 1.6-second window, used five-fold rather than leave-one-list-out cross-validation after a timing-based plan amendment, and does not establish how larger training sets or raw time-series features would perform.

SEEG · free recall · memory decoding · band power · logistic regression · kernel ridge · nonlinear decoding · cross-validation