Encoding band power predicted recall above a free-label null but not beyond serial position in 78 pure-SEEG patients

In free recall, serial position predicts whether a word will be recalled, which can make an encoding-power classifier appear predictive. We evaluated a pre-registered leave-one-list-out ridge classifier on archived encoding band-power features from 78 pure-SEEG patients (FR1 session 0; 8,049 bipolar pairs). The classifier used five bands from 0–1.6 s after word onset. Its mean AUC was 0.521. It exceeded a free-label permutation null by 0.022 AUC (95% bootstrap CI 0.004 to 0.039; Holm p = 0.032), but did not exceed a null that permuted recall labels only among words at the same serial position (difference 0.003, 95% CI −0.012 to 0.019; Holm p = 0.72). The serial-position-only benchmark had mean AUC 0.559, and the serial-position component accounted for 87% of the mean difference from the free-label null. The null result persisted in the tested penalty, band, hippocampal-pair, and within-list analyses. These results provide no evidence that encoding band power predicted recall beyond serial position in this sample and analysis. The study used one linear classifier, five bands, one time window, and session 0 only; the 95% confidence interval excludes group effects larger than about +0.019 AUC. The findings support evaluating recall classifiers against a serial-position-stratified null or benchmark.

free recall · serial position · stereo-EEG · SEEG · band power · ridge regression · cross-validation · permutation test · memory encoding