Within-patient SEEG recall decoders transfer across early and late list positions in FR1 data from 17 patients

We tested whether within-patient SEEG decoders trained on early list positions (1–6) transfer to late positions (7–12), and vice versa, using archived FR1 encoding log-power features from 17 patients. L2 logistic regression (C = 7.2 × 10⁻⁴) was evaluated with leave-one-list-out cross-validation. The pre-specified pooled-AUC transfer cost was −0.027 (95% bootstrap CI −0.069 to 0.011; exact sign-flip p = 0.23), so the study detected no transfer cost. Cross-half AUCs were 0.614 (early-trained on late words) and 0.621 (late-trained on early words); 16/17 patients exceeded 0.5 for each. With AUC comparisons restricted to word pairs at the same serial position, the transfer cost was −0.050 (95% CI −0.094 to −0.011; p = 0.033). An exploratory contrast found that early-trained AUC on early words decreased by 0.052 when position was controlled (p = 4.6 × 10⁻⁵), consistent with position-related signal contributing to within-half accuracy. Training on all positions gave the highest AUC. These results support transfer within this task, feature set, fixed model, and sample; they do not establish generalization beyond those conditions. The original report notes small per-half samples and advises multiple-test correction for the smaller tests.

SEEG · memory decoding · serial position · free recall · logistic regression · AUC · brain-computer interface · cross-validation