In 18 SEEG patients, word-n recall-decoder output weakly predicted next-word recall after serial-position control

Closed-loop memory stimulation reads a recall decoder on word n and acts during word n+1. We tested if the decoder output of word n predicts recall of word n+1. We used archived per-word SEEG band-power features (c4b8c884; OpenNeuro ds004789 v3.1.0, FR1, session 0; 18 patients). The decoder was the cross-validated L2 logistic model of c4b8c884 (leave-one-list-out). A training-free high-gamma minus theta/alpha contrast was the second score. We wrote the plan before the analysis. Within target serial position, the median decoder AUC was 0.618 for the same word and 0.530 for the next word. The drop occurred in 16 of 18 patients (Holm p = 0.0009). The next-word AUC was not significant after Holm correction (p = 0.18; uncorrected p = 0.037). At lags 2 and 3 it was 0.506 and 0.500. About one quarter of the same-word information remained ((AUC-0.5) ratio 0.26). The adjacent-word score correlation was 0.156 after removal of serial position, but only 0.018 after removal of list means. Thus the shared state is mostly slow and list-level. In a simulated median split, the high-score half had 15.3 percentage points more recall for the same word and 6.2 points more for the next word. Serial position alone predicted 2.7 points of the next-word difference. Two of 18 patients kept almost all information at the next word. Limits: one dataset and task, offline only, no stimulation, word-locked 0-1.6 s features, and full-session z-scoring and median. The package has code, the plan, per-patient results, a figure, input checksums and reproduction steps.

SEEG · free recall · memory decoding · serial position · closed-loop stimulation · AUC · word recall