Earlier-session data helps recall decoding with one calibration list, but not with about 20, in a three-patient SEEG pilot

We tested whether adding labeled words from an earlier free-recall session improves decoding in a later session. In three patients with two sessions each, an L2 logistic decoder trained on both sessions had nearly the same held-out AUC as one trained on the later session alone when about 20 later-session training lists were available (mean difference −0.003; bootstrap 95% CI −0.022 to +0.016). With one later-session list, pooling improved AUC by +0.072 in all three patients; the mean benefit decreased to +0.016 with eight lists. Training only on the earlier session reduced AUC by a mean 0.042, although AUC remained above chance. No result was statistically significant: with n = 3, the smallest possible two-sided exact Wilcoxon p value is 0.25. This pilot cannot support population inference. It followed a preregistered nine-patient plan, but only the first three planned session-1 downloads finished within the 45-minute run limit. Results apply to the tested spectral power features and fixed regularization setting; tuning session weights or testing other feature types may produce different results.

free recall · SEEG · memory decoding · brain-computer interface · cross-session decoding · calibration · AUC · pilot study