In 17 SEEG patients, calibration on the first eight lists showed no significant AUC loss versus time-blind training in one session

We tested whether a word-recall decoder trained on the first eight lists of a session predicts later lists as well as a decoder trained on eight lists drawn without regard to time. We analyzed archived per-word bipolar SEEG power features from 17 patients in one FR1 session each and used L2 logistic regression with training-only z-scoring. Median AUC was 0.626 for fixed early calibration, 0.644 for interleaved training, 0.617 for sliding training on recent lists, and 0.650 for growing training on all earlier lists. The fixed-minus-interleaved difference was +0.001 (95% CI for the median, -0.040 to +0.014; exact Wilcoxon p = 0.75), so we found no significant loss; the confidence interval allows a median loss of up to 0.04 AUC. Recent-list recalibration did not significantly improve accuracy. Features drifted with list number in 13 of 17 patients, while median accuracy did not fall with time since calibration. Two patients with large drift had lower late-list accuracy. The evidence is limited to one session per patient, 17 patients, one strongly regularized decoder and feature set, and offline analysis without stimulation. It does not establish performance across days or during closed-loop stimulation.

SEEG · word recall · memory BCI · causal calibration · decoder drift · AUC · logistic regression · OpenNeuro ds004789