Cross-session SEEG decoder AUC declined slightly over 0–72 hours, with no clear day-boundary step

In 10 subjects performing a verbal Sternberg task, we evaluated offline transfer of a shrinkage LDA decoder between sessions using archived late-maintenance log band-power features from OpenNeuro ds004752 version 1.0.1. The analysis included 278 ordered session pairs over gaps of 0–72 hours and contrasted set size 8 with set size 4. With per-session z-scoring, the mean subject-level AUC slope against log2(1 + elapsed hours) was −0.0068 per doubling (95% bootstrap CI −0.0126 to −0.0009; 7/10 slopes negative; exact sign-flip p = 0.066). Same-day transfer exceeded other-day transfer by 0.038 AUC (95% CI 0.002 to 0.075; p = 0.084), while the 1-day versus 2+-day difference was 0.011 (7 subjects; p = 0.77). Mean AUC was 0.685 for within-session cross-validation, 0.715 for same-day transfer, 0.702 at 1 day, and 0.670 at 2+ days. Raw features showed the same direction (slope −0.0101, p = 0.057). The evidence does not show a clear day-boundary step or establish that daily recalibration is necessary; it also does not exclude a loss of about 0.04 AUC over 3 days. The analysis was limited to 10 subjects, one task, label contrast, feature set, and decoder, with a maximum gap of 72 hours. Per-session z-scoring used unlabeled test-session statistics. The analysis plan was not registered before results. Features were copied from prior SCI contributions and were not recomputed from the EDF files.

SEEG · brain-computer interface · cross-session decoding · decoder transfer · AUC · shrinkage LDA · calibration · verbal working memory · OpenNeuro