SEEG maintenance-period band power decodes verbal working-memory load (4 vs 8 letters) across recording days above chance in 3 of 3 tested epilepsy patients (pilot, OpenNeuro ds004752)

Question: Does a SEEG decoder for working-memory load continue to work on a different recording day? This is a necessary condition for cognitive-state BCIs. Data: OpenNeuro ds004752 v1.0.1 (Sternberg verbal working-memory task). We processed 16 sessions; 3 subjects had sessions on 2 or more days. Method: bipolar SEEG, log band power (4-150 Hz) in the 3 s maintenance period, per-session z-scoring without labels, and L2 logistic regression trained on one session and tested on another session (60 ordered pairs). Result: the cross-day ROC AUC was 0.555, 0.621 and 0.630 (each subject at the floor p=0.048 of a 20-permutation test). Same-day AUC was 0.78 vs 0.62 cross-day in one subject, and 0.61 vs 0.63 in another. AUC did not decrease with days between sessions. Limits: n=3 subjects, few permutations and small trial counts. This is a pilot. The package has code, per-trial features and figures to extend the test to the 11 multi-day subjects.

SEEG · verbal working memory · cross-session decoding · band power · ROC AUC · epilepsy · pilot study