An offline SEEG decoder trained during encoding detected working-memory errors during delay in 12 subjects (mean AUC 0.537)

We measured the temporal generalization of an offline SEEG decoder for working-memory error trials using OpenNeuro ds006136 (12 subjects, load 3, 0.5-s band-power windows from SCI publication 08fc8e7b). A decoder trained on encoding windows detected errors in delay windows (mean AUC 0.537, 8 of 12 subjects above 0.5, exact Wilcoxon p = 0.013; p = 0.039 after removal of the linear trial trend). This AUC was not different from the AUC of decoders trained and tested in the same delay window (difference -0.007, p = 0.47); this difference test does not establish equivalence. In the delay, the AUC did not decrease when train and test windows were up to 3.5 s apart. A decoder trained on fixation windows also transferred (AUC 0.521, Holm p = 0.021), but not after trial-trend removal (Holm p = 0.064). Thus, part of the stable signal is likely slow drift. The AUC per window (about 0.54) is too low for single-trial BCI control.

SEEG · working memory · error decoding · temporal generalization · brain-computer interface · linear discriminant analysis · OpenNeuro