In eight subjects, maintenance-period EEG decoded working-memory load, but scalp and depth accuracies did not correlate

We tested whether single-trial band power during maintenance of a verbal Sternberg task decoded set size (4 vs 8 letters) from medial temporal lobe (MTL) stereoelectroencephalography (SEEG) and simultaneous scalp EEG. In eight subjects, using one session each and 20–36 clean trials per subject, repeated stratified five-fold cross-validation with shrinkage linear discriminant analysis gave median maintenance AUCs of 0.70 for SEEG (group Wilcoxon p=0.008) and 0.74 for EEG (p=0.020). The modalities did not differ in group AUC (two-sided p=1.0; median difference −0.03), and their per-subject AUCs did not correlate (Spearman rho=−0.10, p=0.82). Fixation-window decoding was at chance at group level. SEEG single-band decoding was above chance for high gamma (mean AUC 0.68, p=0.020) and gamma (0.64, p=0.027); no EEG single band was significant. Single-channel SEEG effects had no consistent sign across subjects. These findings apply to the analyzed clinically selected MTL sites and within-session folds. The small sample, incomplete analysis of the 15-subject dataset, possible muscle and eye-movement contributions to scalp gamma, and lack of cross-session or cross-subject testing limit conclusions about a practical brain-computer interface.

verbal working memory · medial temporal lobe · stereoelectroencephalography · scalp EEG · band power · decoding · ROC AUC · brain-computer interface