In 13 SEEG sessions of load-3 working memory, burst fraction did not outperform band power (median AUC 0.54 vs 0.63; offline, one threshold)

We tested whether oscillation-burst features predict working-memory errors better than band power in human SEEG. We used OpenNeuro ds006136 (OWM-Dataset), version 1.0.1, from 13 usable sessions across 12 patients, with contacts in the middle frontal and middle temporal gyri. For each channel, we calculated five-band log power and five-band burst fraction (envelope above 2 times the median for at least 3 cycles) during encoding and delay. Shrinkage LDA with 5-fold cross-validation produced channel scores that we averaged within sessions. Band power predicted errors in 12 of 13 sessions (median AUC 0.63; exact sign-flip p = 0.002). Burst fraction was weaker (median AUC 0.54; p = 0.08) and lower than band power in 11 of 13 sessions (paired p = 0.073). Adding burst fraction to band power did not improve AUC (median difference -0.005; mean +0.002; p = 0.92). For this task and burst definition, burst fraction gave no advantage over band power as an offline memory BCI trigger. This result is limited to one dataset, load-3 task, threshold, and whole-window offline features; it is not a causal real-time decoder. The package contains code, S3 provenance, per-session results, a figure, and reproduction steps.

SEEG · working memory · burst fraction · band power · memory BCI · error prediction · AUC · shrinkage LDA · OpenNeuro ds006136