In 15 epilepsy patients, scalp EEG predicted little SEEG band-power variance, but its predicted component carried part of the memory-load signal

We used archived per-trial log band power from simultaneous SEEG (669 bipolar channels) and scalp EEG (8-23 channels) in 15 patients of OpenNeuro ds004752 (maintenance 3-6 s, set size 8 vs 4). Inside each cross-validation training fold, ridge regression predicted each SEEG channel-band feature from all scalp features. The median out-of-fold R2 was below 0 in all five bands (4-150 Hz). Only 11-23% of channels had R2 > 0.1. But when we removed the scalp-predicted part, the SEEG load AUC (shrinkage LDA) decreased from 0.70 to 0.62 (3 of 15 patients increased; sign-flip p = 0.002). A control with shuffled trial pairing decreased the AUC by only 0.016. The SEEG residual still decoded load above chance (p = 0.002). In the other direction, removal of the SEEG-predicted part decreased the scalp AUC from 0.66 to 0.59 (p = 0.001), and the scalp residual stayed above chance (p = 0.026). Thus, for this task and these linear band-power features, most SEEG variance is not visible on the scalp, but the load information is partly shared and partly specific to each modality. The shared part can include global state or common artifacts. The package contains numpy-only code, per-patient and per-channel results, an SVG figure, input checksums and reproduction steps.

scalp EEG · SEEG · verbal working memory · memory load · band power · ridge regression · linear discriminant analysis · epilepsy