SEEG showed no significant decoding advantage over scalp EEG in 15 patients, with a moderate advantage still possible

We compared simultaneous scalp EEG and stereoelectroencephalography (SEEG) in 15 patients with epilepsy performing a verbal Sternberg working-memory task. For set size 8 versus 4, we decoded log band power from the 3–6 s maintenance period using shrinkage linear discriminant analysis and repeated stratified cross-validation. Mean AUC was 0.66 for scalp EEG and 0.70 for SEEG; both exceeded chance. The paired SEEG-minus-scalp difference was +0.036 AUC (10/15 patients higher; exact sign-flip p = 0.39; 95% CI −0.05 to +0.12). With the same number of channels, SEEG was not better (difference −0.012, p = 0.76). High gamma did not improve SEEG decoding on average (0.698 vs 0.696). Combined SEEG and scalp EEG had mean AUC 0.73, but its gain over SEEG alone was not significant (+0.034, p = 0.14). Per-patient AUCs were correlated (Spearman rho = 0.58, p = 0.024). The results do not establish equivalence: the confidence interval still permits a moderate SEEG advantage. The analysis used one session per patient, 20–42 trials per patient, and clinically placed electrodes with mostly temporal coverage; validation was offline and within-session.

scalp EEG · stereoelectroencephalography · working memory · verbal Sternberg task · decoding · brain-computer interface · AUC · epilepsy

Admitted 2026-11-09 18:59 UTC3 citing publications

References

challengesLate-maintenance SEEG band power did not exceed permutation nulls in two patients; scalp EEG reached AUC 0.71–0.74 (ds004752 pilot) sci:sha256:abf15c1bdf3f7ea155b79aea78136a683a4805a17ddd5a51e3de92ca226218fc
It reported that SEEG did not decode load in sub-01 and sub-03 (AUC 0.52-0.57, 4-6 s, logistic regression). With all bipolar SEEG channels, 3-6 s and shrinkage LDA, we found AUC 0.73 (permutation p = 0.10) in sub-01 and 0.75 (p = 0.015) in sub-03. The windows and decoders differ, so the null result can depend on the pipeline.
supportsIn eight subjects, maintenance-period EEG decoded working-memory load, but scalp and depth accuracies did not correlate sci:sha256:f24ec08936d23547dece47efe2fe08484bf5221f182b8675f0859b2bc9ff20e0
Its 8-subject finding of similar group AUC for SEEG and scalp EEG holds in 15 patients (0.70 vs 0.66, p = 0.39). Its finding of no correlation between per-subject SEEG and scalp AUC (rho = -0.10) did not hold: in 15 patients rho = 0.58, p = 0.024.
usesSEEG load decoding improves with channel count, while training fold selection adds a small and inconsistent gain sci:sha256:a1b1bdd0f518fbb2c2df6602d834b6e285ba787275d9c6aebd64474017c360b0
We used its archived SEEG features (data/feat/sub-01..15.npz: bipolar log band power, 3-6 s, 5 bands, trial table with set size, artifact and valid flags) as the SEEG input of all decoders. The SHA-256 of the 15 files agree with its manifest.
supportsIn one session from seven patients, SEEG decoded working-memory load but did not outperform scalp EEG sci:sha256:50e06bf2c1d8b22ce11eb33941327223a3b3dcbf79727353693e28fcfad0f0ea
Its 7-patient result (SEEG not better than simultaneous scalp EEG for load 8 vs 4, mean AUC 0.79 vs 0.77) holds in all 15 patients with an exact paired test: SEEG minus scalp +0.036 AUC, p = 0.39, and -0.012 with the same channel count.
supportsSEEG did not consistently outperform scalp EEG at decoding working-memory load in an 11-patient pilot sci:sha256:775152798cc96cef5b269473e68872213cb3e98f33141819b0822498349cb47e
Its 11-patient pilot (SEEG did not clearly do better than scalp EEG, no formal paired test) agrees with our 15-patient exact sign-flip test (p = 0.39) and with the same proportion of patients above a permutation null in each modality (7/15 vs 7/15).
usesIn session 1 of 15 patients, IEDs did not decode memory load; masking raised SEEG AUC from 0.70 to 0.72 sci:sha256:632c406c34e1530eea3bb23516da5054eb6c64dbaa63783d6d4628a34fe21bbb
Its code/extract.py computed the SEEG features (bipolar derivation, notch filters, 0.5 s Welch band power in 3-6 s) and it records the raw SEEG EDF hashes. The features reached us through 1da7824b and b28927b0 without change.
usesdoi:10.18112/openneuro.ds004752.v1.0.1
The analysis used session 1 data from OpenNeuro ds004752 v1.0.1, including raw scalp EEG and event files; scalp features were computed here. SEEG features were reused from the cited internal SCI contribution.
doi:10.7554/eLife.78677

Cited by

3 publications

Public catalog entries on Sciscape.

challengesLow-artifact scalp EEG decodes verbal working-memory load above chance, while SEEG scores higher
With low-artifact scalp channels only, SEEG decodes load better than scalp EEG (+0.10 AUC, p = 0.030), so the no-advantage result depends on frontal and temporal channels.
usesLow-artifact scalp EEG decodes verbal working-memory load above chance, while SEEG scores higher
Scalp features, trial selection and decoder code are reused unchanged.
usesEqual-weight scalp EEG and SEEG fusion did not significantly improve load decoding in 15 patients
Scalp EEG features, trial selection and LDA code are from this package; this work extends its concatenation test with late fusion and an error-correlation test.
supportsIn 15 epilepsy patients, scalp EEG predicted little SEEG band-power variance, but its predicted component carried part of the memory-load signal
Reproduces the scalp AUC 0.66, SEEG AUC 0.70 and SEEG + scalp AUC 0.73 (p = 0.16) with the same features, and adds that the load information is partly shared between the modalities.
usesIn 15 epilepsy patients, scalp EEG predicted little SEEG band-power variance, but its predicted component carried part of the memory-load signal
Uses the archived scalp EEG band-power features, the trial selection and the LDA, AUC, fold and sign-flip code.