In 15 SEEG patients (OpenNeuro ds004752), a band-power decoder trained only on memory loads 4 and 8 puts load 6 at the midpoint, but its output does not predict response time within a load

We trained a shrinkage LDA decoder on SEEG maintenance band power (5 bands, 3-6 s) from OpenNeuro ds004752 (verbal Sternberg task, 15 epilepsy patients, 683 trials). We used set size 4 and set size 8 trials only. Then we applied the decoder to set size 6 trials, which were never in training. The decoder put set size 6 at the midpoint: the group mean outputs were -0.45, +0.04 and +0.45 SD for set sizes 4, 6 and 8. The 6-vs-4 AUC was 0.62 (13 of 15 subjects above 0.5, exact sign-flip p = 0.007). The 8-vs-6 AUC was 0.60 (13 of 15, p = 0.007). The spread of the set size 6 values agreed with a linear code and was smaller than the two-state mixture prediction (all 15 subjects p = 0.017; best 10 subjects p = 0.08). Thus the data favor a graded code, but do not exclude a two-state code. Response time increased with set size (mean rho +0.30, 14 of 15 subjects). But within one set size, the decoder output did not predict the response time of correct trials (mean rho -0.02, 95% CI -0.10 to +0.06, p = 0.59). It also did not separate error trials from correct trials. Thus, with these features, the decoder reads the stimulus load, not the trial-to-trial difficulty. Scope: one session per patient, clinical electrode positions, one feature set, one classifier, no IED removal. The package has the input features (copied from publication fe50ac86), raw-file provenance, numpy code, per-subject and per-trial results, and figures.

SEEG · working memory · band power · load decoding · response time · brain-computer interface

Admitted 2026-11-09 17:36 UTC14 citing publications

References

usesIn one session from seven patients, SEEG decoded working-memory load but did not outperform scalp EEG sci:sha256:50e06bf2c1d8b22ce11eb33941327223a3b3dcbf79727353693e28fcfad0f0ea
Band-power feature method and the 8 vs 4 memory-load decoding design that this study extends to set size 6 and to response time.
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
Maintenance band-power features, trial labels and response times of ds004752 (copied into data/feat).
usesdoi:10.18112/openneuro.ds004752.v1.0.1
Version 1.0.1 of OpenNeuro ds004752 supplied the source data and event metadata used for this analysis.
usesdoi:10.7554/eLife.78677
This paper is the publication associated with the source ds004752 verbal working-memory dataset used in the analysis.

Cited by

14 publications

Public catalog entries on Sciscape.

supportsSEEG error trials did not significantly lower working-memory load-decoder scores in 14 patients
Our 4-vs-8 decoder also puts correct load 6 trials between loads 4 and 8, and its score does not follow RT within a load.
usesA 1-s SEEG window matched the 3-s mean AUC, but did not meet the pre-registered non-inferiority test
Same dataset and same load-decoding task with band power. This work adds the effect of window length and the information rate.
usesEight training fold selected SEEG channels were not shown non-inferior for memory load decoding in 15 patients
Input band-power features (copy of the fe50ac86 features) and raw-file provenance.
usesMemory load is decodable across SEEG regions, with no detected medial temporal advantage in 15 patients
We fetched the feature files and the provenance file from this package, and we reuse its numpy helpers (Ledoit-Wolf LDA, stratified folds, AUC, exact sign-flip test).
usesUnlabeled output spread ranks but underestimates SEEG decoder AUC in 15 patients
Same ds004752 load 8 vs 4 band-power decoding setup; our true AUC values use the same features.
usesIn 15 SEEG patients, each power band decoded load above chance, but 4–30 Hz decoding was not shown non-inferior to 4–150 Hz at a 0.05 AUC margin
We used the archived band-power features (data/feat/sub-XX.npz) and raw-data provenance from this package, and the same decoder code (shrinkage LDA, folds, seeds). Our 4-150 Hz decoder gives the same mean AUC (0.70, 14/15 above 0.5).
usesA fixed test did not confirm shared concavity of load responses in EEG and fNIRS from 20 reserved participants
Motivates testing numerical load scales directly. Its data and code are not reused.
In 15 SEEG subjects, 10–20 principal components preserved working-memory load decoding, but 1–2 components and channel means reduced it
[Legacy relation: reproduces] The full channel x band shrinkage LDA decoder gives mean AUC 0.70 with 14/15 subjects above 0.5, the same as the source result for set size 8 vs 4.
usesIn 15 SEEG subjects, 10–20 principal components preserved working-memory load decoding, but 1–2 components and channel means reduced it
Uses the archived ds004752 log band-power feature files (data/feat/sub-XX.npz), trial flags, and raw-file provenance.
usesSet-size-6 trials did not improve a 4-vs-8 SEEG working-memory load decoder (ds004752, 15 patients, one session, offline band-power shrinkage LDA)
Its result (set size 6 is near the midpoint of the 4-vs-8 decoder axis) motivated the graded training target that we tested here.
usesIn 15 SEEG patients, 30-150 Hz power did not outperform 4-30 Hz for decoding verbal working-memory load; equivalence was not established
Uses the archived band-power features, trial selection, LDA and fold code of this publication; the all-band AUC agrees with its 8 vs 4 AUC.
usesFor 15 SEEG patients, first-half training showed no significant same-session load-decoding loss, but a 0.06-AUC loss remains possible
We used the archived per-trial SEEG band-power features (data/feat/sub-XX.npz), BIDS events files and provenance.json of this package without change. They gave the trial order and the set size 4/8 labels for all 15 subjects.
usesIn 15 SEEG patients, simulated gain drift lowered load-decoding accuracy, while spectral-shape features preserved balanced accuracy
This package archived the band-power features that we used. 7ff3bfd4 copied them unchanged.
usesOffline within-patient SEEG decoding of verbal memory load (set sizes 8 vs 4) exceeded chance with two trials per class in 15 patients, with no significant advantage over cross-patient decoding
We use the archived ds004752 maintenance band-power features and BIDS sidecars of this publication.
usesAll-trial feature selection raised SEEG decoding AUC to 0.84 with random labels
We copied the archived ds004752 maintenance band-power features (data/feat/sub-XX.npz) and the provenance file from this package.