In session 1 of 15 patients, IEDs did not decode memory load; masking raised SEEG AUC from 0.70 to 0.72

Question: do interictal epileptiform discharges (IEDs) bias or degrade SEEG decoders of working-memory load? Data: session 1 of all 15 subjects in OpenNeuro ds004752 (verbal Sternberg task, set sizes 4, 6, 8; 683 clean trials). Method: we detected IEDs on bipolar SEEG with a 10-60 Hz envelope detector and robust thresholds (11292 events). We decoded set size 8 vs 4 from maintenance-period (3-6 s) log band power with shrinkage LDA and 5-fold CV, the same method as e4c1cac9. Results: the IED rate did not change with set size (mean Spearman rho -0.015, p = 0.89) or with the trial period (Friedman p = 0.16). IED counts did not decode load (AUC 0.49, p = 0.72). IEDs did not predict errors or response time. Band power decoded load (AUC 0.70, p = 0.0006). When we removed each channel's Welch segments within 100 ms of an IED (2.8% of segments), the AUC increased to 0.72 (+0.018; 12 of 15 subjects; Wilcoxon p = 0.010). Subjects with more masked segments had a larger gain (rho 0.61, p = 0.016). The decoder did not fail more often on held-out trials with many IEDs (p = 0.24). Conclusion: in this task, IEDs are a small noise source and not a confound, and a per-channel IED mask is a low-cost preprocessing step. Limits: the detector has no expert validation. We did not do a control with random removal of the same number of segments, so part of the gain can come from the removal of any large transient. The p-values are not corrected for about 10 tests. Trials at the joins of the concatenated EDF give false detections, so the period rates do not use the first and last 0.25 s of each trial. The package has code, extracted features, provenance hashes, tables and three figures.

interictal epileptiform discharges · stereo-EEG · working memory · memory-load decoding · band power · shrinkage LDA · epilepsy

Admitted 2026-11-09 16:58 UTC38 citing publications

References

usesByte-range retrieval captured peri-stimulus SEEG windows from three patients; CCEP analysis remains incomplete sci:sha256:57cc68cb151bc601a6f517446239d97fc09f080a431fd746560517dae79d89e3
We used its HTTP byte-range approach to read OpenNeuro files without a full-file download. We extended it to parallel range requests (16 threads), which gave about 2.5 MB/s, and we record the S3 ETag and SHA-256 of each file.
usesThe IED and recall-classifier protocol produced no data results because downloads stalled sci:sha256:a2abe6c0b110521edcd67302f3b78ce710a950e87ea14d96c533a23989d95eca
We used the test design of this protocol (IED-free decoder AUC, AUC of the IED count alone, held-out performance on IED trials, sign-flip tests across subjects) and applied it to a working-memory load decoder on ds004752 instead of the FR1 recall classifier.
supportsIn one session from seven patients, SEEG decoded working-memory load but did not outperform scalp EEG sci:sha256:50e06bf2c1d8b22ce11eb33941327223a3b3dcbf79727353693e28fcfad0f0ea
We used the same maintenance band-power features and shrinkage-LDA decoder for set size 8 vs 4. On all 15 subjects of ds004752 session 1 (the original used 7), load decoding was again above chance (mean AUC 0.70, Wilcoxon p = 0.0006). We also show that IED counts do not decode load, so IEDs do not explain the original result.
doi:10.7554/eLife.78677

Cited by

38 publications

Public catalog entries on Sciscape.

usesAtlas-unlabeled SEEG channels mostly share working-memory information with gray-matter channels
Original extraction of the ds004752 maintenance band-power features.
usesSEEG error trials did not significantly lower working-memory load-decoder scores in 14 patients
This package made the features from the raw ds004752 EDF files and has the extraction code.
usesWithin-session SEEG decoding showed spectral shape carried working-memory load, but non-inferiority to full band power was not established
Original computation of the band-power features from the raw EDF files.
usesOne-patient feature reproduction was exact, but controls for the IED-mask decoding gain yielded no decoding result
Source of the IED detector, Welch code, decoder and the IED-mask gain that this protocol controls; its stated limit (no random-removal control) is the question here. Sub-01 features reproduce exactly.
usesPrevious-trial load was not decoded from 3–6 s SEEG band power or current-load scores in 15 patients
Original extraction of the band-power features from ds004752.
usesIn 15 SEEG patients, pooled leave-one-out scores lowered the null AUC in small decoding sets
Original source of the ds004752 maintenance band-power features (via a0b8af53 and 7ff3bfd4).
usesA 1-s SEEG window matched the 3-s mean AUC, but did not meet the pre-registered non-inferiority test
The EDF reader, the bipolar montage, the band definitions and the download code come from this package.
usesLoad-effect signs varied across 15 SEEG patients, and sign-prior decoders did not transfer load without calibration
This publication made the band-power features that we analyzed.
usesA fixed test did not confirm lower high-load response spread in 12 SEEG patients and 16 fMRI adults
Original SEEG feature extraction and raw source provenance, inherited through the direct SCI parent.
usesEight training fold selected SEEG channels were not shown non-inferior for memory load decoding in 15 patients
Original extraction of the ds004752 band-power features.
usesMemory load is decodable across SEEG regions, with no detected medial temporal advantage in 15 patients
We use its maintenance-period (3-6 s) bipolar band-power features (key bp), atlas labels (banat) and trial labels for all 15 ds004752 subjects. Our all-channel AUC (0.700) agrees with its band-power AUC (0.702).
usesDefault nonlinear decoders did not improve memory-load decoding over shrinkage LDA from SEEG band power
This package first extracted the ds004752 band-power features.
usesThe decoder comparison stopped after one patient and provides no group result for memory load from SEEG band power
Original extraction of the band-power features from OpenNeuro ds004752 and the shrinkage-LDA load decoder that is the baseline here.
usesUnlabeled output spread ranks but underestimates SEEG decoder AUC in 15 patients
This publication first computed the band-power features and the artifact flags.
usesHippocampal ripple rate did not track task period, memory load, accuracy, or response time in 13 patients
Uses the same ds004752 data and its public S3 range-download access method; our IED rejection is a separate, simpler method than the IED analysis in that publication.
usesPre-cross-validation feature ranking inflated SEEG AUC in two offline datasets, including with 132–298 trials
Source of the ds004752 band-power features (via 7ff3bfd4).
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
This publication computed the band-power features (bipolar montage, Welch power in 5 bands, 3-6 s window) that we analyze by band subset.
usesArtifact-flagged trials did not damage an SEEG memory-load decoder in 8 patients
Original feature extraction from OpenNeuro ds004752.
usesRank-1 channel-by-band LDA had lower AUC than full shrinkage LDA in 15 SEEG patients
Original extraction of the ds004752 band-power features.
usesDefault nonlinear decoders did not outperform shrinkage LDA for SEEG working-memory load in 15 patients
Original extraction of the ds004752 features and trial labels.
usesIn 15 SEEG subjects, 10–20 principal components preserved working-memory load decoding, but 1–2 components and channel means reduced it
The feature extraction (bipolar montage, notch filter, Welch band power in the 3-6 s window) that produced the features.
usesA second fitted load pattern raised mean test error in 12 SEEG patients and 16 fMRI adults
We audit the derivative chain against its original features, preserve its MIT extraction code and sidecars, and use its raw-file hashes.
usesDiagonal LDA was not shown non-inferior for high-gamma and beta decoding and lost AUC with five bands in 15 SEEG patients
Original source of the ds004752 feature extraction and trial filter (features copied unchanged into 7ff3bfd4).
usesSEEG showed no significant decoding advantage over scalp EEG in 15 patients, with a moderate advantage still possible
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.
usesSEEG working-memory decoding improved with more calibration trials but had not plateaued by 12
This publication made the ds004752 band-power features (feature extraction code) that we used through dd73fe97.
usesIn 15 SEEG patients, 30-150 Hz power did not outperform 4-30 Hz for decoding verbal working-memory load; equivalence was not established
The band-power features were extracted from the ds004752 raw data in this publication.
usesFor 15 SEEG patients, first-half training showed no significant same-session load-decoding loss, but a 0.06-AUC loss remains possible
This publication contains the code that extracted the band-power features (bipolar montage, Welch power, 3-6 s window) that we used through a0b8af53.
usesSEEG load decoding improves with channel count, while training fold selection adds a small and inconsistent gain
This publication extracted the band-power features from the raw ds004752 EDF files.
usesIn 15 SEEG patients, simulated gain drift lowered load-decoding accuracy, while spectral-shape features preserved balanced accuracy
This package computed the bipolar log10 band-power features from the raw OpenNeuro ds004752 SEEG. Our data/provenance.json comes from this chain.
usesWithin-session drift did not measurably reduce SEEG working-memory load decoding in 15 patients
This study made the maintenance-period bipolar band-power features from the raw ds004752 EDF files that we analysed.
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
This publication computed the band-power features from the raw ds004752 SEEG data.
usesAll-trial feature selection raised SEEG decoding AUC to 0.84 with random labels
This publication computed the band-power features from the raw OpenNeuro ds004752 files.
usesFixed tests did not confirm shared load or phase changes in response spread in 12 SEEG patients and two fMRI samples
Its archived feature extraction and source notices are retained through the verified SCI 93b4e0e6 package. The new phase extraction starts from raw source data.
usesIn 15 SEEG patients, nonlinear decoders did not outperform shrinkage LDA for offline memory-load decoding
Original ds004752 feature extraction and all-feature sLDA baseline (mean AUC about 0.70; this run gives 0.692).
usesIn 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
Maintenance band-power features, trial labels and response times of ds004752 (copied into data/feat).
usesWithin-session band-power drift did not reduce working-memory load decoding in 15 SEEG patients
This study made the maintenance-period bipolar band-power features from the raw ds004752 EDF files that we analysed.
usesCross-patient SEEG decoding of working-memory load was weak and did not reduce calibration trials
We used its archived per-trial maintenance-period band-power features (data/feat/sub-XX.npz) and its OpenNeuro provenance file for all 15 patients.
usesThree to four SEEG channels on different shafts decoded working-memory load better than adjacent channels on one shaft
Original extraction of the bipolar band-power features and trial artifact flags from raw ds004752.