Encoding and delay SEEG band power predicted later working-memory errors in 12 of 12 subjects (median within-session AUC 0.64)
In 12 subjects with stereo-electroencephalography (SEEG) from bipolar middle temporal and middle frontal gyrus contacts during a load-3 object working-memory task, a shrinkage linear discriminant ensemble using encoding- and delay-period band power predicted later errors above chance in every subject (median cross-validated AUC 0.642; one-sided Wilcoxon p = 0.00024). Seven of 13 eligible sessions passed a 200-permutation test (Fisher combined p = 2.7e-8). Fixation-only decoding was at chance (median AUC 0.507, p = 0.37), and error clustering over trial time did not explain decoder performance. Before errors, encoding low-gamma power was higher in 12/12 subjects (Bonferroni p = 0.005); theta and alpha power were lower, but these effects did not survive Bonferroni correction. The effect was modest and supports within-session prediction from the sampled cortical contacts; it does not establish reliable single-trial prediction, transfer across sessions or subjects, causality, or benefit from stimulation. Trials were selected and artifact-rejected by dataset authors, no new artifact rejection was applied, and the analysis was not preregistered.
working memory · stereo-electroencephalography · SEEG · error prediction · linear discriminant analysis · band power · brain-computer interface · cross-validation
References
2- usesdoi:10.18112/openneuro.ds006136.v1.0.1
The analysis used OpenNeuro ds006136 version 1.0.1, as recorded in the publication source and exact-version dataset metadata.
- doi:10.1016/j.neuroimage.2026.121718
Cited by
10 publicationsPublic catalog entries on Sciscape.
- usesMatched-count MTG and MFG SEEG channels did not differ significantly in working-memory error prediction
Error label and shrinkage-LDA ensemble method; this work adds a matched-count region comparison.
- usesSingle-channel SEEG predicted working-memory errors above chance in 12 patients, but the LMFG–RMTG advantage was not significant after correction
Same task, labels and band-power error decoding approach on ds006136; we apply it per region.
- supportsAcross 12 SEEG patients, cross-patient decoding predicted working-memory errors with median AUC 0.59
We reproduce its within-subject result (median AUC 0.60 vs 0.64) with the same band and window choices, and extend it: a decoder trained on other subjects gives the same AUC, so the error signal is shared across patients.
- usesSEEG band power up to 1 s into the working-memory delay predicts later recall errors (median AUC 0.60, 12 subjects); the rest of the delay adds no significant accuracy after correction
Same decoder, features, inclusion rule and label; this work adds the time course and a causal decision-time analysis.
- supportsIn 13 sessions, evoked SEEG voltage predicted working-memory errors less well than band power, while induced and total band power performed similarly
With its unchanged pipeline on all 13 sessions with at least 10 errors (it reported 12), encoding+delay band power predicted errors with median within-session AUC 0.651, 13 of 13 sessions above 0.5. This agrees with its reported median AUC 0.64.
- usesIn 13 sessions, evoked SEEG voltage predicted working-memory errors less well than band power, while induced and total band power performed similarly
We use its ds006136 loader, Welch band-power features, per-channel shrinkage-LDA ensemble and stratified cross-validation without change (code/owm_decode_18254913.py). With this code, band power gave median within-session AUC 0.651 in 13 sessions, which agrees with its reported 0.64. We add evoked-voltage and induced-band-power feature sets to the same pipeline.
- 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)
[Legacy relation: reproduces] We used the same dataset and found again that band power predicts working-memory errors (median AUC 0.63 vs 0.64 in that work), with a different pipeline.
- supportsA leave-one-subject-out SEEG decoder predicts working-memory errors above chance in 10 of 12 patients (median AUC 0.60)
Leave-one-subject-out weights have the same direction in all folds (low gamma up, theta/alpha down before errors), which supports its univariate finding and extends it to transfer across subjects.
- usesA leave-one-subject-out SEEG decoder predicts working-memory errors above chance in 10 of 12 patients (median AUC 0.60)
Reuses its band-power feature code, shrinkage LDA, label definition, inclusion rule and data recipe for ds006136.
- supportsAperiodic slope predicts working-memory errors, but residual band features retain cross-patient information in 12 SEEG patients
With a different pipeline on the same dataset, within-subject band-power decoding of errors is above chance (median AUC 0.604, 10/12 subjects; lower than its 0.64 and 12/12), and fixation-window aperiodic features do not predict errors.
- usesIn 12 SEEG patients, first-half calibration gave later-trial AUC 0.56, not above chance after Holm correction or significantly below random-split AUC 0.59
Uses the within-patient error-prediction task and the decoder design (per-channel shrinkage LDA on SEEG band power) that this publication introduced for ds006136.
- supportsIn 18 session-0 SEEG patients, pre-word band power did not predict free recall within serial position or improve the encoding decoder
18254913 found that pre-stimulus SEEG power did not predict working-memory errors in ds006136, but encoding and delay power did. We found the same pattern for delayed free recall in a different dataset (ds004789, 18 patients): pre-word power did not predict recall within serial position (median AUC 0.516, p = 0.25), but encoding power did (0.618, p = 5e-5).