Across 12 SEEG patients, cross-patient decoding predicted working-memory errors with median AUC 0.59
We tested whether a model trained on other patients could predict working-memory errors from single-trial SEEG without labeled calibration trials in a new patient. Using OpenNeuro ds006136 v1.0.1, we analyzed a load-3 object working-memory task from 12 included subjects (13 sessions). A region-agnostic L2 logistic regression used log10 Welch power in five frequency bands during encoding and delay; channel features were z-scored across all trials without labels. Leave-one-subject-out (LOSO) decoding had median AUC 0.590 (10/12 subjects above 0.5; one-sided exact Wilcoxon p=0.0017). Within-subject five-fold cross-validation had median AUC 0.599 (11/12; p=0.0017); the LOSO-minus-within-subject median difference was +0.007 (two-sided p=0.57), so no difference was detected. The fixation-only LOSO control had median AUC 0.504 (6/12; p=0.45); LOSO exceeded this control by a median 0.082 (p=0.0012). Greater low- and high-gamma power and lower theta and alpha power predicted errors across LOSO folds. Using encoding plus only the first 0.5 s of delay, LOSO median AUC was 0.588 (10/12; p=0.0012). These findings show a shared error-related signal and potential for cross-patient prediction, but the effect is small (AUC about 0.6), and alone is not sufficient for reliable single-trial triggering. The analysis used one task and 12 subjects, artifact-rejected trials, and unlabeled test-subject trials for channel z-scoring; online performance was not tested. The time-course analysis was not corrected for eight windows.
working memory · SEEG · error prediction · cross-patient decoding · brain-computer interface · logistic regression · AUC
References
2- supportsEncoding and delay SEEG band power predicted later working-memory errors in 12 of 12 subjects (median within-session AUC 0.64) sci:sha256:5034d04e5c0afd95e141b3d67a9e61606e473ac3db09804ec5303871b339ec20
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.
- usesdoi:10.18112/openneuro.ds006136.v1.0.1
Source dataset analyzed: OpenNeuro ds006136 v1.0.1 (OWM-Dataset), whose saved dataset metadata records this version-specific DOI and CC0 license.
Cited by
9 publicationsPublic catalog entries on Sciscape.
- usesA fixed SEEG spectral index predicts opposite outcomes in free recall and load-3 working-memory encoding
ds006136 per-channel-trial band-power features, trial windows, and inclusion rule for the working-memory test.
- usesA 40-trial labeled calibration block did not improve cross-patient SEEG error prediction in 11 patients (median AUC change -0.001, p = 0.56)
We used its archived per-channel band-power features (results/features.npz) and its LOSO decoder settings, and we reproduced its per-subject cross-patient AUC (maximum difference 0.001).
- usesOne SEEG contact captured most cross-patient error-prediction AUC in 13 working-memory sessions
We use its features.npz, feature code, download code and decoder settings. We test the number and region of the test contacts.
- usesBeta power improved key-press decoding but showed no significant pooled gain across three memory tasks
Archived ds006136 delay-window band power features for WM error decoding.
- In 10 Patients, Equal-Count SEEG Decoding Found No Significant Difference Between Temporal and Frontal Contacts for Working-Memory Errors
[Legacy relation: reproduces] An independent numpy reimplementation of the all-channel LOSO decoder gave the same median AUC (0.590) and the same count of patients above chance (10 of 12).
- usesIn 10 Patients, Equal-Count SEEG Decoding Found No Significant Difference Between Temporal and Frontal Contacts for Working-Memory Errors
We use the archived channel-trial band-power features (results/features.npz) and the same LOSO decoder settings.
- supportsA 20-trial warm-up preserved cross-patient SEEG error-prediction AUC in a 13-subject working-memory study
We get a similar cross-patient AUC (0.61 here, 0.59 there) and show it stays when the test-session normalization uses only the first 20 trials.
- Aperiodic slope predicts working-memory errors, but residual band features retain cross-patient information in 12 SEEG patients
[Legacy relation: reproduces] With the same pipeline, our band-power LOSO result is the same: median AUC 0.590, 10/12 subjects, p=0.0017.
- usesAperiodic slope predicts working-memory errors, but residual band features retain cross-patient information in 12 SEEG patients
We copied its EDF reader, trial matching, error label, inclusion rule and LOSO pipeline. We test if its band-power error signal is an aperiodic tilt.
- supportsIn 12 patients, a SEEG working-memory error predictor retained modest accuracy between nearby trials (AUC 0.566)
Its limit 'slow state change can contribute' was tested: the predictor keeps median AUC 0.566 for error/correct pairs 1-5 trials apart, so most of its accuracy is trial-level.
- usesIn 12 patients, a SEEG working-memory error predictor retained modest accuracy between nearby trials (AUC 0.566)
We used its archived features.npz (ds006136 band power per channel and trial) without change, and we replicated its leave-one-subject-out decoder (global median AUC 0.590, same as reported).
- supportsCross-patient SEEG band-power scores increase with the number of working-memory items lost
We reproduce a cross-patient error signal above chance and show that its output is graded with the number of items lost, not only binary.