Aperiodic slope predicts working-memory errors, but residual band features retain cross-patient information in 12 SEEG patients

We tested whether the cross-patient working-memory error signal in stereo-electroencephalography (SEEG) is only an aperiodic (1/f) spectral tilt. The analysis used 12 subjects, 13 sessions and 19,617 channel-trials from OpenNeuro ds006136 v1.0.1, with bipolar middle and frontal gyrus contacts during a load-3 object working-memory task. An error was fewer than three of three items correct. Using leave-one-subject-out decoding, band power gave median AUC 0.590 (10/12 subjects above 0.5). Aperiodic slope and level alone predicted errors (median AUC 0.545; 10/12; p=0.001), but performed below band power (median AP−BP difference −0.036; AP better in 2/12; p=0.042). Band residuals after subtracting a straight-line fit to the log-log spectrum also predicted errors (median AUC 0.617; 10/12; p=0.005). The aperiodic slope was flatter on error trials during encoding in 11/12 subjects. A fixation-window aperiodic control did not predict errors (AUC 0.505; p=0.34). These results indicate that a flatter aperiodic spectrum is part of the error signal, while band-limited changes, mainly lower theta/alpha power, retain information beyond the aperiodic fit. A cross-patient error monitor should retain band features. The study is limited to 12 subjects, a single straight-line aperiodic model without a knee, bipolar MTG and MFG contacts, and reading effects; it does not examine error type or causal/stimulation effects.

stereo-electroencephalography · working memory · error decoding · aperiodic spectrum · band power · leave-one-subject-out · neural signal processing

Admitted 2026-11-09 16:54 UTC4 citing publications

References

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
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.
Across 12 SEEG patients, cross-patient decoding predicted working-memory errors with median AUC 0.59 sci:sha256:38c12432900cb2ceb1750b02e0755c7a987e2ada08cb5cc4d95a36e5d09d7827
[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.
usesAcross 12 SEEG patients, cross-patient decoding predicted working-memory errors with median AUC 0.59 sci:sha256:38c12432900cb2ceb1750b02e0755c7a987e2ada08cb5cc4d95a36e5d09d7827
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.
usesdoi:10.18112/openneuro.ds006136.v1.0.1
The analysis reports OpenNeuro ds006136 v1.0.1 as its SEEG source; its data manifest records the source keys, ETags, sizes and dates for 121 dataset files.

Cited by

4 publications

Public catalog entries on Sciscape.

usesIn 7 SEEG subjects, white-matter channels had 2–3× less broadband power than gray-matter channels, but the 1/f exponent did not separate them
That work uses the 1/f exponent as a channel feature. Here the exponent did not mark tissue type, but broadband power did.
supportsIn 9 patients, amygdala band power decoded aversive blocks, but the 1/f exponent alone did not
As in that ds006136 working-memory result, the 1/f slope decoded less than band power, and band residuals after the straight-line 1/f fit still carried information. We show the same pattern for an aversive state in amygdala SEEG (ds003374).
usesIn 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)
Earlier analysis of non-oscillatory features on the same dataset. We compare a different non-band-power feature (burst fraction).
supportsIn FR1, broadband slope predicts the channel-mean theta decrease in 18 SEEG patients; 8-band power decodes recall better
In a different task and dataset (ds004789 FR1 recall), we find the same pattern: a 1/f line fit decodes less well than band power, and band residuals after the line fit keep information (residual decoder AUC 0.638 vs 8-band 0.669).