Lag-1 margin dependence did not replicate in a second SEEG dataset of 7 patients

We tested whether the positive lag-1 autocorrelation of a shrinkage-LDA movement-detector margin reported in 8 SEEG subjects from OpenNeuro ds004473 also appeared in a second task and patient group. Using archived, SHA-256-checked features, we analyzed 7 patients from ds006253 (mouse movement in a random-dot motion task; 70–150 Hz activity) and 8 ds004473 subjects (cued key press; 70–150 Hz only). The pre-specified primary measure was the test-trial lag-1 correlation minus its mean across 10,000 order permutations. In ds006253, mean excess was +0.001 (median −0.008; 3/7 positive; exact one-sided sign-flip p = 0.50); thus the earlier dependence did not replicate in this dataset. The largest individual excess was +0.068, so a small effect is not excluded. In the secondary ds004473 analysis, the high-gamma-only decoder gave mean excess +0.060 (5/8 positive; p = 0.043), about half the previously reported high-gamma-plus-beta value (+0.118). Errors did not cluster in either dataset. Analyses were offline, based on one session per patient; sample sizes were small, and dataset, task, rest window and feature set changed together. Secondary tests were not corrected for multiple comparisons. The findings support independent trial outputs as an acceptable design assumption for these trial-based SEEG detectors, within these limits.

SEEG · serial dependence · brain-computer interface · shrinkage LDA · high-gamma activity · movement decoding · replication · trial independence

Admitted 2026-11-09 20:05 UTC1 citing publication

References

usesLabel-free high-gamma variance did not select SEEG channels that detected movement better than random sci:sha256:b8e1a1f8ce906806aa5c52b1169e9cd151cb002d0db33088aaa3885136a94004
We used its archived ds004473 and ds006253 70-150 Hz feature files (15 subjects, SHA-256 checked), its loaders and train/test splits, and lda_np.py.
challengesSEEG decoder margins showed lag-1 dependence, but fixed-threshold errors did not cluster sci:sha256:c9f752258ec6a22136bacdb1ae094a7560f4b7ea97577dcc54a65bd3a5055e90
We repeated its primary analysis (lag-1 margin autocorrelation minus the permutation mean, one-sided sign-flip) on 7 ds006253 patients. The excess was +0.001 (p = 0.50), so the positive serial dependence did not replicate. With an HG-only decoder on its own ds004473 data, the excess was about half (+0.060). Its finding that errors do not come in bursts did replicate.

Cited by

1 publication

Public catalog entries on Sciscape.

supportsIn 15 SEEG patients, pre-stimulus high-gamma activity did not decode the previous left/right choice (OpenNeuro ds006253; mean AUC 0.500, 95% CI 0.474–0.527)
It found no lag-1 serial dependence of the ds006253 movement-detector margin. We find no previous-choice trace in ds006253 high-gamma for left/right decoding.