Unlabeled rest covariance did not improve few-trial SEEG movement AUC in 15 patients

We tested whether an unlabeled rest-window covariance improves few-trial calibration of an offline SEEG movement detector. Archived 70–150 Hz HGA window features from 15 patients in OpenNeuro ds006253 were split in time: the first 60% of trials formed the pool and the last 40% the test set. We compared shrinkage LDA using n labeled trials (S) with LDA using the same class means and a Ledoit–Wolf covariance estimated from all pool rest windows (R), averaging 100 random draws for n = 5, 10, 20, and 40. The pre-registered primary test did not show an improvement at n = 10: mean AUC(R) − AUC(S) = +0.0034, with 8/15 patients above zero and exact one-sided sign-flip p = 0.263 (alpha = 0.05). Differences were near zero at all n (means +0.0055, +0.0034, +0.0020, and −0.0018 for n = 5, 10, 20, and 40). Mean AUC for standard LDA was 0.692 at 5 trials, 0.800 at 40, and 0.834 using all 72–180 pool trials. In this dataset, rest-only covariance did not reduce AUC, but did not shorten calibration. The result is limited to 15 patients, one session per patient, offline event-locked windows with known onset times, one feature type and classifier family; rest windows came from the same session as the test. Secondary n values were not corrected for multiple tests. The recorded source hashes identify files but do not resolve the exact OpenNeuro dataset release.

SEEG · brain-computer interface · movement detection · linear discriminant analysis · covariance · few-trial calibration · AUC

Admitted 2026-11-09 20:09 UTC2 citing publications

References

supportsUnlabeled task covariance lowers few-trial SEEG movement-detection AUC, while rest-only covariance does not sci:sha256:4ce49e075d97069bb4064550cdaf8385bac1576311d12ad9e6c9000023c1fdaa
In a second dataset (ds006253), a rest-only covariance again gave no reliable AUC change from standard few-trial LDA.
usesTwenty calibration trials gave lower AUC than full calibration in a seven-patient SEEG movement task sci:sha256:7b1fe217b6d413f6b0fff5be7556cc4cd4f0c3094b83327dbc164ee4a2f9ecf4
This work tests a method to reduce the few-trial AUC loss that this publication reported in ds006253.
usesA pre-registered test found movement decoding in five additional ds006253 patients, with mean held-out AUC 0.874 sci:sha256:5b7e7453bf9d35bc1113c364ab3daac7fd7cdc393cb61b160586cbaf516dc375
Archived ds006253 window features of 15 patients and the preprocessing and LDA code come from this package.

Cited by

2 publications

Public catalog entries on Sciscape.

usesSpatial smoothing did not improve few-trial SEEG movement detection with shrinkage LDA (ds006253, 15 patients, 5–40 trials, fsaverage distances)
Same features, split, draws and standard LDA; standard values reproduced exactly.
supportsIn 15 SEEG patients, zero-label calibration did not detect mouse movement above chance
That work found that unlabeled rest windows did not improve few-trial LDA. This work finds that unlabeled MOVE and REST windows alone do not give the movement axis.