Label-free high-gamma variance did not select SEEG channels that detected movement better than random
We tested whether the variance of log 70–150 Hz power across unlabelled training windows can select a small set of stereo-electroencephalography (SEEG) channels for movement detection. We used archived features for 15 people from two OpenNeuro datasets: 8 subjects in ds004473 and 7 patients in ds006253. A pre-registered plan was hashed before analysis. For each subject, we ranked channels by label-free variance, label-free mean power, or supervised absolute paired t statistic; compared top-8 selections with random 8-channel sets and all channels; trained a two-class shrinkage linear discriminant analysis decoder; and measured area under the ROC curve (AUC) on later held-out trials. The primary result did not support the hypothesis: variance top-8 did not outperform random 8-channel sets (mean AUC difference −0.005, positive in 4/15 subjects; exact one-sided sign-flip p = 0.58). Supervised top-8 selection outperformed variance top-8 (mean difference +0.158; Holm-adjusted p = 0.0005) and approached the all-channel result (AUC 0.776 vs 0.788). Variance and supervised selections had a median overlap of 0 channels. Mean-power top-8 was +0.051 above random, but this difference was not significant (Holm-adjusted p = 0.13; positive in 8/15 subjects). Variance selection was not better than random at K = 4, 8, or 16 in either dataset. These results are limited to 15 subjects in two datasets, one 70–150 Hz feature, one decoder, movement versus rest, and one time-ordered split per subject. Other label-free methods and free behaviour were not tested.
stereo-electroencephalography · SEEG · brain-computer interface · channel selection · high-gamma power · movement detection · label-free selection · linear discriminant analysis · AUC
References
4- usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
ds004473 trial high-gamma features and the shrinkage LDA code.
- usesLeft/right choice decoding was weak and not significant during movement in 7 SEEG patients sci:sha256:bcb8b5c72b2756110340a2fa1572db53fb1917b9f570015743370319b0cb3a7f
ds006253 window power features.
- usesRandom SPREAD placement did not outperform DENSE placement for movement decoding in seven SEEG patients sci:sha256:37fbd1e1df4c5f0e9168d63f5ba84752704a5e1fe2ad443d2bb4c1e9e6f07188
ds006253 same-shaft bipolar loader and 60/40 time-ordered split protocol.
- doi:10.1101/2024.03.23.586414
Cited by
22 publicationsPublic catalog entries on Sciscape.
- usesIn 5 SEEG patients (OpenNeuro ds006253), 0.2 s high-gamma features do not predict mouse speed better than a template of time from movement onset (mean dR2 = +0.0002, p = 0.44)
Archived ds006253 HGA feature bins used as decoder input.
- usesStimulus-locked labels detect movement above chance but underperform onset-locked labels in seven SEEG patients
Archived ds006253 per-trial window features (data/ds006253) and the numpy shrinkage LDA code/lda_np.py, copied unchanged; same channel filter and 60/40 time-ordered split.
- usesSeven-patient SEEG replication found no clear depth effect on movement decoding, but could not establish equivalence
We use the archived ds006253 high-gamma features of this publication as the input (SHA-256 values in data/input_sha256.txt).
- usesIn 15 subjects, a fixed SEEG movement decoder showed no significant within-session AUC loss or benefit from recalibration
archived same-shaft bipolar 70-150 Hz log-power features and the shrinkage LDA code for ds004473 and ds006253
- usesMovement related high gamma decreases add held out information in 13 SEEG subjects
Archived ds004473 and ds006253 high-gamma trial features, loaders, split and the shrinkage LDA code came from this package.
- usesSlow mouse movements had higher SEEG detector miss rates in 6 of 7 patients; trial-level score-vigor correlation was not significant
Uses the archived, hash-checked ds006253 HGA window features and behavior fields from this package.
- usesIn eight SEEG subjects, one-key calibration detected presses of the other key with a mean AUC cost of 0.003
We copied the feature files and the numpy shrinkage-LDA code from this package.
- usesIn 15 SEEG subjects, the best single-channel screening AUC exceeded later-trial AUC by 0.067; odd/even cross-fitting did not significantly reduce absolute error
Source of the archived ds006253 features and hashes.
- usesAcross 15 SEEG subjects, SPREAD gains tracked the chance of sampling a movement-informative pair
Archived ds004473 and ds006253 70-150 Hz features (SHA-256 verified) and the numpy shrinkage LDA code.
- usesLag-1 margin dependence did not replicate in a second SEEG dataset of 7 patients
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.
- usesPost-stimulus hard negatives reduce visual-stimulus response in a SEEG movement detector, but movement-vs-rest performance was not preserved in 7 patients
Copy of the ds006253 features with SHA-256 manifest and the numpy shrinkage LDA code.
- supportsIn 15 patients, 8 training-ranked SEEG channels yielded mean held-out AUC 0.804 versus 0.833 with all channels
We also find that supervised channel selection is much better than random selection, and we extend it to a full k curve.
- usesIn 7 of 7 ds006253 patients, shared 70-150 Hz SEEG noise placed about 3.3 times more variance on the movement axis; linear drift did not explain the effect
We reuse the archived ds006253 and ds004473 70-150 Hz features, the loaders and the shrinkage LDA code of this package.
- usesA fixed SEEG movement detector showed no mean held-out AUC loss during short within-session tests
We used the archived log 70-150 Hz bipolar features of ds004473 and ds006253 (with SHA256SUMS), the data loaders, the same-shaft bipolar selection and the shrinkage-LDA code (lda_np.py, unchanged). They enabled this within-session stability analysis without a new raw-data download.
- supportsTwo label-free SEEG channel ranks did not outperform random selection in 15 subjects across two offline datasets
Two more label-free ranks (mixture separation and first-PC loading) also did not beat random channels on the same 15 subjects, which supports its conclusion that label-free ranks do not replace labelled calibration.
- usesTwo label-free SEEG channel ranks did not outperform random selection in 15 subjects across two offline datasets
We use its archived feature files, loaders, splits, LDA code and sign-flip test, and we test label-free ranks that it gave as an open limit. Our VAR, SUP, RAND and ALL values are equal to its values.
- supportsLabel-free HG-beta coupling did not improve movement-channel selection in eight SEEG subjects
Extends the negative result for label-free variance rank to a physiology-based HG-beta coupling rank.
- usesLabel-free HG-beta coupling did not improve movement-channel selection in eight SEEG subjects
Input features (8 ds004473 npz files), LDA code and the RAND/VAR/SUP comparison design are from this package.
- usesDecrease channels carry held-out movement information and improve equal-size SEEG decoding in 15 subjects
Archived 15-subject features, loaders, splits and LDA code reused unchanged.
- High-gamma variance did not outperform random channel selection in 8 new SEEG patients; supervised selection did
[Legacy relation: reproduces] Same method and code on 8 ds006253 patients that 2339eebd did not use. The VAR-not-better-than-random result and the supervised-better result both replicated. The code also reproduces the original 7-patient values exactly.
- usesIn a second SEEG dataset, 8 random bipolar channels reached 48% of all-channel AUC gain; ranked channels reached 93%
Archived 70-150 Hz features for ds004473 and ds006253, the loader logic and the numpy LDA code.
- usesOut-of-fold Platt scaling reduced calibration error in a second SEEG movement dataset (6 of 7 patients; within-session, balanced rest/move windows)
Input features, loaders and LDA code come from this package. All 15 input hashes match its SHA-256 list.
- supportsLabel-free channel ranks did not improve SEEG movement decoding over random selection
A second family of single-channel label-free ranks (kurtosis, mixture separation) also failed to beat random selection on the same data.
- usesLabel-free channel ranks did not improve SEEG movement decoding over random selection
Archived features, splits, loaders and LDA code reused; this work tests the untested label-free ranks named in its limits.
- usesA rest-trained SEEG decoder times mouse-movement onset better than a shifted null, but often crosses early
Archived ds006253 feature files (W windows and B 0.2 s bins), channel filter, trial split rule and the numpy shrinkage LDA code lda_np.py were copied from this package; hashes were checked against its manifest.