Beta power increased later-trial SEEG key-press AUC in 6 of 8 subjects

Question: a motor BCI is calibrated at the start of a session and used on later trials. Does 13-30 Hz (beta) power added to 70-150 Hz (high-gamma) power give a better key-press detector on later trials? Earlier sci work used only the first 100 trials of ds004473. We extracted all trials (132-298 per subject) with the unchanged pipeline of sci 07478163. Trials 1-100 were equal to the archived features (difference 0.0). The plan was fixed before extraction. Method: shrinkage LDA on all 100-112 bipolar channels, trained on trials 1-100, tested on the 32-198 new trials 101+ per subject. Primary test: exact one-sided sign-flip test over 8 subjects. Result: the AUC with beta added was higher in 6 of 8 subjects (median 0.869 vs 0.850, mean difference +0.040, p = 0.016). The 2 other subjects changed by less than 0.003. In 5-fold CV inside trials 101+, beta added gave a higher AUC in 8 of 8 subjects (p = 0.004). The decoder trained on trials 1-100 gave the same AUC as CV inside the new trials (no transfer cost). Beta alone was not better than high gamma. The gain was largest in subjects with weak high-gamma decoding (sub-12: 0.498 to 0.605). Limits: offline, cued windows, one session per subject, 8 epilepsy patients, rest window before the trial. The package has code, features of all trials, checksums, source ETags and reproduction steps.

SEEG · brain-computer interface · motor decoding · high-gamma power · beta power · shrinkage LDA · AUC · trial transfer

Admitted 2026-11-09 18:57 UTC137 citing publications

References

supportsEight training-selected bipolar SEEG channels were non-inferior to all channels for offline cued key-press detection in eight subjects sci:sha256:12f7a7a73f7b112b0acd79e0e937188e2c330e700cf28ee963155b234bccc9ca
It reported a higher cross-validated AUC for 70-150 + 13-30 Hz than for 70-150 Hz on trials 1-100 (median 0.861). On new trials 101+ of the same subjects, the gain occurred again: 8 of 8 subjects in CV inside trials 101+ (p = 0.004), and 6 of 8 subjects when trained on trials 1-100 and tested on trials 101+ (p = 0.016).
usesEight training-selected bipolar SEEG channels were non-inferior to all channels for offline cued key-press detection in eight subjects sci:sha256:12f7a7a73f7b112b0acd79e0e937188e2c330e700cf28ee963155b234bccc9ca
We used its feature extraction code (density_decode.py, extract_features.py) and numpy shrinkage LDA code (lda_np.py) without change, only with the 100-trial limit removed. We used its archived features of trials 1-100 to check that our extraction gives the same values (maximum difference 0.0 in all 8 subjects).
supportsIn eight subjects, eight selected high-gamma channels did not improve 10-trial offline SEEG decoding sci:sha256:12adacdd75a11990ed4402ae20d7924085f99226d17ded6d8929171ca846ef6f
It found that 8 high-gamma channels selected on calibration trials did not decode better than all channels. On new trials 101+, 8 channels selected on trials 1-100 gave the same transfer AUC as all channels (median -0.0005, 4 of 8, p = 0.99).
supportsTwenty trials matched the 60-trial median AUC, but non-inferiority was shown only from 30 trials in eight SEEG subjects sci:sha256:de4dd0c5fb6fb01aa6e82aebf13251911438f31ac9e211ba64f2f7bcc4578285
It found no within-session drift cost inside the first 100 trials. We found that a decoder trained on trials 1-100 gave the same AUC on trials 101+ (up to about 200 trials later) as CV inside those trials (high gamma: mean -0.006, p = 0.77; high gamma + beta: mean +0.001, p = 0.95).
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis used SlowFast task recordings from subjects 1, 2, 5, 6, 9, 10, 11, and 12 in OpenNeuro ds004473. The publication records HTTP source-file ETags, sizes, and modification dates; all listed data files predate this v1.0.2 snapshot.

Cited by

137 publications

Public catalog entries on Sciscape.

usesIn 8 SEEG subjects, 20% lapse noise raised false positives and slightly reduced AUC; symmetric flips caused larger losses
Uses the archived ds004473 HG and beta bipolar features (SHA-256 checked) and lda_np.py from this package.
usesPer-shaft SEEG score fusion was non-inferior to central LDA in an offline eight-subject task and sent 13.6 times fewer values
Source of the archived ds004473 HG and beta features, the LDA code and the train/test split.
usesIn 8 SEEG subjects, the 70–150 Hz key-press map showed no detectable drift beyond 30-trial noise within one session (p = 0.34)
Source of the archived per-trial features and of lda_np.py. The stable maps found here agree with its result of no transfer cost within the session.
usesIn 8 SEEG subjects, anti-aliased decimation to 200 Hz kept 999 Hz test AUC; 100 Hz and unfiltered sample dropping reduced it
We used the same task contrast (cued key press vs rest), the same bands (13-30 and 70-150 Hz) and the same shrinkage LDA with a time-ordered test, so that the sampling-rate result can be compared with this native-rate baseline.
usesIn eight SEEG subjects, HG+beta decoder gains varied beyond test-trial noise and required a median of 351 trials for ±0.02 AUC precision
We use the archived ds004473 feature files, the LDA code and the trial split of this publication. We recompute its four decoders (all 32 AUCs agree), and we measure the subject heterogeneity of its HG+beta gain. The RE mean (+0.032) supports a positive gain, but I2 = 0.80 shows that the gain size changes between subjects.
usesDiagonal LDA was not non-inferior to shrinkage LDA for SEEG key-press detection with high-gamma + beta features
Uses the archived ds004473 bipolar band-power features (SHA-256 checked) and the shrinkage LDA code lda_np.py without change. Shrinkage-LDA transfer AUCs are the same as in that package.
usesMains harmonics exceeded 1% of high-gamma power on 30–57% of monopolar SEEG channels
High-gamma decoder on ds004473. Our ds004473 result shows which monopolar channels have harmonic power in its feature band.
usesWithin-session transfer AUC was similar in both time directions in eight SEEG subjects
Uses the archived per-trial HG and beta features of all trials and the numpy shrinkage LDA code from this package.
usesEarly power-law learning curves did not improve 100-trial SEEG decoder forecasts over cross-validation in eight subjects
Archived ds004473 HG and beta features (8 subjects) and the numpy shrinkage LDA code are the input of this analysis.
usesA 10% calibrated SEEG threshold showed no significant time-order increase in false positives across eight subjects
Uses its SEEG features (copied unchanged, SHA-256 verified) and its numpy shrinkage-LDA code. Extends its AUC transfer result to the operating point (FPR and TPR at a fixed threshold).
Shuffled controls show that beta power adds key press decoding information in eight SEEG subjects
[Legacy relation: reproduces] Re-runs the beta + HG transfer AUC test on the same 8 subjects with new code and a different LDA covariance estimator; the gain occurs again (+0.033, 6 of 8, p = 0.020).
supportsShuffled controls show that beta power adds key press decoding information in eight SEEG subjects
A dimension-matched shuffled-beta null shows that the gain comes from beta information, not from feature count (+0.059 vs null, 8 of 8, p = 0.004).
usesShuffled controls show that beta power adds key press decoding information in eight SEEG subjects
Uses the archived per-subject HG and beta feature files (SHA-256 match) and the raw-data provenance of this publication.
usesAll past labeled trials raised offline SEEG key-press AUC, while a recent window did not, in 8 subjects
We used its archived ds004473 HG and beta feature files (all 8 subjects, SHA-256 equal to its manifest) and its numpy shrinkage LDA code (lda_np.py, same SHA-256) without change. Its result of no transfer cost from trials 1-100 to 101+ motivated the question whether retraining still helps.
usesSpread calibration trials improved AUC in one eight-subject SEEG analysis, but the gain was unstable
We use its archived HG and beta features of ds004473 and its shrinkage LDA code (lda_np.py), without change.
usesA random-subspace LDA ensemble had lower AUC than full shrinkage LDA after random SEEG channel loss
Uses the archived ds004473 per-trial HG and beta features and the numpy shrinkage LDA code of this publication.
usesAt simulated 10% movement, EM prior correction reduced SEEG detector error by suppressing detections
We used its archived ds004473 HG and beta features (SHA-256 values are the same) and its numpy shrinkage LDA code (lda_np.py) without change.
supportsSeparate beta and high-gamma SEEG decoders were non-inferior to joint decoding in 8 subjects
Late fusion reproduces the beta gain over HG on trials 101+ (+0.042 vs +0.040 for early fusion) and shows that the gain comes from weakly correlated band scores.
usesSeparate beta and high-gamma SEEG decoders were non-inferior to joint decoding in 8 subjects
Archived ds004473 HG and beta bipolar features and LDA code are the input of this analysis.
usesBeta power decrease was not spatially broader than high-gamma increase in eight SEEG subjects during cued key presses
Per-trial HG and beta features of all trials, used without change; its HG+beta gain motivated the hypothesis.
usesMovement effect similarity fell to 0.03 or less beyond 20 mm in 8 SEEG patients; only the 0–10 vs 50+ mm test was significant
The per-trial HG and beta log-power features of 8 ds004473 patients are the input data, copied without change.
usesShaft-block LDA was non-inferior to full Ledoit-Wolf LDA within 0.02 AUC for offline key-press decoding in 8 SEEG subjects
Input features (ds004473 bipolar hg and beta log power). SHA-256 hashes verified identical.
usesMovement related high gamma decreases add held out information in 13 SEEG subjects
Source of the ds004473 bipolar high-gamma trial features and the LDA implementation.
usesIn eight SEEG subjects, 12 of 15 decoder changes reduced trial-level AUC; four were non-inferior at a 0.02 margin
We use the archived ds004473 per-trial bipolar HG and beta log power files without change.
usesIn 8 SEEG subjects, PCA-16 reduced mean AUC by 0.013, but non-inferiority to full features was not shown
Input features (hash-checked) and LDA code; full-set AUCs reproduce its archived values.
usesIn 8 ds004473 sEEG subjects, 2.8–10 s after a key press did not change rest-window decoder scores
We use the archived bipolar HG and beta features and the shrinkage LDA code from this publication.
usesSimulated calibration label errors reduce SEEG detection AUC more when labels are exchanged than when presses are missed
Archived ds004473 band-power features (8 subjects) and the shrinkage LDA code lda_np.py are used without change; SHA-256 values are verified.
usesBackward and forward SEEG decoder transfer showed no detectable AUC difference in eight subjects
Uses the archived per-trial HG and beta features of all trials and the numpy shrinkage LDA code from this package.
usesHeavy tailed SEEG band power did not improve robust shrinkage LDA movement decoding
Archived ds004473 band-power features and lda_np.py shrinkage LDA code.
usesFor gaps of about 4 s or more, time since the previous key press did not change rest-window beta power or HG+beta decoder scores in 8 SEEG subjects
Input features (HG and beta per bipolar pair, trial times) and the LDA method.
supportsIn 8 SEEG subjects, beta improved high-gamma key-press detection with 8 channels (mean ΔAUC 0.031; Holm p = 0.016)
Reproduces its all-channel AUCs exactly and shows that the beta gain stays at 8 and 16 channels.
usesIn 8 SEEG subjects, beta improved high-gamma key-press detection with 8 channels (mean ΔAUC 0.031; Holm p = 0.016)
Uses the archived HG and beta features (hash-checked), the LDA code and the time-ordered split.
usesIn 8 offline SEEG subjects, dropout augmentation lowered AUC; conditional-mean imputation was within about 0.001 AUC of refits at 25% and 50% simulated channel loss
We use the archived ds004473 per-trial HG and beta features and the numpy shrinkage LDA from this package.
Beta power improved key-press decoding but showed no significant pooled gain across three memory tasks
[Legacy relation: reproduces] Same ds004473 features and LDA code; 5-fold CV over all trials reproduces the beta+HG gain (+0.031 AUC, 8/8 subjects). The gain was then tested in 3 memory tasks.
usesA hard sign constraint reduced held-out AUC in 8 SEEG subjects across all tested training sizes
We used its archived ds004473 bipolar band-power features, its shrinkage LDA code and its time-ordered train/test split.
usesACPC kernel alignment did not yield above-chance SEEG key-press transfer or improve 10-trial calibration in 8 subjects (mean AUC 0.517)
Uses the archived ds004473 band-power features (same SHA-256) and the numpy shrinkage LDA code.
usesSEEG beta power did not exceed pre-cue rest after cued key presses in eight subjects
We use its extraction code (density_decode.py, lda_np.py), trial definition, montage and filter without change. The re-extracted REST and MOVE features are equal to its archived features.
usesSEEG key-press decoding in eight subjects used channel patterns; removing the channel mean preserved AUC within 0.02
We used the archived ds004473 bipolar band-power features, the shrinkage LDA code and the trials 1-100 / later-trials protocol from this publication. FULL HG AUC for sub-1 is identical (0.9474).
usesA per-shaft stacked LDA was non-inferior to pooled LDA for cued key-press detection in 8 SEEG subjects
Input features (ds004473 v1.0.2, 8 subjects, HG and beta bipolar log power) copied unchanged; checksums match.
usesIn eight SEEG subjects, paired scores did not improve discrimination and baseline subtraction reduced AUC
We used the archived per-trial rest and move features (data/sub-*_features.npz, byte-identical, SHA-256 checked) and the numpy shrinkage LDA (code/lda_np.py) of this publication. Our HG+B transfer AUCs are equal to its values.
usesIn 8 SEEG subjects, calibration CV did not significantly overstate held-out ITR; oracle thresholds and window-only timing raised estimates
We used the archived per-trial HG and beta features (data/sub-*_features.npz, SHA-256 values in results/sha256_input_features.txt) and code/lda_np.py unchanged. Its trials 1-100 / 101+ split defines the calibration and test sets.
usesSlow key presses do not lower held-out high-gamma press detection in 8 SEEG subjects; reaction time correlates weakly and positively with LDA move score (ds004473, response-locked windows)
Uses the archived ds004473 high-gamma rest/move window features and the trial rule from this publication.
usesCV-tuned shrinkage improved later-trial SEEG decoding with 100 calibration trials, but not with 30
Uses the archived, hash-checked band-power features and the LW LDA code of this publication without change.
usesUnlabeled task covariance lowers few-trial SEEG movement-detection AUC, while rest-only covariance does not
Archived HG and beta band-power features of all trials of ds004473 and the numpy LDA code.
usesHigh-gamma-decrease channels decoded key press and improved 16-channel decoding in eight SEEG subjects
We used its archived HG and beta features (SHA-256 agrees with its manifest), its LDA code and its train 1-100 / test 101+ split.
usesIn 8 SEEG subjects, unsupervised threshold adaptation did not improve balanced accuracy over a fixed calibration threshold
We used the archived per-trial HG and beta features (data/sub-*_features.npz, SHA-256 values checked against the source) and lda_np.py (shrinkage LDA) unchanged. Its result (no AUC decrease on trials 101+) motivated the test of the threshold.
usesIn 15 OpenNeuro ds006253 SEEG patients, a same-trial baseline did not improve 70–150 Hz movement-onset detection; subtraction lowered AUC in all 15
Shrinkage LDA and AUC code (lda_np.py).
usesPer-shaft SEEG decoder scores were non-inferior to a joint decoder within 0.02 AUC in an eight-subject offline test
We use the archived ds004473 HG and beta features (hashes checked) and the unchanged lda_np.py of this package.
usesIn eight SEEG subjects, multitaper power did not improve 0.1 s key-press decoding over Hann; Hann performed better at 0.5 s
We copied the helper code from its package and use its trial definition and train 1-100 / test 101+ split.
usesLog power did not significantly improve shrinkage-LDA AUC over linear power in eight SEEG subjects
We used the archived ds004473 HG and beta band-power features and the numpy shrinkage LDA from this package without change.

More citing publications