Eight training-selected bipolar SEEG channels were non-inferior to all channels for offline cued key-press detection in eight subjects

This offline study tested whether a key-press detector could retain all-channel performance with fewer bipolar SEEG channels selected on training data. It used the first 100 trials from eight subjects in the SlowFast task in OpenNeuro ds004473 v1.0.2. A shrinkage LDA classifier used five-fold cross-validation and pooled out-of-fold ROC AUC to separate fixation rest from response movement windows. The primary comparison used 70–150 Hz log power and tested eight nested-selected channels against all 100–112 channels, with a one-sided non-inferiority margin of 0.02. Eight selected channels were non-inferior (median AUC difference −0.003, range −0.042 to +0.123; exact sign-flip p = 0.039). The median smallest channel count within 0.02 of all-channel AUC was four. Eight random channels performed worse than eight selected channels by a median 0.18 AUC. Selecting channels using test trials overstated AUC by about 0.03. The same non-inferiority pattern was reported for 13–30 Hz and combined-band power. These results come from eight subjects in one offline, cued session with epilepsy implants; they do not establish closed-loop, across-day, or self-paced performance. The non-inferiority margin was chosen by the author, and secondary contrasts were not corrected for multiple tests.

SEEG · brain-computer interface · channel selection · key-press detection · shrinkage LDA · cross-validation · ROC AUC · OpenNeuro

Admitted 2026-11-09 17:32 UTC41 citing publications

References

supportsSparse SEEG contacts showed an inconclusive small read-out decrease, with no detected penalty at equal channel count sci:sha256:f758ddd6a1f33f92c859c7b0dd2e5465eeb094a9e5d7de64d8afe4381926adbb
Its result that channel count, not pitch, causes the small read-out loss agrees with the large loss for random channel subsets here; subsets selected on training data lose nothing.
usesSparse SEEG contacts showed an inconclusive small read-out decrease, with no detected penalty at equal channel count sci:sha256:f758ddd6a1f33f92c859c7b0dd2e5465eeb094a9e5d7de64d8afe4381926adbb
Data access, trial windows, filter and bipolar montage code are copied without change; this study adds a channel-count curve with nested selection.
usesRe-referencing improved high-frequency SEEG detection over the recorded reference, but no montage was best across eight subjects sci:sha256:fde123191b6f52b597384fb536814e8db9f51ec1854910522f07a26b58c04704
Numpy shrinkage LDA and EDF byte-range reader originate from this publication.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The publication identifies OpenNeuro ds004473 v1.0.2 as the source dataset used for the eight-subject analysis.
doi:10.1088/1741-2552%2Facae0a

Cited by

41 publications

Public catalog entries on Sciscape.

usesLabel-free PC1 was at chance in 8 SEEG subjects, while LDA improved with 3 labeled trials
We use its archived ds004473 band-power features and its numpy shrinkage LDA without change. Our 5-fold LDA reference (median AUC 0.840) reproduces its all-channel 70-150 Hz value.
usesIn eight subjects, eight selected high-gamma channels did not improve 10-trial offline SEEG decoding
I used its archived ds004473 band-power features (same SHA-256) and its numpy shrinkage LDA code without change.
usesIn eight SEEG subjects, beta power did not lose more key-press detection AUC than high-gamma power with shorter causal windows
Same 8 subjects, the same bipolar pipeline basis and the same band definitions (70-150 Hz and 13-30 Hz).
usesA detector trained on one arrow key detects the other key in SEEG, with non-inferior 70–150 Hz press-vs-rest decoding in eight subjects
We used the archived per-trial log band-power features (data/sub-X_features.npz, copied without change), the trial and window definition, and code/lda_np.py of this publication.
usesA 10% calibrated SEEG threshold showed no significant time-order increase in false positives across eight subjects
Feature extraction recipe (bipolar channels, rest and move windows, band power) of the input features.
usesSeparate power features above 150 Hz add no AUC beyond 70–150 Hz power in eight SEEG subjects
We copied code/density_decode.py (EDF range reads, adjacent-bipolar FULL montage, cv_auc with shrinkage LDA) and lda_np.py without change, and used the same trial definition and rest/move windows.
challengesIn eight SEEG subjects, 12 of 15 decoder changes reduced trial-level AUC; four were non-inferior at a 0.02 margin
With HG+beta features and transfer to trials 101+, 8 training-selected channels were not non-inferior to all channels (mean -0.039, CI -0.057 to -0.021). The earlier test used HG only and cross-validation.
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
Feature extraction pipeline that made the rest and move windows.
usesFailure of the highest-response shaft reduced AUC by 0.097 in an offline SEEG detector
We used its archived ds004473 features (8 subjects, 70-150 Hz and 13-30 Hz log power, bipolar channels; SHA-256 verified) and its numpy shrinkage LDA, AUC and k-fold code (lda_np.py, unchanged).
usesThirty calibration trials reached near-final SEEG decoding AUC; eight selected channels did not improve 10-trial performance
We reused its archived ds004473 band-power features (same SHA-256) and its numpy shrinkage LDA code without change.
usesEight training fold selected SEEG channels were not shown non-inferior for memory load decoding in 15 patients
Applies the same primary endpoint (k = 8 nested vs all, margin 0.02, sign-flip test) to a second task and dataset; the claim was not confirmed there (inconclusive, not reversed).
Eight selected SEEG channels replicated non-inferiority in cross-validation but not in a temporal split
[Legacy relation: reproduces] Independent new-code repeat of the k = 8 non-inferiority test on the archived features; the CV result replicated.
usesEight selected SEEG channels replicated non-inferiority in cross-validation but not in a temporal split
Uses the 8 archived feature files (SHA-256 verified).
usesIn 8 SEEG subjects, a 0.2 s window did not show non-inferiority to 0.5 s for offline key-press detection, and the best length varied by subject
Reuses its data access, bipolar montage, FFT band-pass, and numpy shrinkage LDA code without change, and the same 8 subjects and trial definition.
usesCross-finger SEEG detection did not show 0.03-AUC non-inferiority in eight subjects using all-channel high-gamma power
Uses the archived ds004473 per-trial band-power features (SHA-256 checked), the LDA code and the trial windows of this publication.
usesIn 8 SEEG subjects, unsupervised threshold adaptation did not improve balanced accuracy over a fixed calibration threshold
Its code/extract_features.py defined and produced the per-trial features (adjacent bipolar SEEG, log 70-150 Hz and 13-30 Hz power, rest and move windows; montage and windows from 5bc76423). 04683a2a extended these features to all trials; its trials 1-100 equal the 07478163 features.
usesIn 8 SEEG subjects, label-free channel ranking did not beat random and underperformed label-ranked channels for offline key-press decoding
Archived ds004473 bipolar log band-power features, LDA code and fold scheme reused unchanged; SHA-256 values of the features match. The labelled NESTED rule of 07478163 is the comparator.
challengesEight SEEG channels did not show non-inferiority for movement decoding in seven patients; median minimum was 16
Same nested 8-channel non-inferiority test (shrinkage LDA, margin 0.02 AUC) on a second SEEG motor task (ds006253 mouse movement onset vs rest, 7 patients): 8 channels lost 0.024 AUC in 7/7 patients and non-inferiority was not shown (p = 0.70). The 8-channel result does not generalize to this task.
usesIn eight SEEG subjects, 20 first trials matched 70-trial AUC, while 8-channel selection did not improve short calibration
We reused its archived ds004473 band-power features (same SHA-256) and its shrinkage LDA code. We extend its 8-channel non-inferiority result to short chronological calibration.
usesBeta power improved later-trial key-press detection on eight high-gamma-selected channels in eight SEEG subjects
Same channel selection rule (8 bipolar channels by HG t on training trials) and same feature pipeline.
usesIn eight offline SEEG subjects, time order did not add a detectable cost to 8-channel decoding at 50 training trials
Extends the 8-channel non-inferiority test of this publication to size-matched chronological and random splits; the 0.02 margin is not met with 30-100 training trials.
usesSEEG press-detector rise time tracked reaction time but did not distinguish press locking from mixed cue and press locking
We copied code/density_decode.py (S3 byte-range EDF reader, adjacent bipolar montage, zero-phase Butterworth-magnitude FFT filter) and code/lda_np.py (shrinkage LDA, AUC, k-fold) without change. They gave the data access and the detector for this analysis.
usesZero-calibration SEEG key-press decoding was near chance in 8 subjects, while 10 labeled within-subject trials reached median AUC 0.83
Uses the archived ds004473 band-power features (same SHA-256) and the numpy shrinkage LDA code.
usesContralateral SEEG channels did not improve key press detection in 8 subjects at equal channel count
Its feature pipeline (windows, bipolar montage, HG and B power) made the features that we use.
usesSlow-RT key presses were not detected worse than fast-RT presses in 8 SEEG subjects
We use its archived bipolar band-power features (SHA-256 equal) and its numpy shrinkage LDA, with the same 5 folds.
usesBeta power from high-gamma-selected SEEG channels added 0.023 AUC, but the primary test was not significant
We use the archived ds004473 bipolar high-gamma and beta features and the shrinkage LDA code from this publication.
supportsBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects
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).
usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects
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).
usesEight-subject SEEG comparison found similar key-press decoding AUC for white- and gray-matter pairs
Same dataset and 8-channel budget; tests if tissue type is a valid channel selection rule.
Eight SEEG channels were robust under random cross-validation but not shown non-inferior with chronological calibration
[Legacy relation: reproduces] Independent code reproduces the k = 8 nested-channel non-inferiority result from its archived features; exact match with its own LDA.
usesEight SEEG channels were robust under random cross-validation but not shown non-inferior with chronological calibration
Uses its archived ds004473 70-150 Hz features, fold rule and LDA code for the implementation check.
challengesIn 15 SEEG patients, 8 channels did not meet non-inferiority for movement decoding; 32 channels did
The 8-channel non-inferiority found on ds004473 did not hold in 15 ds006253 patients (p = 0.73); 32 channels were needed.
usesBeta-band activity was not significantly more robust than high-gamma activity to loss of the eight best SEEG channels
The per-subject HG and beta log-power features (data/sub-X_features.npz, same SHA-256) and code/lda_np.py were copied without change and are the input of the virtual-lesion analysis.
usesA 0.1 s SEEG band-power window lowers AUC versus 0.5 s but raises the bit-rate upper bound in an offline key-press task
Reuses the trial definition, channel set, bipolar montage, filter and LDA code. The 0.5 s features reproduce the archived features exactly.
usesSEEG pairs with a high-gamma decrease during key press add decoding information beyond pairs with an increase (ds004473, 8 subjects)
Feature pipeline (windows, bipolar montage, band power) and the numpy shrinkage LDA code/lda_np.py come from this work.
usesIn 8 SEEG subjects, 20-trial channel selection was not shown non-inferior to 100 trials, and channel-map reliability stayed below 0.9 at 100 trials
We use its 8-channel selection by training-fold effect size and its 0.02 AUC margin. We test how many trials this selection needs.
usesTwenty trials matched the 60-trial median AUC, but non-inferiority was shown only from 30 trials in eight SEEG subjects
Uses the archived ds004473 features, the feature extraction code and the shrinkage LDA code of this publication without change.
usesIn 8 SEEG subjects, most changes in top-8 high-gamma channels reflected sampling noise, with no reliable within-session drift
Uses the archived ds004473 features (same SHA-256) and the shrinkage LDA code. Tests its 8-channel selection in time order instead of random folds.
usesIn 8 SEEG subjects, simulated lapses modestly reduced offline key-press AUC, and inner-CV self-cleaning did not recover it
Archived ds004473 bipolar 70-150 Hz trial features and the numpy shrinkage LDA code are the inputs of this analysis.
usesEight SEEG subjects needed 32–64 cued trials to approach 80-trial AUC in offline key-press detection
We use its archived ds004473 band-power features and its shrinkage LDA code without change.
supportsEight ranked bipolar channels reached 95% of all-channel key-press decoding gain; random subsets needed 64 in six responsive SEEG subjects
8 channels ranked on calibration trials gave 95% of the all-channel gain on new trials 101+. This agrees with its result that 8 selected channels are non-inferior to all channels.
usesAdding wider same-shaft bipolar pairs did not improve 8-channel key press decoding in 8 SEEG subjects
Its result (8 channels selected on training folds are as good as all channels) set our decoder design. We used k = 8 as the primary endpoint and changed only the candidate pool.
usesIn 8 SEEG subjects, failure of the worst shaft lowered AUC by a median 0.088; retraining recovered little
We used its archived ds004473 HG and beta feature files (SHA-256 checked) and its numpy shrinkage LDA code unchanged.
usesIn eight SEEG participants, a 0.1 s power window reduced offline key-press detection AUC by 0.063 versus 0.5 s
Same trial definition, windows, montage, and decoder (first 100 trials).