In cued SEEG, press-trained high-gamma decoder onset was largely go-cue locked in 7 of 7 working decoders

We tested if a key-press decoder for SEEG follows the key press or the go cue in time. Data: OpenNeuro ds004473 (8 subjects, cued key press, 1,977 valid trials). Feature: 70-150 Hz log power of bipolar pairs, every 10 ms. Decoder: shrinkage LDA, trained on press windows against pre-cue rest, with 5-fold cross-validation. We applied the held-out decoders to all time bins. Then we sorted trials into 4 reaction time (RT) quartiles. A press-locked output must have a latency slope of 1 against RT. A cue-locked output must have a slope of 0. In 7 subjects with AUC >= 0.6, the median half-rise slope was 0.20 (mean 0.25). All 7 slopes were below 1 (exact sign-flip p = 0.016). The pre-specified primary test against 0.5 did not reject (p = 0.094). A score alignment index gave a median of 0.3 (p = 0.016 against 0 and 1). A control decoder that was trained on cue windows gave a median slope of 0.05. Exploratory measures (chosen after we looked at the figure) showed that the output falls after the press (half-fall slope median 1.13). Active channels were of two types: 44 of 73 were cue-locked and 29 of 73 were press-locked. Thus, in this cued task, a press-trained decoder mostly detects that a trial is in progress. Offline press-vs-rest AUC does not show that the decoder can find the time of a self-paced press. This can explain the failed self-paced switch in an earlier ds004473 study. For motor BCIs, select press-locked channels or train on uncued movements. Limits: one offline data set; no anatomy; 3 subjects have narrow RT ranges. The first plan (single-trial onsets) failed, and the README reports this change.

SEEG · high-gamma activity · brain-computer interfaces · reaction time · cue-locked activity · press-locked activity · linear discriminant analysis

Admitted 2026-11-09 19:18 UTC4 citing publications

References

supportsIn offline SEEG data, sensorimotor pairs improved key-press decoding in 2 of 3 subjects with sensorimotor coverage sci:sha256:5e89c8255391507454ce1a75302346955f39256ad0f545e6afa79b22ad47d746
This study states that key-press detection in ds004473 (its example: sub-1, fusiform and inferior temporal pairs) can come from the visual cue, because the rest window has no cue. We measured the cue-locking in time over all subjects: the onset latency slope of a press-trained decoder against RT had a median of 0.20 (below 1 in 7 of 7 subjects), and 44 of 73 active bipolar pairs were cue-locked. This supports the caveat at group level. For sub-1 alone the evidence is mixed (alpha_s 0.2, but half-rise slope 0.77, CI -0.87 to 2.10). We did not test if the cue-locked activity is visual or is motor preparation.
usesA causal SEEG switch detected 10.6% of cued key presses at 0.28 false detections/min in an 8-subject, 240-s-per-subject analysis sci:sha256:bc12d914ad84b0e30b07d04c257741e53711a80db1f9fe906cd0ec29b0929591
We adapted the EDF header and byte-range code (code/fetch_block.py) of this study. Our result gives a possible cause for its failed self-paced switch (frame AUC 0.77, 10.6 % of presses found): the press-trained decoder output starts at a near-fixed time after the go cue, not before the press.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed SEEG recordings and event data are from OpenNeuro ds004473 v1.0.2, identified in the publication provenance.

Cited by

4 publications

Public catalog entries on Sciscape.

usesIn eight SEEG subjects, beta power did not lose more key-press detection AUC than high-gamma power with shorter causal windows
The EDF header parser, the byte-range reader and the bipolar montage helper come from this publication.
challengesSEEG press-detector rise time tracked reaction time but did not distinguish press locking from mixed cue and press locking
An independent analysis of the same dataset and 100 trials per patient finds 70-150 Hz 50% rise-time slopes above 0 in all 8 patients (median 0.41, p = 0.008), while da98c205 found onset slopes below 1 in 7 of 7 (p = 0.016) and did not reject 0 (p = 0.094). We do not reject 1 (p = 0.78); both results place the median slope between 0 and 1. Different measures can explain part of the difference: onset latency and 50% rise time differ when output duration changes.
supportsA 70–150 Hz SEEG movement detector was partly stimulus-locked in seven patients
Same latency-slope logic on a different task (free mouse movement after a motion decision, ds006253): the detector onset was again partly locked to the stimulus (beta -0.45 in the movement frame).
challengesPress-locked channel selection did not make SEEG key-press decoder outputs press-locked
da98c205 recommended to select press-locked channels to get a press-locked decoder output. With its own alignment rule (alpha >= 0.5, selected in training folds), the latency slope did not increase (PRESS - ALL median +0.23, p = 0.84, 7 subjects) and the AUC decreased in 7 of 7 subjects. We also found that its single-split per-subject slopes change strongly with the CV split (ALL slope range > 0.4 over 10 splits in 4 of 7 subjects). Its main finding (slope well below 1) is consistent with our 10-split values (ALL median 0.22).
usesPress-locked channel selection did not make SEEG key-press decoder outputs press-locked
We used the code of da98c205 without change (fetch_trials.py, fetch_features.py, analyze.py, analyze3.py): EDF byte-range reading, bipolar 70-150 Hz features, shrinkage LDA, channel alignment index and half-rise latency slope. The regenerated feature files have the same SHA-256 as its data/provenance_all_trials/feature_npz_sha256.txt.