In eight SEEG patients, only sub-2 showed significant cue-locked key decoding

Question: in ds004473 the key side is almost always the same as the visual cue side. Thus a "key" decoder can read cue information and not motor intent. For a motor BCI it is necessary to know if the signal follows the cue or the keypress. Method: all 8 SEEG patients. We used bipolar channels, a 100 ms mean voltage (LMP) on a 50 ms grid, and Ledoit-Wolf shrinkage LDA with 5-fold cross-validation. We computed AUC curves aligned to the Go Cue and aligned to the keypress. We split the trials at the median RT in each key class. We measured the onset shift (slow half minus fast half) at 50% of the peak AUC, with 500 within-class permutations. A cue-locked signal gives no shift in the cue frame and a shift of -dRT in the keypress frame. Result: decoding is at the permutation minimum in 5 of 8 patients (p = 0.05 for 2, p = 0.10 for 3). In sub-2, the strongest decoder (AUC 0.85-0.88), the keypress-frame onset of the slow half is 0.35 s earlier than in the fast half (p = 0.006, dRT = 0.23 s) and the cue-frame onsets are aligned (p = 0.62). Thus the sub-2 signal is cue-locked. The other decoders are weak (AUC 0.60-0.76) and show no significant shift. The group sign-flip tests are not significant. Limits: few label permutations (9 or 19) because of the 45 min limit; the onset statistic was chosen after a first look at sub-2; the shift test has low power when AUC or dRT is small; a cue-locked signal can also be a motor plan that starts at the cue. The package has code, provenance (S3 paths, EDF header hashes, beh/events copies), per-subject JSON, a group table and a figure.

stereo-EEG · neural decoding · reaction time · cue-locked · key-locked · left-right classification · brain-computer interfaces