Slow-RT key presses were not detected worse than fast-RT presses in 8 SEEG subjects

We tested if a movement-detection BCI decoder detects slow key presses less well than fast key presses. We used the archived bipolar SEEG features of sci 07478163 (OpenNeuro ds004473 v1.0.2, 8 epilepsy patients, first 100 trials). We used Go Cue and Response times from events.tsv to get the reaction time (RT). A shrinkage LDA on 70-150 Hz log power of all bipolar channels gave out-of-fold scores (5-fold CV). In each subject, we split the trials at the median RT and compared the AUC of each half against the same set of rest windows. The slow half was not detected worse: median AUC 0.855 (slow) vs 0.827 (fast), mean difference +0.022 AUC (95% CI -0.004 to +0.056, exact sign-flip p = 0.16, 3 of 8 subjects lower). The 13-30 Hz and the two-band decoders also showed no difference. The 70-150 Hz move score had a small positive correlation with RT (mean Spearman rho 0.08, 6 of 8 subjects, p = 0.031, not corrected). Thus, in this RT range (median 0.39 to 1.06 s), slower responses inside one session do not decrease offline detection. The response-locked window can contain cue-evoked activity in fast trials. The package has the plan, one documented amendment, code, data hashes, results and a figure.

SEEG · brain-computer interface · movement detection · reaction time · high gamma · shrinkage LDA · cross-validation · AUC