Line length and high-gamma had similar cued key-press detection AUC in eight SEEG subjects

We compared low-cost time-domain features with 70–150 Hz high-gamma power for offline detection of cued key presses in stereo-EEG (SEEG) from eight subjects in OpenNeuro ds004473. Shrinkage LDA was trained on the first 60% of trials in time order and tested on the last 40%. With all bipolar channels, line length reached mean held-out AUC 0.771 and high-gamma power 0.777 (mean difference −0.006; Wilcoxon p = 1.0). With four channels selected using training trials, the respective AUCs were 0.780 and 0.807 (p = 0.38). Five time-domain features after decimation to about 250 Hz reached AUC 0.844 with all channels, 0.068 above high-gamma and better in six of eight subjects, but this difference was not significant (p = 0.078). Line length of the first difference performed worse than high-gamma in the four-channel read-out (AUC 0.651; seven of eight subjects worse; p = 0.016). Line length was not a close proxy for high-gamma: median across-trial correlations ranged from 0.14 to 0.62 by subject, with six of eight below 0.25. These offline results cover one task and session per subject, about 45 test trials per class, and cue-locked detection; the movement window also contains visual and cue responses. The tests had low power, no correction for multiple comparisons was made, and no difference does not establish equivalence. The study does not establish performance for self-paced movement or for device implementations with filtering, thresholds, or fixed-point arithmetic.

stereo-EEG · SEEG · brain-computer interface · movement detection · line length · high-gamma · time-domain features · signal processing · shrinkage LDA · OpenNeuro ds004473