Random and time-ordered cross-validation gave similar key-press AUC in eight SEEG subjects

In eight subjects from OpenNeuro ds004473, we compared random cross-validation with time-ordered evaluation of an offline SEEG key-press-versus-rest detector. Each subject contributed 60 trials sampled across one 23–57 minute session. With 70–150 Hz features, the median AUC was 0.701 for random five-fold cross-validation and 0.706 for five contiguous time blocks; the mean random-minus-blocked difference was −0.026 (95% CI −0.069 to +0.016; random higher in 2/8 subjects; sign-flip p = 0.18). With 70–150 Hz plus 13–30 Hz features, medians were 0.785 and 0.771, and the mean difference was −0.002 (95% CI −0.016 to +0.012; random higher in 3/8; p = 0.73). Training on the first 30 trials and testing on the last 30 was not worse than random half splits: forward AUC was higher in 7/8 subjects with 70–150 Hz features (p = 0.078). Rest-window high-gamma drift was small (median early-versus-late AUC 0.57; range 0.42–0.64). These results indicate that, in this within-session offline setting with trials spaced 23–48 s apart, random-fold AUC was a fair estimate of forward AUC. The study used one session per subject, one random seed, and eight subjects; it did not test between-session drift. The press window also included visual and cue responses, so the result concerns the evaluation scheme, not a motor-specific signal. The archived feature files came from ds004473, but the exact dataset release used is not recorded.

SEEG · cross-validation · time-ordered evaluation · AUC · brain-computer interface · key press detection · signal drift · shrinkage LDA · OpenNeuro ds004473