In eight offline SEEG subjects, time order did not add a detectable cost to 8-channel decoding at 50 training trials

In a single-session SEEG key-press-versus-rest task, we compared chronological training on trials 1–50 with 200 size-matched random splits in eight subjects. With 70–150 Hz power and shrinkage LDA, the mean 8-channel cost was similar for chronological and random splits (dAUC −0.0103 and −0.0105; Td = +0.0002, exact two-sided sign-flip p = 1.000). The all-channel AUC difference was −0.003 (p = 0.617). Neither split showed 8-channel non-inferiority at the 0.02 AUC margin (p = 0.125 and 0.133); one random split per subject met the test in only 13.5% of repeats. Across tested sizes of 30–100 trials, no time-order cost was significant for k = 4, 8, or 16. These results indicate that the earlier non-inferiority failure does not require a time-order effect, but eight subjects provide low power: a null result does not establish zero non-stationarity. Random splits interleave time and do not measure slow drift shared across both halves. The analysis is offline, from one session and task, and has no closed-loop test.

SEEG · non-inferiority · time-order effect · shrinkage LDA · AUC · random split · channel selection · OpenNeuro ds004473