A 1-s SEEG window matched the 3-s mean AUC, but did not meet the pre-registered non-inferiority test

We decoded working-memory load (set size 8 versus 4) from SEEG band power in 15 patients in session 1 of OpenNeuro ds004752. We changed only the window length, starting each window at maintenance onset. Mean AUC was 0.725 for 1 s and 0.720 for 3 s (difference +0.006; 95% CI -0.076 to +0.085), but the pre-registered non-inferiority test at margin 0.05 did not pass (sign-flip p = 0.107). The 0.25-s window had AUC 0.529, 0.19 below 3 s (Holm p = 0.0002). The 1-s window had the highest information rate (0.14 bits/s versus 0.04 bits/s for 3 s). A sliding 0.5-s window had the highest AUC, 0.68–0.69, from 0.5 to 1.5 s after maintenance onset, then fell to 0.52 at maintenance end. The result does not prove that 1 s is sufficient. Trial counts were 20–42 per subject, and per-subject variance was large. Set size also changed the encoding sequence, so the maintenance-onset signal may include a stimulus offset or after-effect and may not be a pure memory signal. Information rate used window time only; the event time was known. Data were read from the public S3 mirror; the exact dataset release was not established.

SEEG · working memory · memory load · band power · analysis window · AUC · non-inferiority · information rate · brain-computer interface · OpenNeuro ds004752