A rest-anchored threshold reduced false-alarm error for HG+beta SEEG detection, but not for HG alone

We tested whether a decision threshold for an SEEG key-press detector keeps its target false-alarm rate (FPR) on later trials. We used archived high-gamma (HG, 70–150 Hz) and beta (13–30 Hz) bipolar features from 8 subjects in OpenNeuro ds004473 v1.0.2. A shrinkage LDA threshold was calibrated on out-of-fold rest scores from trials 1–100 and evaluated on trials 101 onward. At a 10% target, the mean FPR error for HG was −0.019 (exact sign-flip p=0.35), while subject FPRs ranged from 0.00 to 0.157 and mean absolute error was 0.048. At a 5% target, 2 of 8 subjects had more than twice the target FPR. In an exploratory control using random rather than first-100 calibration trials, mean absolute error was 0.041 for HG at a 10% target (p=0.45), consistent with finite calibration-sample noise accounting for much of the error. A running threshold based on the last 50 cued rest scores reduced mean absolute error for HG+beta to 0.022 at a 10% target and 0.018 at a 5% target (p=0.047 for both comparisons); the reduction was not significant for HG alone. TPR did not change significantly (all p>0.4). The study includes 8 subjects, one session each, and one cued task; its windows are trial-locked, and tests were not corrected for multiple comparisons.

SEEG · brain-computer interface · false-alarm rate · threshold calibration · shrinkage LDA · high-gamma · beta band