In 8 SEEG subjects, 20% lapse noise raised false positives and slightly reduced AUC; symmetric flips caused larger losses

We simulated calibration-label noise in an offline SEEG move-versus-rest decoder for one cued key-press task. The analysis used HG and beta power features from 8 subjects in OpenNeuro ds004473 v1.0.2, a shrinkage LDA decoder, 100 calibration trials, and clean later test trials. At 20% lapse noise, mean test AUC decreased by 0.016 in all 8 subjects (maximum decrease 0.022); non-inferiority at a 0.02 margin was not shown (exact sign-flip p = 0.055). False-positive rate at threshold 0 increased by 0.035 in 7/8 subjects (p = 0.008), while hits did not decrease. Symmetric 20% label flips caused a larger mean AUC decrease of 0.082 and FPR increase of 0.080, each in all 8 subjects. An out-of-fold posterior guard recovered 0.035 AUC for flips and had no detected cost on clean data, but did not significantly help for lapses. A rest-quantile threshold kept FPR below 0.10 under lapses but reduced TPR by 0.134. This is a simulation in one dataset, task, decoder type, and calibration split; it is offline and does not test closed-loop use. The analysis used 15 repeats per condition rather than the 50 planned repeats.

brain-computer interface · SEEG · label noise · shrinkage LDA · calibration · false-positive rate · AUC · lapse · OpenNeuro