In eight SEEG subjects, 20% missed cues kept median AUC loss below 0.02; cleaning added 0.011 AUC at 30% missed cues

We measured how much calibration-label noise a shrinkage-LDA SEEG movement detector tolerates. We used the archived HG and beta log-power features of publication 04683a2a (OpenNeuro ds004473 1.0.2, 8 subjects, cued key press vs rest). We trained on trials 1-100 with simulated noise and tested on trials 101 and higher with the cue labels. MISS noise: a missed cue, where the move window is a rest window but keeps the label move. FLIP noise: symmetric label flips. There were 40 seeds for each subject, model and noise fraction. The plan was hashed before the run. Results (subject is the unit, exact sign-flip tests): with MISS noise, the median AUC drop was 0.008 at 10% and 0.016 at 20%, and 0.040 at 50% (8/8 subjects). MISS noise mostly increased the false-positive rate at the trained threshold (median 0.169 to 0.270). FLIP noise was much more harmful: the median drop was 0.031 at 10% and 0.139 at 30%. A cleaning step (remove windows that a 5-fold CV decoder classifies against their label, then refit) increased the AUC by +0.011 at 30% MISS (primary endpoint, 6/8 subjects, one-sided p = 0.039). With FLIP noise, it increased the AUC by +0.018 to +0.058 at 10-40% (p <= 0.031). On clean data, it changed the AUC by -0.002. Limits: 8 subjects, offline analysis, simulated noise, one linear decoder.

brain-computer interface · SEEG · label noise · shrinkage LDA · movement decoding · signal classification · cross-validation cleaning