In eight SEEG subjects, simulated lapses cost 0.016 AUC at 20%, while label swaps cost 0.077

We tested how simulated calibration-label errors affected a shrinkage linear discriminant analysis (LDA) detector of cued key press versus rest in eight subjects with stereo-EEG (SEEG). We used archived per-trial features from OpenNeuro ds004473: bipolar log power in 70–150 Hz and 13–30 Hz bands. The classifier was trained on trials 1–100 and evaluated on later trials with their original labels. Across 40 random draws per error rate, at 20% simulated lapse errors the mean test AUC loss was 0.016; seven of eight subjects had loss below the 0.02 margin (exact one-sided sign-flip p = 0.012). With 70–150 Hz features alone, mean loss was 0.013 (p = 0.020). At 20% label swaps, mean loss was 0.077, or 0.061 more than for lapses (all eight subjects; two-sided p = 0.008). The largest tested rate with mean loss below 0.02 was 20% for lapses and 5% for swaps. Removing the 20% of training trials with the lowest cross-validated score margin did not yield a reliable gain at 20% lapses (+0.002 AUC, p = 0.34) or a reliable cost on clean labels (−0.003, p = 0.46). These results are from simulated noise on real features in an offline analysis of one session per subject, with fixed windows, one decoder, 100 training trials, and a coarse noise-rate grid. The 0.02 AUC margin was adopted from earlier work and is not a clinical standard.

stereo-EEG · brain-computer interface · label noise · shrinkage LDA · AUC · motor decoding · OpenNeuro