In 8 SEEG subjects, simulated lapses modestly reduced offline key-press AUC, and inner-CV self-cleaning did not recover it

In an offline cued key-press versus rest detector, we simulated calibration lapses in archived 70–150 Hz bipolar log-power features from 8 SEEG subjects (100 trials each, 100–112 channels). A lapse was a move cue with no movement, so the move-labelled window contained rest activity. With shrinkage LDA, 5-fold trial-wise cross-validation, and 10 repeats, median AUC with correct labels was 0.847. Lapses in 10%, 20%, 30%, and 40% of training trials reduced mean AUC by 0.008, 0.015, 0.025, and 0.035; at 20–40%, all 8 subjects showed a loss (exact sign-flip p = 0.008). At 20% lapses (10% of training windows wrong), the loss was 0.015, compared with 0.026 for 10% symmetric label flips, or about 0.6 times the loss per wrong window. The pre-specified primary hypothesis was not supported: inner-cross-validation self-cleaning did not recover AUC at 20% lapses (clean minus plain = −0.003, 95% subject-bootstrap CI −0.012 to 0.004, 4/8 subjects better, p = 0.56), and it did not help at other lapse rates (Holm p = 1.0). The safety hypothesis was also not shown: with correct labels, self-cleaning removed 27% of training windows and changed AUC by −0.006 (95% CI −0.016 to 0.001), whose interval crosses the −0.01 non-inferiority margin. The cleaning rule identified lapse windows only about 2.5 times more often than chance (precision 0.25 at a 0.10 base rate); an oracle removing only wrong windows recovered +0.011 AUC at 20% lapses. Exploratory, uncorrected results showed +0.018 to +0.032 AUC recovery for symmetric flips at 20–40% (6–7 of 8 subjects), 16–35% of the oracle gain. These offline simulations use one dataset, one session per subject, one feature type, and one linear decoder; real lapses may include partial or late movement. For this strong motor high-gamma detector, the results do not support adding misclassification-based self-cleaning to protect against no-movement lapses.

SEEG · brain-computer interface · motor decoding · label noise · calibration lapse · shrinkage LDA · AUC · cross-validation · self-cleaning · OpenNeuro