Simulated calibration label errors reduce SEEG detection AUC more when labels are exchanged than when presses are missed

In an offline analysis of archived band-power features from 8 subjects in one SEEG key-press task, simulated exchange of rest and move labels in 20% of calibration trials reduced median test AUC by 0.074 (8/8 subjects; exact one-sided sign-flip p = 0.004) from a clean-label median AUC of 0.869. The loss increased with the exchange rate, reaching 0.248 at 40%. Simulated missed presses, represented by replacing move examples with same-trial rest windows, reduced median AUC by 0.017 at 20%. A training-only self-cleaning step that removed 15% of examples with the lowest cross-validated signed margins recovered 0.019–0.032 AUC under exchanged labels, less than half the 20% exchange loss; on clean labels, its median change was -0.004 (p = 0.63). Noise was simulated, tests beyond the primary comparison had no multiple-comparison correction, and the analysis covered 8 subjects, one session, and one task. These offline results may not represent real label errors such as late presses.

stereo-EEG · brain-computer interface · label noise · calibration · linear discriminant analysis · AUC · signal processing