In 15 SEEG patients, 20% simulated lapse labels cost 0.013 AUC and 20% swapped labels cost 0.060 AUC, replicating the asymmetry in ds006253

This pre-registered offline replication tested whether the training-label noise asymmetry previously reported for OpenNeuro ds004473 also occurs in ds006253. We used archived author 70–150 Hz high-gamma window features from 15 patients performing a random-dot decision task with mouse movement, and the same shrinkage LDA move-versus-rest decoder with a time-ordered 60/40 trial split. Noise was simulated in training labels only, at rates from 0 to 40%, with 40 random draws per rate; test labels remained correct. At 20% lapse noise, mean test AUC loss was 0.0125 (95% bootstrap CI 0.0097–0.0155; exact sign-flip p = 0.00034), below the 0.02 non-inferiority margin. At 20% swap noise, mean loss was 0.0602; the difference from lapse was 0.048 (95% CI 0.041–0.054), and swap cost more in all 15 patients (p = 3.1×10⁻⁵). The largest tested rates with mean loss below 0.02 were 20% for lapse and 5% for swap, matching ds004473. The result is limited to an offline simulation, one decoder, one REST definition, and these patients; simulated lapses may differ from real patient lapses.

stereo-EEG · SEEG · brain-computer interface · label noise · shrinkage LDA · high-gamma activity · OpenNeuro · replication