In 8 offline SEEG subjects, dropout augmentation lowered AUC; conditional-mean imputation was within about 0.001 AUC of refits at 25% and 50% simulated channel loss

We tested if dropout-augmented training keeps the AUC of an SEEG decoder after electrode loss, without a refit. Data: archived features of OpenNeuro ds004473 (8 subjects, key press vs rest, 70-150 Hz and 13-30 Hz log power of bipolar pairs). Train on trials 1-100, test on trials 101+. We simulated random loss of 0-75 % of bipolar pairs (50 draws) and loss of one shaft. The plan was pre-registered. The primary hypothesis was not supported. At 25 % loss, dropout-augmented LDA (p = 0.25) had a lower AUC than standard LDA with mean fill: mean -0.027, 7 of 8 subjects lower, one-sided sign-flip p = 0.99. It also had a lower AUC with no loss (-0.024), at 50 % loss (-0.034), and after shaft loss (-0.025). Two label-free methods increased AUC in 8 of 8 subjects at 25 % loss: an analytic dropout covariance (+0.010) and conditional-mean imputation of the lost features from the surviving features (+0.011). Conditional-mean imputation was within about 0.001 AUC of an oracle refit on the surviving channels at 25 % and 50 % loss (exploratory). Exploratory: augmentation decreased the automatic Ledoit-Wolf shrinkage by about 10 times. This is a possible cause. Limits: offline, simulated loss with known lost channels, one linear decoder, one task, 8 subjects. The package has the code, features, results, a figure, and reproduction steps.

SEEG · brain-computer interface · channel loss · shrinkage LDA · dropout augmentation · conditional-mean imputation · AUC