In 18 FR1 SEEG patients, 10% simulated random label flips lowered median held-out AUC by 0.020; 20% exceeded the effective-sample rule

We measured how much a within-patient SEEG word-recall decoder loses when some of its training labels are wrong. Data: 18 patients from OpenNeuro ds004789 (FR1, bipolar SEEG), 4944 words, 1120 recalled. Features: log power at 8 frequencies (3-140 Hz) on all channels, 0-1.6 s after word onset. Decoder: L2 logistic regression, C = 7.2e-4, 5-fold cross-validation by list. We changed only the training labels; the test labels were always correct. Clean median AUC was 0.673. Random symmetric flips with probability p = 0.05, 0.10, 0.20, 0.30, 0.40 decreased median AUC by 0.009, 0.020, 0.046, 0.079, 0.114. The pre-registered primary test (p = 0.10) found a loss in 17/18 patients (median -0.020, 95% CI -0.027 to -0.013, exact Wilcoxon p = 4e-5). Non-inferiority with a 0.02 AUC margin was not shown. For p = 0.05 and 0.10, the loss was equal to the loss from training on a fraction (1-2p)^2 of words with correct labels (paired difference +0.001 and -0.003, CIs include 0). At p = 0.20 the flips were worse than this rule (-0.022, CI -0.040 to -0.008, p = 0.007). Missed-recall errors (recalled words labeled not recalled, q = 0.10, 0.25, 0.50) decreased AUC by 0.004, 0.013, 0.031. Per wrong label, this is approximately 1.3 times the loss of symmetric flips (q = 0.25 and 0.50; not tested statistically). Thus, for BCI calibration, a 10% label error rate has approximately the cost of discarding 36% of the training words. Limits: errors were simulated at random, one task, one linear decoder, offline, 4 seeds per condition.

training-label errors · label noise · SEEG · free recall · brain-computer interface · AUC · logistic regression · effective sample size