In 15 SEEG patients, 10% training-label flips reduced median held-out AUC by 0.047; the equal-sample rule was not confirmed

This cross-task replication tested whether the held-out AUC loss from symmetric training-label flips follows the (1−2p)² equal-sample rule in a verbal Sternberg working-memory load decoder. It used archived maintenance-period band-power features from 15 patients with SEEG, a shrinkage LDA decoder, and 20 repeats of stratified five-fold cross-validation; only training labels were changed. Clean median AUC was 0.663. With 10% of training labels flipped, median AUC decreased by 0.047 (95% CI 0.007–0.058; 12/15 patients; exact sign-flip p=0.004), about twice the 0.020 loss reported for an FR1 recall decoder. The decoder retained a median 81% of above-chance AUC among 10 patients with clean AUC above 0.55. Differences between flipped-label and equal-sample conditions were not statistically significant at p=0.10 or 0.20, but confidence intervals exceeded the prespecified ±0.02 AUC equivalence margin, so equivalence was not shown. Limits include simulated symmetric errors, one offline session per patient, 20–42 trials per patient, and five patients with little clear load signal. The recorded OpenNeuro URLs and hashes identify the source files; their exact immutable dataset release was not established.

SEEG · brain-computer interface · label noise · working memory · shrinkage LDA · AUC · cross-validation · calibration trials