In 8 SEEG patients, a label-free GMM estimate did not outperform calibration CV under simulated channel loss

A brain-computer interface decoder can lose channels after calibration, while labels for new trials may be unavailable. We tested whether a label-free estimate of held-out area under the ROC curve (AUC) could detect this loss better than calibration cross-validation (CV) AUC. We used archived high-gamma (70–150 Hz) and beta (13–30 Hz) bipolar band-power features from 8 patients performing a key-press versus fixation task in OpenNeuro ds004473. A z-scored Ledoit-Wolf shrinkage linear discriminant analysis decoder was trained on trials 1–100 and evaluated on trials 101 onward. At test, we simulated loss of 0%, 25%, 50%, or 75% of bipolar pairs by filling both bands with their training means, with 30 random draws for each nonzero loss. The label-free estimate fit a two-component Gaussian mixture with weights fixed at the design prior of 0.5 to unlabeled test scores and converted the fitted separation to an implied AUC. At 50% loss, its mean absolute error was not smaller than calibration-CV error (0.086 versus 0.072; better in 3 of 8 patients; one-sided exact sign-flip p = 0.79). From 0% to 50% loss, true mean AUC fell by 0.051, while the mean GMM estimate fell by 0.008 (p = 0.39). Across conditions, within-patient Spearman rho between GMM estimate and true AUC averaged 0.43 (one-sided p = 0.027; range −0.17 to 0.91). Exploratory results showed rho of 0.78–0.91 in 4 patients whose true AUC was at least 0.88 without loss, and −0.17 to 0.30 in the 4 weaker patients; this split was identified after seeing the data and needs retesting. The result is limited to one dataset and task, 8 patients, one decoder, one feature set, simulated mean-filled channel loss, a known class prior, and one train/test split. The method cannot detect a decoder sign flip.

brain-computer interface · AUC estimation · Gaussian mixture model · channel loss · intracranial EEG · cross-validation