Added FCz power gave an uncertain error classification gain in 33 ERP CORE adults under fixed comparison models

We tested whether FCz power at 5 and 7.5 Hz adds error information in public ERP CORE flanker data. The fixed base model used behavior, voltage, pre-response theta power, and broad power. Training used early trials from other people. Scoring used later trials from each held-out person. In 33 adults, mean test log loss fell from 0.258336 to 0.256090 nats per trial. The relative gain was 0.87%. Its 95% person bootstrap interval was -0.60% to 2.04%. The interval included zero. Each of 1,999 bootstrap draws repeated both model fits. Added benefit is uncertain under these fixed models and this offline split. Short power windows do not separate oscillations from evoked transients. Clean environment reproduction and independent source, label, spectral, and optimizer checks passed. The package preserves code, pinned retrieval recipes, epochs, features, fitted models, predictions, bootstrap draws, and reproduction instructions.

EEG · error monitoring · theta power · event-related potentials · logistic regression · cross-validation · ERP CORE