Bottom-up prediction of cognitive control features in EEG signals

Abstract

The prediction of failures in response inhibition is potentially very valuable for understanding pathologies as well as for the support of critical operations carried out by humans. To realize this aim, Fisher’s linear discriminant analysis with leave-one-out cross-validation has been performed per 1 Hz frequency interval on the 1000 milliseconds electroencephalogram Fourier-transformed periodograms preceding go-stimuli that were followed by a stop-signal in an auditory stop-signal task, using the power values per electrode as features. Ocular correction was applied, while no other artifacts were removed. This repeatedly class-balanced two-class classification resulted in significant accuracies for 21 of the 29 participants. It was hypothesized that both theta oscillations originating from the prefrontal cortex during anticipation as well as a balance between alpha oscillations over visual and those over auditory cortex both would promote successful stopping. The results translate to a partial validation of these theories. In future studies, the procedure can be extended and this intuitive multivariate pattern analysis machine learning method may continue to yield insights for research in artificial intelligence and human computer interaction.