Online processing of electrical signals in the brain may have clinical utilities. Using machine learning, it has become possible to predict an action, desire, or event, from electroencephalogram (EEG) activity. For example, the Error-Related Negativity (ERN), an electrophysiological response to errors, has been associated with risk for anxiety disorders and studies have used ERN as an indirectly modifiable bio-marker of anxiety. Thus, it may be possible to reduce the ERN if feedback is provided to participants online as they perform tasks that engender errors. As part of a larger study, pattern matching algorithms formed models from EEG data in order to locate precisely when a participant made an error during an Eriksen Flanker-like task. Such a feat is accomplished by locating ERN online and using its position in time as an indicator of error occurrence after being trained using markers. Preliminary analyses indicate that this approach was able to predict responses to the task with an average accuracy of 91.9% when implementing the Dual Augmented Lagrangian (DAL) method (N = 112). Such promising results on pseudo-online data using sophisticated pattern matching prompted a pilot study to explore the viability of a simple, yet robust test of single-trial ERN detection for use in neurofeedback. This pilot study aimed at: 1) determining the reliability of an online calculated ERN to the corresponding offline ERN; 2) reducing ERN using ERN using rudimentary operations which compose machine learning algorithms. Undergraduates from San Diego State University (N=33; mean age=19) were attached to an EEG system and asked to perform a pre and post-assessment flanker-like task consisting of five horizontal arrows in either congruent or incongruent arrangement. Offline ERN was correlated to online ERN (r = 0.44) with only minimal pattern matching and digital signal processing implemented. An algorithm processed trial-by-trial ERN and composed a rolling average ERN which was compared to the participant’s baseline. If the rolling average online ERN differed significantly from the ERN produced by the participant during the pre-assessment, the algorithm reacted and made a prediction on the change of anxiety state of the participant. Feedback was not given, but was calculated for reliability. Future iterations of the study will present the feedback in attempts to modify ERN.
Alessandro D'Amico– Coinvestigator / Research Assisstant, Center for Understanding and Treating Anxiety; San Diego State University