Who is WithMe? EEG features for attention in a visual task, with auditory and rhythmic support

Abstract

Introduction: The study of attention has been pivotal in advancing our comprehension of cognition. The goal of this study is to investigate which EEG data representations or features are most closely linked to attention, and to what extent they can handle the cross-subject variability.

Methods: We explore the features obtained from the univariate time series from a single EEG channel, such as time domain features and recurrence plots, as well as representations obtained directly from the multivariate time series, such as global field power or functional brain networks. To address the cross-subject variability in EEG data, we also investigate persistent homology features that are robust to different types of noise. The performance of the different EEG representations is evaluated with the Support Vector Machine (SVM) accuracy on the WithMe data derived from a modified digit span experiment, and is benchmarked against baseline EEG-specific models, including a deep learning architecture known for effectively learning task-specific features.

Results: The raw EEG time series outperform each of the considered data representations, but can fall short in comparison with the black-box deep learning approach that learns the best features.

Discussion: The findings are limited to the WithMe experimental paradigm, highlighting the need for further studies on diverse tasks to provide a more comprehensive understanding of their utility in the analysis of EEG data.