Direct comparison of EEG resting state and task functional connectivity patterns for predicting working memory performance using connectome-based predictive modeling

The use of machine learning techniques combined with advanced neuroimaging methods to reveal a complex relationship between patterns of neuronal activity and behavioral characteristics is an actively developing area in modern neuroscience. Investigating brain activity data non-invasively recorded in resting-state conditions has substantially enhanced our understanding of the neuronal foundations of cognitive functions. However, results of more recent studies using functional magnetic resonance imaging suggest that task-based paradigms may outperform resting state ones in predicting cognitive outcomes. To date, no studies have tested this hypothesis using EEG data. The aim of this study was to experimentally compare, for the first time, the predictive power of models built on high-density EEG data acquired during rest and an auditory working memory task execution, while utilizing different data processing pipelines to ensure the robustness and reliability of the findings. In accordance with previous studies, we found that modeling performance was slightly better for task EEG data compared to resting state recordings, except for high working memory load, where resting state data was superior to task-based functional connectivity in terms of predictive capacity. In general, models derived from both conditions showed overall high modeling accuracy, reaching up to R2 = 0.31. Functional connectivity in the alpha frequency band was the most efficient predictor of working memory scores, followed by that of theta and beta bands. Finally, we demonstrated that the choice of parcellation atlas and functional connectivity method significantly impacted the ultimate results, which should be taken into consideration when designing future experiments.