Sparse representation of working memory processes based on fMRI data

Cognitive processes, such as working memory, are widely considered as dynamic, and they are believed to involve complex functional information flows in large-scale brain networks. However, traditional voxel-based fMRI time series analysis methods, which essentially assume that the hemodynamic responses of involved brain regions follow the block or event-related paradigms, are limited in recognizing and modeling this complex process that has been changing in both spatial and temporal spaces. In this paper, we propose a novel computational framework to explore the potential mechanisms underlying the working memory process. Specifically, we adopt the Transfer Entropy (TE) as the measure to model the directional functional interactions based on a series of structurally-consistent cortical landmarks. Then an effective and carefully-designed sparsity learning procedure was applied to derive the most representative interaction patterns for further analysis. Our results show a few cortical landmarks displaying significantly higher interaction strength during the whole fMRI scan and suggest that they might act as hubs in coordinating working memory process. Moreover, four prominent interaction patterns associated with the Default Mode Network (DMN) are found to be attenuated in the task performance period.