Analysis of Resting-State Functional Connectomes Based on the Graph Embedding Pipeline Rest2vec

Resting-state functional connectomics has risen as a novel way for assessing the interactions between brain regions. This discipline aims to study connectomes, complex maps that describe the set of functional connections within the brain, to understand how different areas are functionally correlated during rest. To this end, spontaneous fluctuations in neural activity can be used to determine the functional macro-scale topology of the human brain. Resting-state functional magnetic resonance imaging has been employed to analyze the endogenous neural activity. The quantification of the correlation between Blood Oxygenated Level Dependant (BOLD) timeseries of different brain regions may be employed to evaluate the relationships between different regions. Graph theory models have been applied to analyze the brain as a complex network, but several challenges still remain. Functional connectomes are high dimensional data mathematically described as graphs, in which the regions are considered as nodes and the correlations are referred to as edges. This characteristic may cause phenomena related to the "curse of dimensionality". Moreover, the meaning of negative correlations is still unclear, and they are often not considered in simpler models. Thus, the novel graph embedding method rest2vec has been developed to address these issues. Rest2vec uses a phase angle representation, the Phase Angle Spatial Embedding (PhASE) framework, to model the correlation between regions, considering both negative and positive edges, linked with a nonlinear dimensionality reduction method (Isomap) to map brain regions into a low-dimensional functional embedding, in which regions are defined according to their functional organization. This work aimed to apply rest2vec to analyze the functional connectivity information of a dataset of healthy subject, obtained by processing rs-fMRI data provided by the Human Connectome Project. The Glasser cortical atlas was used as brain parcellation, which allowed to evaluate the representation of biological-relevant patterns in the embedding space, comparing them to previously identified subnetworks. Moreover, a binary gradient-based partition was performed by using maximum mean discrepancy (MMD) as metric to assign the regions to one of module. This approach has outlined two subsets resembling the putative task-positive and task-negative network. Finally, the functional connectivity of different groups was compared using the PhASE framework, outlining statistically significant differences related to gender and age.