Decoding Brain’s Global Spatiotemporal Dynamics using a scalable k-means clustering method

Motivation: Technical advancements in fMRI data acquisition have continuously increased both the spatial and temporal resolution while our current approaches to dynamically analyzing the spatiotemporal patterns of hemodynamic activity have been hindered by limited computational power. Goal(s): Investigating whole brain global spatiotemporal dynamics in large fMRI datasets. Approach: We developed a scalable k-means method to work on large fMRI data. Results: Our results demonstrate the feasibility of our clustering method on terabyte (TB) size data resulting in six global spatiotemporal patterns. Those global patterns are either in the spatial or temporal domain, alluring to the fundamental characteristics of BOLD dynamics at resting state. Impact: We demonstrate the power of our large data clustering method on high-resolution fMRI data. Six global spatiotemporal patterns are obtained, showing that our method allows for the joint estimation of linked spatial or temporal patterns in one forward step.