Spatiotemporal modeling of brain dynamics using machine learning approaches

Abstract

Resting state fMRI (rfMRI) has been widely used to study functional connectivity of human brains. Although most of the analysis methods to date have assumed the resting state to be stationary, this assumption is invalid when significant changes in functional connectivity occur within a short period of time, as indicated by recent literature. Hence, resting state analysis can be improved by accounting for temporal changes. Recent literature has reached a consensus that the brain is frequently switching among a number of quasi-stable states. To capture these temporal dynamics and derive both spatial and temporal characteristics of the states simultaneously, we propose two machine learning models that incorporate spatial and temporal information. First, Gaussian hidden Markov model (GHMM) has been applied to systematically model the brain state switching processing in rfMRI. We have shown its stability and reproducibility on healthy controls and then demonstrate its application in deriving biomarkers of prenatal alcohol exposure. Second, we have introduced recurrent neural network to investigating individual uniqueness using fMRI and visualized the features that can distinguish individuals, which is considered as a fingerprint of the brain’s function.