Relating high-dimensional structural networks to resting functional connectivity with sparse canonical correlation analysis for neuroimaging

Abstract

Human brain mapping is increasingly faced with the need to efficiently interrogate small sample size, but high-dimensional (“short and wide”) data-sets. Such data may derive from rare or difficult to identify populations wherein we seek to detect subtle network changes that precede disease. Few prior hypotheses may exist in these cases and yet, due to small sample size, exploratory analysis is power challenged. We overview how to use sparse canonical correlation analysis to produce biologically principled low-dimensional representations before proceeding to hypothesis testing. This strategy conserves power by taking advantage of the underlying neurobiological covariation across modalities to compress large data-sets. We provide an example that maps voxel-wise cortical thickness measurements to resting state network correlations in order to identify structure-function sub-networks with little further supervision. The resulting network-like, sparse basis functions allow one to predict traditional univariate outcomes from multiple neuroimaging modalities even when sample sizes are relatively small. Importantly, these data-driven functions are anatomically and edge-wise specific, allowing a nearly traditional neuroscientific interpretation.