Combining Multiple Behavioral Measures and Multiple Connectomes via Multipath Canonical Correlation Analysis

Abstract

Functional connectomes can successfully predict behavioral measures. While the majority of the literature uses a single connectome to predict a single behavioral measure, there is ample evidence that combining different connectomes and behavioral measures reveals more robust neural correlates. Here, we proposed a prediction framework that combines connectomes from multiple sources (e.g. task and resting-state fMRI) and predicts a latent phenotype, derived from a battery of behavioral measures. The framework relies on a novel generalization of canonical correlation analysis with both a closed-form and an iterative solution. We applied the framework to data from the Human Connectome Project (HCP) to predict a latent, general intelligence factor. Prediction accuracy was higher for this latent factor than any single measure of intelligence, showing the advantage of combining multiple connectomes and behavioral measures in a single predictive model.