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.
CITATION STYLE
Gao, S., Shen, X., Todd Constable, R., & Scheinost, D. (2019). Combining Multiple Behavioral Measures and Multiple Connectomes via Multipath Canonical Correlation Analysis. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11766 LNCS, pp. 772–780). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-32248-9_86
Mendeley helps you to discover research relevant for your work.