Characterizing brain-phenotype relationships in health and disease

Date of Award

Fall 10-1-2021

Document Type


Degree Name

Doctor of Philosophy (PhD)


Interdepartmental Neuroscience Program

First Advisor

Constable, R. Todd


Relating whole-brain activity patterns to human behaviors, traits, and clinical symptoms holds the promise of revealing the macroscale neural bases of these complex phenotypes. In this work, I advance predictive modeling approaches to characterize such brain-phenotype relationships. In the first empirical chapter, I demonstrate that in-scanner tasks amplify these relationships; models built from fMRI data acquired while participants performed various tasks better predict intelligence measures than do models built from resting-state fMRI data. In the second empirical chapter, I explore how tasks have this effect, using psychophysiological interaction (PPI) and predictive modeling analyses to show that spatially distributed, task-induced changes in functional connectivity predict phenotype independent of activation. Activation, however, is useful for prediction only if the in-scanner task is related to the predicted phenotype. Further, I developed an inter-subject PPI analysis to demonstrate that tasks have this effect, in part, by synchronizing the BOLD signal. In the final empirical chapter, I use fMRI and phenotypic data from a clinically and demographically heterogeneous sample, collected in a study that I designed and oversaw, to explore the failure of these models, asking whether there are groups of individuals with distinct brain-phenotype relationships. Results suggest that there is no single brain-phenotype relationship across all individuals. Rather, using functional connectivity to predict a range of phenotypic measures, I find that model failure is consistent for a given individual, generalizes across related phenotypic measures and datasets, and is associated with performance that is inconsistent with the archetypal profile for high and low scorers. These results have important implications for model generalizability and bias. Together, this work offers a framework to identify robust, individualized macroscale networks underlying a range of phenotypes with relevance in both health and disease.

This document is currently not available here.