Date of Award

Fall 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Interdepartmental Neuroscience Program

First Advisor

Scheinost, Dustin

Abstract

Machine learning models trained on human neuroimaging data are increasingly used as tools for predicting individual behavioral phenotypes and enhancing precision medicine. However, multiple limitations hinder their current scientific and future clinical utility. In this dissertation, I address key challenges surrounding "brain-behavior" predictive model interpretability, generalization, and bias. In Chapters 2 and 3, I show that prevalent feature selection practices overlook features that are meaningful for both prediction and neurobiological interpretation. This suggests that prevailing approaches lead to feature overinterpretation and a misrepresentation of the neurobiological bases of brain-behavior associations. In Chapters 4 and 5, I demonstrate that brain-behavior predictions can survive across the dataset idiosyncrasies that will be inherent to real-world use scenarios. In Chapters 6 and 7, I identify neuroimaging predictive model performance bias in rural populations. I also explore additional means through which predictive models in healthcare more broadly may exhibit decreased accuracy when deployed in rural populations and settings. Collectively, the insights presented in this dissertation lay the groundwork for developing more interpretable, generalizable, and fair brain-behavior predictive models.

Download

Share

COinS