Date of Completion

Spring 6-9-2020

Thesis Advisor(s)

Fumiko Hoeft

Honors Major

Individualized Major

Disciplines

Behavior and Behavior Mechanisms | Medicine and Health Sciences

Abstract

Over the last 20 years, advances in computational neuroimaging and computational power have made it feasible to create predictive models (Woo et al. Nature Neuroscience 2017). Predictive modeling is an approach that uses pattern recognition techniques (machine learning) to develop models using brain data to predict clini­cal (or educational) outcomes, differential diagnosis and subtyping, and inform development of new treatments (Doyle et al Royal Society 2015, Haynes Neuron 2015, Orrù et al. NBR 2012; Woo et al. Nature Neuroscience 2017). In recent years, machine learning algorithms have been implemented to develop a model (pattern classifier) using neuroimaging data to predict reading outcomes in children with a wide range of reading ability (Hoeft et al. Behav Neurosci 2007) and those diagnosed with reading disorders (RD) (Hoeft et al. PNAS 2011). In their studies, they showed that models combining neuroimaging and behavior were superior to just behavioral measures (Hoeft et al. Behav Neurosci 2007), and that neuroimaging data was able to predict reading outcome in RD more quickly and efficiently or when behavioral measures failed to do so (Hoeft et al. PNAS 2011).

For this project, we used resting state functional MRI (rsfMRI) data coupled with multivariate pattern analysis (MVPA) to develop models that predict RD diagnosis in a large population of children. rsfMRI uses blood oxygen level dependent (BOLD) signals to provide information about functional activation and connectivity between both local and nonlocal brain regions. Through MVPA, in particular support vector machines (SVMs) and random forest classifiers, patterns of temporal connectivity that differentiate between RD and non-RD children were identified and the accuracy of the model was calculated. Further exploratory analyses are performed to identify patterns that differentiate RD and controls in younger versus older children such that potential compensatory mechanisms and developmental differences are identified. Such tools may offer clinicians the ability to, in conjunction with behavioral techniques, more quickly and accurately diagnose children not just with RD but with a wide range of neurocognitive disorders and allow for better diagnostic criteria in the future.

Share

COinS