Deep Learning Models of the Developing Brain: Brain Age, ADHD, and Brain Behavior Associations in the ABCD Study: Mohammadreza Bayat's Dissertation Defense

About

UNIVERSITY GRADUATE SCHOOL BULLETIN ANNOUNCEMENT

Florida International University

University Graduate School

Doctoral Dissertation Defense

Abstract

Deep Learning Models of the Developing Brain: Brain Age, ADHD, and Brain Behavior Associations in the ABCD Study

by

Mohammadreza Bayat

Two children of the same age can differ markedly in how quickly they learn, how well they hold information in mind, and whether they go on to develop a psychiatric condition. Large consortium datasets such as the Adolescent Brain Cognitive Development (ABCD) Study, which follows nearly 11,860 children from age 9 with multimodal magnetic resonance imaging (MRI) and extensive behavioral assessment, now make it possible to study these differences with unprecedented statistical power. This dissertation examines how the brain is represented before analysis and how models that link brain to behavior are built, validated, and interpreted, using deep learning models applied to whole-brain MRI from the ABCD Study. The research addresses two key domains: (1) brain maturation in attention-deficit/hyperactivity disorder (ADHD), using brain-age models to test the maturational lag hypothesis; and (2) prediction of cognition and other phenotypes from brain images, investigating whether the apparent ceiling on brain-behavior prediction reflects limitations of the brain or of the analytical approach. Study 1 provides a comprehensive review of the brain-age framework, tracing the development of regression-based and deep learning approaches, surveying applications across neurodevelopmental, psychiatric, and neurodegenerative conditions, and identifying unresolved challenges including sex differences, head motion, and generalization across datasets. Study 2 employs ten deep learning architectures to test the maturational lag hypothesis of ADHD, examining how model accuracy relates to clinical sensitivity and whether conclusions depend on a single training run. Study 3 asks whether the apparent ceiling on brain-behavior prediction reflects the brain or how it is represented, using a hybrid three-dimensional convolutional-attention model trained on unreduced voxelwise maps to predict working memory performance, composite cognition, and waist circumference. Together, these studies address how much of what neuroimaging can reveal about individual differences in development and behavior depends on how brain data are represented and how models are built, validated, and interpreted. The findings have direct implications for designing more robust brain-behavior studies and improving the translation of neuroimaging research into clinical insight.