Regression with Structured Features at Multiple Scales to the Study of General Cognition in Children
Abstract
This article is motivated by an application, where we aim to comprehend the neural underpinnings of general cognition, a pivotal indicator of healthy brain development, by examining the relationship between structural task-based brain activation maps and resting-state brain connectivity graphs in children aged 9-10 years old. While prior studies have identified certain brain regions linked to general cognition, these findings predominantly rely on analyses focusing on a single image modality, such as the resting-state graph alone. Moreover, no structured regression technique currently exists to assess the collective impact of both structural and graph features on general cognition while preserving linkage between their topology. To address this gap, this article focuses on developing a regression model with a scalar outcome and two sets of imaging features obtained at different scales: (a) a \emph{graph}-valued feature with ``labelled" nodes at a coarse scale, quantifying interconnections between nodes in the form of a brain connectome graph from resting state functional magnetic resonance imaging (fMRI); and (b) \emph{structural} features at a finer scale \emph{nested} within each graph node in the form of task-based brain activation maps. We introduce a novel flexible Bayesian regression framework that harnesses the relational information of nodes in the graph-valued feature and the nested architecture between graph and structural features through a novel joint prior structure on coefficients. We refer to the proposed framework as Bayesian Multi-Object Feature Regression (BMFR). The framework enables inference on significant nodes in the graph predictive of the outcome, coefficients for features at both scales, and predictive inference for the outcome, each accompanied by precise characterization of uncertainty. The implementation utilizes an efficient Markov Chain Monte Carlo algorithm. Results from simulations showcase the framework's excellent performance in terms of influential node inference, regression coefficient estimation, and outcome prediction, outperforming popular competitors such as high-dimensional regression approaches, tree-based models, and deep neural networks. Application of BMFR to the multi-modal imaging data identifies two parieto-frontal resting state networks and constituent structural regions activated during a working memory task that provide new evidence to support existing theories of neuronal integration.
Department
StatisticsCollections
Citation
Gutierrez, Rene; Guhaniyogi, Rajarshi; Scheffler, Aaron (2024). Regression with Structured Features at Multiple Scales to the Study of General Cognition in Children. Available electronically from https : / /hdl .handle .net /1969 .1 /200962.
The following license files are associated with this item: