Project Summary/Abstract
Significance. Due to a global rise in life expectancy, Alzheimer’s disease and other neurodegenerative disorders
(NDs) represent a looming health crisis necessitating improvements in disease monitoring and treatment. NDs
cause cognitive decline through disruption of structure and connectivity in healthy brains which can only be de
tected using brain images from multiple modalities. Specifically, images contain either structural information or
network information which must be linked via joint models to provide principled clinical inference. However, there
are few statistical models that integrate both types of information due to the difficult multimodal structure which
includes highdimensional signals, complex correlations, and heterogenous data types. This lack of appropriate
methods not only limits the interpretation of clinical findings, but also has been shown to bias estimated effects,
reduce statistical efficiency, and increase sensitivity to noise. To this end, hierarchical Bayesian methods allow
structured information to be shared among brain images via their joint prior structure but have not been embraced
due to a lack of theoretical guarantees, computational bottlenecks, and difficulty specifying prior structures. This
proposal develops a new Bayesian framework to address the theoretical, inferential, and computational chal
lenges of jointly modeling multimodal brain images. Viewing multiple brain images as multiobjects, we propose a
Bayesian objectoriented modeling (BOOM) framework for analyzing multiobject data that exploits object topology
while leveraging linkages among objects to perform inference, clustering, and prediction. Overall, the proposed
framework provides statisticians with much needed inferential and computational tools for highdimensional im
age analysis and allows clinical neuroscientists to fully leverage existing multimodal imaging datasets. The BOOM
framework will be distributed to the neuroimaging community via accessible software developed to study impactful
multimodal data collected by our collaborators to study NDs at the UC San Francisco Memory and Aging Center.
Innovation. Our proposal will be the first to develop Bayesian regression and supervised probabilistic clustering
with multiobject image predictors or responses, where at least one image is network valued and the rest are
a collection of spatially correlated cells. This provides significant innovation in estimation, Bayesian asymptotic
theory, and computation for structured Bayesian highdimensional regression models.
Aims. We advance the following aims: 1) (Scalar on multiobject regression) To develop, apply, and evaluate
a Bayesian regression framework that provides inference and prediction for a scalar outcome as a function of a
multiobject predictor; 2) (Supervised clustering of multiobject data) To develop, apply, and evaluate a Bayesian
multiobject mixture modeling framework that enables supervised clustering of multiobject responses via their as
sociation with a vector predictor; 3) (Object on multiobject regression) To develop, apply, and evaluate a Bayesian
regression framework that provides inference and prediction for an object outcome as a function of multiobject
predictor.