Research

Disease progression models (DPMs) are critical tools for characterizing the etiology of neurodegenerative disorders and capturing treatment effects in prospective clinical trials. Their development is complicated by subject-level variability in age of onset, rate of progression, and signal amplitude.

We develop Bayesian DPMs that model progression across heterogeneous biomarkers by explicitly modeling subject-specific variation in phase, amplitude, and rate. These produce inference on the timing and ordering of biomarkers and partition that variation between subject-level risk factors and individual variation. Current work explores threshold-aligned joint models for Alzheimer’s disease, extensions to time-to-event modeling, and the computational challenges of high-dimensional multi-modal imaging.

Neurodegenerative disorders cause cognitive decline by disrupting structure and connectivity in healthy brains, changes detectable only across multiple imaging modalities. Images carry either structural or network information, and must be linked through joint models to support principled clinical inference. Few statistical models integrate both, because the multimodal structure combines high-dimensional signals, complex correlations, and heterogeneous data types.

This gap does more than limit interpretation: it biases estimated effects, reduces efficiency, and increases sensitivity to noise. We develop Bayesian frameworks that treat multiple brain images as multi-objects, exploiting object topology while leveraging linkages among objects to perform inference, clustering, and prediction, alongside deep generative and explainable-AI approaches to spatial and network images. This work is motivated by imaging studies in primary progressive aphasia at the ALBA Language Neurobiology Laboratory.

Functional data analysis offers a powerful framework that embraces the underlying structure of these data by assuming the basic unit of observation is a signal observed over a continuous domain. This lets us model variation along frequency, spatial region, and time simultaneously rather than collapsing each dimension to a scalar summary.

This research is motivated by electroencephalography studies in children with autism spectrum disorder conducted with collaborators at The Jeste Developmental Neurophysiology Lab and the Autism Biomarker Consortium for Clinical Trials.