Title: Graph-informed disease modeling from neuroimaging to EHR data

Speaker:Yize Zhao, Associate Professor, Yale University.

Time: Wednesday, November 18, 15:00 PM,  Austin 313.

Link: https://teams.microsoft.com/meet/2262859116495162?p=kNRyhsdYOkh9DB3pbp

Abstract: Neurodegenerative and complex chronic diseases emerge from interactions that span biological networks and populations. I will introduce two Bayesian frameworks that quantify these interactions, from within-brain propagation to across-disease comorbidity. In the first project, aimed at characterizing Tau protein spread along functional networks in the early course of Alzheimer’s disease generated from the A4 study, we jointly model tau propagation, functional connectivity structure, and subgroup heterogeneity using cross-sectional data. By integrating graph-constrained infection dynamics with connectivity patterns, the model infers plausible propagation pathways and subgroup-specific infection sequences. In the second project, using UK Biobank EHRs, we represent disease relationships through a latent hypergraph, where each hyperedge captures the higher-order clusters of diseases that share covariate-dependent risk. By uncovering disease hyperedges and their associated risk factors, we characterize how biological and lifestyle factors jointly influence sets of related conditions. To scale posterior inference while preserving uncertainty, we employ amortized variational inference with neural parameterization. Together, these studies yield probabilistic, graph-informed and interpretable views of disease pathology and organization.

Bio of  Yize Zhao: Yize Zhao, PhD, is a tenured Associate Professor of Biostatistics at Yale School of Public Health, with a secondary appointment in Biomedical Informatics & Data Science at Yale School of Medicine. Her research advances statistics, artificial intelligence, machine learning, and data science by bringing Bayesian inference together with deep learning, graph learning, and multimodal data integration. Her methodological contributions address how to learn from complex, interconnected data while accounting for individual differences, incomplete information, and uncertainty.