Multimodal AI for aortic disease

Project Summary/Abstract The aorta is central to cardiovascular health, with its structure and function significantly influencing disease risk and organ damage. Aging leads to structural changes such as elastin degradation, fiber rupture, collagen deposition, and calcification, resulting in adverse hemodynamic effects, including increased cardiac workload and microvascular damage. While aortic stiffening has been extensively studied, the effects of age-related changes in aortic geometry on cardiovascular health are less understood. Advances in artificial intelligence, particularly convolutional neural networks, have enabled automated quantification of aortic structures, uncovering prognostic phenotypes and genetic loci associated with aortic structure. These initial findings suggest that aortic structure has significant predictive value for incident cardiovascular events, and that aortic structure is influenced by genetic factors; however, time-varying remodeling patterns of aortic geometry, and their relationship towards cardiovascular disease and genetics are unknown. Generative multimodal modeling offers a solution by learning the underlying data distribution and creating task-agnostic, low-dimensional representations that can be leveraged for multiple downstream tasks. This research will address key gaps in our understanding of aortic structure by developing generative and multimodal AI methods to simulate age- and disease-associated remodeling and creating an organ-specific and genetically-informed foundation models. Aim 1 seeks to develop a novel generative method to simulate three-dimensional aortic remodeling patterns. Using imaging, clinical, and genetic data from the UK Biobank, a novel variational autoencoder (VAE) will be designed to integrate imaging and clinical-genetic features. Aortic remodeling patterns will be generated by varying clinical input variables and analyzing point-cloud shape variations. Through these simulations, novel phenotypes will be identified based on 3D mesh variations. Novel phenotypes will be validated using time-to-event analyses. Aim 2 focuses on constructing a foundation model to improve diagnostic and prognostic tools for aortic and cardiovascular diseases. Using multimodal contrastive learning, this model will integrate over 1,000,000 CT scans, paired radiology reports, and genomic data from the Penn Medicine Biobank to align imaging features with organ-specific measurements and genetic information. Segmentation and phenotype extraction protocols will be employed to compute organ-specific features. Genetic information will be incorporated using 100+ polygenic scores. The foundation model will support advanced classification tasks, predicting incident cardiovascular events with improved precision and generalizability compared to task-specific models. Together, these efforts will enhance understanding of aortic remodeling processes, inform mechanisms underlying disease progression, and provide innovative tools for cardiovascular risk assessment and intervention.