Neural Operator Learning to Predict Aneurysmal Growth and Outcomes

PROJECT SUMMARY
Despite continuing advances in medical genetics, medical imaging, and surgical interventions, thoracic aortic
aneurysms (TAAs) are increasingly responsible for significant morbidity and mortality. Large clinical studies
reveal the complexity of the disease, which typically presents sporadically in older individuals, with uncontrolled
hypertension amongst the key risk factors, while also presenting in younger individuals having genetic or
congenital predispositions. Standard methods (including multivariate regressions) have failed to improve
prediction of life-threatening acute aortic syndromes (dissection and rupture) and current AHA/ACC guidelines
based on maximum aortic diameter fail to predict risk. Further complicating the situation, recent data show that,
although life-saving, surgical repair of the proximal aorta with a prosthetic graft increases incidence of distal
aortic disease and acute events, thus emphasizing the need to time surgery appropriately – that is, either
unnecessary delays due to adherence to current guidelines or pre-mature intervention may increase risk to
patients. There is a dire need for a better approach for predicting thoracic aortic growth and potential outcomes.
This proposal is significant for it is designed to resolve this unmet clinical need; it is innovative for we propose a
novel mechanobiological and biomechanical data-driven approach to develop a next-generation (neural operator
based) machine learning tool that can better predict TAA growth and certain outcomes, including drug efficacy.
We will combine a novel repurposing of extant murine and human data, generation of ~25000 new synthetic data
sets, and collection of unique new murine data (12 models of TAAs) to identify the best machine learning
approach, then combine extant and prospective clinical imaging data (~300 patients) to train and test the final
neural network (a deep operator neural network, or DeepONet). Our proposed unique meta-learning framework
is simply not possible with standard neural networks. We will exploit multi-fidelity training so that both low
resolution data and relatively inaccurate models can be used in training when combined with high-fidelity real or
synthetic data and uncertainty quantification via functional priors (the most informative Bayesian priors) that are
learned by combining historical data, biophysical models, and GANs (generative adversarial networks). This
unique combination allows us to learn posteriors with few samples (e.g., 2 or 3 new medical images), hence
predictions can be made for new cases with minimal (clinical) information. This project is possible given our
highly collaborative team of physician-scientists, bioengineers, and applied mathematicians having a strong track
record of successful research (grants, papers) and training of diverse students, post-docs, and residents.

Grant Number: 5R01HL168473-03
NIH Institute/Center: NIH
Principal Investigator: Roland Assi