CAREER: Bayesian Inference Networks for Model Ensembles

点击翻译按钮获取中文摘要

Integration of cardiovascular models, expert opinion and clinical measurements in a coherent system for probabilistic reasoning represents the next frontier of model-aided diagnostics to inform treatment in personalized medicine. However, several fundamental limitations must be addressed: (1) deterministic models are inadequate to characterize the effect of uncertainty in various properties of the cardiovascular system; (2) cardiovascular models are computationally expensive and therefore a stochastic treatment of these models may be too computationally expensive; (3) predictive models are needed in complex decision workflows to directly answer questions of clinical relevance in justifiable, probabilistic terms. This project proposes key advances in three complementary areas that are essential to make this next generation of model-aided diagnostics a reality: (1) highly scalable computational methods able to efficiently handle multiple instances of cardiovascular models with uncertain parameters; (2) new Monte Carlo estimators that leverage computationally inexpensive low fidelity surrogates; (3) new inference systems based on variables organized in networks, accommodating non-linear hemodynamic models, expert opinion and data. Since cardiovascular disease is the leading causes of death worldwide, this project serves the national interest by advancing the national health, prosperity and welfare, as stated by NSF's mission. The proposed research at the interface of computational mathematics and physiology offers an ideal framework to educate a diverse and globally competitive STEM workforce at the high school, undergraduate and graduate levels. Workshops and mini-symposia will also facilitate the exchange of ideas on stochastic cardiovascular modeling within the scientific community. Computational models are increasingly being adopted to inform treatment in personalized medicine but innovation in model-based diagnostics is still hindered by three main problems: (1) deterministic simulations, i.e., simulations with certain outputs providing a false sense of confidence; (2) stochastic simulations are typically associated with a dramatic increase in computational cost; (3) current paradigms in uncertainty quantification (UQ) need generalization to provide coherent inference frameworks combining physics-based models, expert opinion, observational data and experiments. Thus, the main goal of this CAREER proposal is to develop the next generation of efficient computational tools to accelerate inference from models and data in computational hemodynamics as well as in a wide range of applications. This goal is achieved through the following objectives: (1) development of efficient ensemble solvers for hemodynamics running on modern CPU/GPU hybrid architectures; (2) research in generalized approximate control variate Monte Carlo estimators to drastically reduce the time required to solve direct and inverse problems in uncertainty analysis; (3) extend Bayesian Networks combining numerical models, expert opinion and data in a coherent inference framework. This research will provide the scientific basis to construct the first model-based inference framework including experimental evidence, expert opinion, and to develop systems directly applicable to the clinical decision making process. Two prototype systems will be finally developed, with a focus on applications to pediatric surgery.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.