PROJECT SUMMARY/ABSTRACT In this proposal, we will use modern machine learning techniques to combine and enhance computational modeling predictions. We will overcome the physics deficiencies that are inherent in modeling assumptions by including ground-truth clinical measurements, but in turn provide predictions that are more informative (higher spatial and temporal resolution) than the original clinical measurements. Furthermore, we will implement this framework as a surrogate model that can be used in real time and can replace current models with prohibitively high computation cost. If successful, the proposed research will enable lab-to-bedside deployment of a vast array of existing and future computational models and it ultimately could lead to a paradigm shift in health care workflow. Our overarching hypothesis is that the statistical correlations between computational models and clinical measurements can be exploited in a probabilistic data-fusion framework for more accurate predictions. Our multi-fidelity framework is based on an autoregressive Gaussian Process (GP) scheme. Our proposed scheme is a non-parametric Bayesian machine learning technique that has a probabilistic workflow and estimates uncertainty at different levels of fidelity in a principled manner. As a template for other clinical applications, we will develop this framework for perfusion scanning of brain hemodynamics in healthy and stroke populations, which has a significant health application. In Aim 1, we will simulate cerebral perfusion in healthy and stroke populations based on CT and MR angiography (CTA and MRA) scans. We will simulate and validate cerebral blood perfusion in healthy and stroke gender-balanced subjects. In Aim 2, we construct subject-specific multi-fidelity models by combining computational results and perfusion scans. We propose to leverage the multi-fidelity model to reduce scan time and radiation exposure by incorporating simulated perfusion maps with CT perfusion scans.