Physics-Informed and Data-Driven Inference of Cardiovascular Hemodynamics, Thrombosis Risk, and Model Parameters

Abstract

Cardiovascular hemodynamics are governed by blood transport through complex cardiac chambers, veins, and arteries. Quantities such as velocity, pressure, residence time, and vascular resistance are closely linked to cardiac function, thrombosis risk, and clinical decision-making, yet many are difficult or invasive to measure directly in vivo. Medical imaging provides complementary but incomplete information, while computational fluid dynamics (CFD) can simulate realistic blood flow but often requires accurate boundary conditions, additional measurements, and substantial computational resources. These limitations motivate approaches that combine physical laws, medical imaging, and data-driven inference to recover cardiovascular hemodynamics from sparse and indirect measurements. This thesis develops physics-informed neural network frameworks and data-driven methods for patient-specific inference of cardiovascular flow, thrombosis risk, and model parameters. The central idea is to embed governing physical principles, including the incompressible Navier–Stokes equations, indicator dilution theory, and lumped-parameter models, into machine learning and model-based inference frameworks. These constraints enable reconstruction of hemodynamic quantities that are not directly measurable from standard medical imaging, without requiring comprehensive and computationally expensive direct numerical simulations. A major focus of this work is AI-VFM, a physics-informed framework for left ventricular (LV) vector flow mapping (VFM) from color-Doppler echocardiography. AI-VFM uses mass conservation, momentum balance, and boundary conditions to infer 2D intraventricular velocity and pressure fields from standard Doppler acquisitions. Validation against CFD data and application to clinical echocardiographic acquisitions demonstrate its ability to recover velocity and pressure fields, perform phase unwrapping, fill missing data, and generate super-resolved hemodynamic maps. This framework is further extended to time-resolved 3D reconstruction using triplane color-Doppler imaging, where sparse velocity measurements from three intersecting planes are used to infer volumetric LV velocity and pressure fields. Validation against CFD simulations shows that 3D AI-VFM recovers dominant flow features, including vortical structures, velocity patterns, and pressure fluctuations. This thesis also investigates left atrial flow and thrombosis risk using contrast-enhanced computed tomography (CT). In atrial fibrillation, blood stasis in the left atrial appendage (LAA) contributes to thrombus formation and stroke risk, but blood stasis remains difficult to quantify non-invasively. To address this challenge, a physics-informed neural network is used to reconstruct LA flow and residence time fields from contrast dynamics, while an indicator dilution theory framework provides rapid residence time estimation from contrast concentration curves. Benchmarking against patient-specific CFD simulations demonstrates that contrast-enhanced 4D CT contains sufficient transport information to infer atrial flow, LAA stasis, and elevated thrombosis-related risk. Finally, this physics-informed inference paradigm is extended to reduced-order modeling of post-Norwood single-ventricle circulation. A lumped-parameter-model-informed neural network is developed to infer patient-specific hemodynamic states and model parameters from clinically measurable quantities. Using synthetic data from a physiologically derived Norwood circulation model, the framework recovers flow, pressure, and volume waveforms while estimating sensitive vascular parameters such as pulmonary vascular resistance and shunt resistance. Together, these methods advance a unified framework for cardiovascular hemodynamic inference from sparse, noisy, and indirect clinical data. By combining medical imaging, physics-informed neural networks, indicator dilution theory, and lumped-parameter modeling, this work demonstrates how physical constraints can transform limited measurements into quantitative biomarkers of flow, pressure, residence time, thrombosis risk, and patient-specific cardiovascular parameters.