Cardiovascular Hemodynamic Digital Twins for Diagnostics, Continuous Monitoring, and Real-Time Intervention Planning

Abstract

Cardiovascular disease (CVD) remains the leading global cause of mortality and morbidity and is projected to impose a sevenfold increase in healthcare costs by 2050 due to its prevalence in aging populations. Despite advances in diagnostics and treatment, current clinical paradigms remain largely reactive, episodic, and limited in their ability to detect or prevent long-term complications such as in-stent restenosis—a re-narrowing of previously treated vessels-before substantial physiological damage occurs. While computational modeling has shown promise in identifying hemodynamic phenomarkers associated with disease, most approaches operate at isolated time points, are disconnected from real-time patient data, and lack the scalability required for routine clinical integration. These limitations highlight the critical need for personalized, predictive models that can continuously assess vascular health and inform proactive treatment decisions over clinically relevant timescales. This dissertation develops and validates the first, to our knowledge, cardiovascular digital twin framework capable of tracking 3D patient-specific hemodynamics over millions of heartbeats, forecasting intervention outcomes, and integrating continuous biometric data streams to inform dynamic clinical decision-making. We address three fundamental challenges in realizing this vision: (1) improving the accuracy of single timepoint phenomarker recovery, (2) extending the temporal domain of simulations to enable longitudinal capture of hemodynamics, and (3) supporting real-time prediction of patient-specific responses to therapy for proactive clinical planning. A cardiovascular digital twin is a continuously updated, patient-specific computational model that simulates individualized hemodynamics to reflect evolving physiology and support real-time clinical decision-making. Before realizing such digital twins, there are unmet needs in even accurately recovering single timepoint metrics. Current non-invasive approaches for estimating hemodynamic phenomarkers often operate in an open-loop fashion and fail to prioritize input features known to impact clinical accuracy. As a result, there are significant opportunities to improve diagnostic precision by identifying the minimal set of governing parameters necessary to recover phenomarkers and then refine predictions. Once single timepoint accuracy is optimized, the next challenge lies in scaling up our computational framework for longitudinal timescales. To enable digital twins that track evolving cardiovascular hemodynamics, computational models must be coupled with biometric data streams from wearable sensors to dynamically update a patient's physiological state. Simultaneously, the underlying simulation framework must be accelerated to support timely and tractable estimation of complex hemodynamics. To close the loop between continuous monitoring and decision-making, digital twins must also predict patient-specific responses to intervention in real time—enabling clinicians to plan and adapt therapies proactively in clinical settings. Therefore, there is an urgent need to address these foundational challenges in order to reduce long-term complications in cardiovascular disease, particularly in coronary and peripheral artery disease, through the development of the first cardiovascular hemodynamic digital twin.Toward the realization of a cardiovascular digital twin, this dissertation required the development of frameworks that expand the temporal domains over which we can tractably capture hemodynamics, continuously inform the digital twin of a patient’s evolving physiological state, and accurately forecast responses to intervention. To achieve this vision, we first addressed the challenge of optimizing the accuracy of single timepoint phenomarkers, focusing specifically on fractional flow reserve (FFR) in coronary artery disease. Through global sensitivity analyses encompassing over 5 million simulations, we identified key anatomical features—such as stenosis radius—and physiological parameters—such as cardiac output—as dominant contributors to FFR accuracy. These findings informed the development of a machine learning (ML)-refined framework that augments initial estimates from physics-based models with anatomical data to enhance diagnostic performance.To enable long-term dynamic flow capture and assess changes in cardiovascular physiology over extended periods, we introduced the Longitudinal Hemodynamic Mapping Framework (LHMF), which integrates wearable sensor data to dynamically update patient-specific models of physiological state and establish a parallel-in-time strategy that decomposes temporal dependence, enabling independent and high-throughput simulations across timepoints. Using this approach, we achieved near-perfect agreement with an explicit computational fluid dynamics (CFD) simulation of 750 heartbeats—the longest contiguous simulation of its kind reported to date. We further improved LHMF by introducing LHMF-C, a refinement that clusters hemodynamically similar simulations to reduce redundancy. This approach enabled the capture of 4.5 million heartbeats in just 43 hours using widely available cloud computing resources. LHMF can now transform conventional single-heartbeat analyses into spatiotemporal representations, which we term Longitudinal Hemodynamic Maps (LHMs). These maps overlay clinically relevant phenomarkers onto patient-specific anatomy across time, allowing us to visualize the duration and spatial distribution of exposure to disease-prone conditions. We anticipate that this framework will eventually facilitate the discovery of novel temporal phenomarkers and shift analyses from static to dynamic assessments of vascular health.To predict hemodynamic outcomes following intervention, we developed HARVEY Virtual Intervention (HarVI)—a real-time ML surrogate that emulates geometric interventions and predicts their hemodynamic impact within minutes. HarVI demonstrated high accuracy relative to physics-based ground truth across both coronary and peripheral artery disease, meeting the stringent intraoperative timelines required for clinical translation. In peripheral disease, where treatment planning is challenged by complex multi-lesion anatomy, HarVI demonstrated the first real-time, clinically validated prediction of post-intervention flow metrics.Collectively, these contributions establish a scalable, accurate, and clinically actionable digital twin platform for cardiovascular disease. By enabling proactive monitoring, real-time therapeutic planning, and dynamic forecasting of disease progression, the work lays the foundation for a paradigm shift from reactive to personalized and proactive cardiovascular care.