Digital Twins in Coronary Artery Disease: A Mathematical Roadmap

1. Problem Statement

The paper addresses the critical need for personalized medicine in the treatment of Coronary Artery Disease (CAD), the leading cause of death in the Western world. Current clinical assessments rely heavily on the Fractional Flow Reserve (FFR), an invasive metric requiring catheterization and pharmacological stress to measure pressure drops across stenoses. While effective for short-term decisions, FFR lacks long-term predictive power regarding plaque progression and future adverse events (e.g., myocardial infarction).

The core challenge is the estimation of Wall Shear Stress (WSS), a hemodynamic quantity known to be a strong predictor of CAD development and plaque instability. However, WSS cannot be directly measured non-invasively. Existing methods face a trade-off:

• Purely Data-Driven: Direct numerical differentiation of noisy ultrasound (US) velocity data amplifies errors.
• Purely Model-Driven (CFD): Computational Fluid Dynamics (CFD) requires precise boundary conditions (pressure/flow) which are often unavailable or invasive to measure, leading to uncertainty.

The authors propose a Digital Twin (DT) framework to bridge this gap, creating a bidirectional system that integrates patient-specific data with mathematical models to predict WSS and guide treatment without excessive invasiveness.

2. Methodology

The proposed Digital Twin operates on a bidirectional feedback loop: P2D (Physical-to-Digital) for personalization and D2P (Digital-to-Physical) for decision support.

A. P2D: Data Assimilation (Personalization)

The goal is to estimate the WSS by fusing sparse, noisy non-invasive velocity data (from Ultrasound) with the governing physics (Navier-Stokes Equations). The paper evaluates three mathematical approaches to achieve this:

1. Variational Reduced-Order Modeling (VDA-ROM):

   • Concept: Treats missing boundary conditions (inlet/outlet pressures) as control variables. It minimizes a cost function representing the mismatch between the CFD solution and observed data.
   • Optimization: Uses a Lagrange multiplier approach to solve the Karush-Kuhn-Tucker (KKT) system.
   • Acceleration: To make this clinically viable (minutes rather than hours), the authors employ Reduced Order Models (ROMs) using Proper Orthogonal Decomposition (POD). This separates the process into an "offline" stage (generating snapshots) and an "online" stage (rapid projection), reducing computational cost by orders of magnitude. Preliminary results suggest optimization can be achieved in ~15 minutes.

2. Continuous Reduced-Order Modeling (CDA-ROM):

   • Concept: Based on "nudging," this method inserts observational data directly into the dynamical system equations as a feedback term ($\mu(I_H(u) - I_H(v))$).
   • Advantage: Proven to converge exponentially to the true solution even with unknown initial conditions.
   • Application: The authors propose a CDA-ROM that handles under-resolved and convection-dominated flows (common in CAD with occlusions). It uses a nudging term to steer the ROM solution toward the sparse data, significantly improving accuracy over standard G-ROMs, especially when data is scarce.

3. Physics-Informed Neural Networks (PINNs) in DA:

   • Concept: A Neural Network is trained to minimize the residual of the Navier-Stokes equations while simultaneously fitting the observed data.
   • Enhancement: The standard PINN loss function is augmented with a data-mismatch term ($J^{DA}_{PINN}$).
   • Results: Preliminary tests on the Kim and Moin 2D problem show that DA-PINN drastically reduces $L_2$ errors for both velocity and pressure compared to standard PINNs, offering a fast, mesh-free alternative for WSS estimation.

B. D2P: From Data to Decision-Making

Once the digital state (WSS, hemodynamics) is estimated, the system must guide clinical decisions.

• Probabilistic Graphical Models (PGMs): The authors propose using PGMs to represent the DT as a graph of random variables (State $S$, Observations $O$, Digital State $D$, Actions $U$, Rewards $R$).
• Memory Integration: Unlike standard Markovian models (where $t+1$ depends only on $t$), the proposed framework uses Multi-Step Markov Chains. This allows the model to account for the rate of change in pathology (e.g., plaque growth speed), which is critical for distinguishing between stable and rapidly deteriorating conditions.
• Multifidelity & Uncertainty Quantification (UQ): The framework integrates data of varying fidelity (wearables vs. invasive CT/MRI) and model resolution. It utilizes a hierarchical graph structure to combine high-fidelity PDEs with low-fidelity surrogates and data, explicitly quantifying uncertainty to support robust decision-making.

3. Key Contributions

• Mathematical Roadmap for CAD DTs: The paper provides a structured, rigorous mathematical framework specifically for constructing Digital Twins for Coronary Artery Disease, moving beyond theoretical concepts to actionable algorithms.
• Novel WSS Estimation via DA: It proposes a hybrid strategy combining Data Assimilation with ROMs and PINNs to estimate the unmeasurable WSS from sparse US data, overcoming the boundary condition limitations of pure CFD.
• Advanced PGMs with Memory: The introduction of memory (multi-step dependencies) into PGMs for clinical DTs is a significant innovation, allowing the system to predict long-term pathology evolution based on historical trends rather than just instantaneous states.
• Multifidelity Integration: The proposal to use nonlinear manifolds and latent spaces to unify data and models of different fidelities addresses the challenge of heterogeneous data sources in clinical settings.

4. Results

• Computational Efficiency: The integration of ROMs with Data Assimilation reduces the time required for personalization from hours to approximately 15 minutes, making it compatible with clinical workflows.
• Accuracy Improvements:
  • CDA-ROM: Demonstrated significantly lower errors in velocity, lift, drag, and kinetic energy compared to standard G-ROMs in convection-dominated flows (Re=500, 1000).
  • DA-PINN: Showed a dramatic reduction in $L_2$ error (e.g., velocity error dropping from 0.074 to 0.0023 in benchmark tests) when data assimilation was added to the PINN loss function.
• Robustness: The methods demonstrated the ability to handle noisy data, sparse measurements, and complex geometries (including stented arteries), which are critical for real-world application.

5. Significance

This paper represents a pivotal step toward personalized, predictive, and preventive cardiology.

• Clinical Impact: By enabling non-invasive, rapid, and accurate estimation of WSS, the proposed DT could replace or augment invasive FFR measurements, reducing patient risk and healthcare costs.
• Long-term Prognosis: Unlike current tools that focus on immediate blockage, this framework targets the long-term development of CAD, potentially preventing myocardial infarction by identifying high-risk patients earlier.
• Interdisciplinary Bridge: It highlights the critical role of applied mathematics (numerical analysis, optimization, stochastic modeling) in modern medicine, arguing that the transition from "replicas" to "twins" requires rigorous mathematical integration of data and physics.
• Future Direction: The paper sets the stage for future research in geometric atlases, stabilization of ROMs in convection-dominated regimes, and the standardization of multi-center data protocols to validate these Digital Twins.