Cardiovascular disease is the leading global cause of mortality, with coronary artery disease (CAD) as its most prevalent form, necessitating early risk prediction. While 3D coronary artery digital twins reconstructed from imaging offer detailed anatomy for personalized assessment, their analysis relies on computationally intensive computational fluid dynamics (CFD), limiting scalability. Data-driven approaches are hindered by scarce labeled data and lack of physiological priors. To address this, we present PINS-CAD, a physics-informed self-supervised learning framework. It pre-trains graph neural networks on 200,000 synthetic coronary digital twins to predict pressure and flow, guided by 1D Navier-Stokes equations and pressure-drop laws, eliminating the need for CFD or labeled data. When fine-tuned on clinical data from 635 patients in the multicenter FAME2 study, PINS-CAD predicts future cardiovascular events with an AUC of 0.73, outperforming clinical risk scores and data-driven baselines. This demonstrates that physics-informed pretraining boosts sample efficiency and yields physiologically meaningful representations. Furthermore, PINS-CAD generates spatially resolved pressure and fractional flow reserve curves, providing interpretable biomarkers. By embedding physical priors into geometric deep learning, PINS-CAD transforms routine angiography into a simulation-free, physiology-aware framework for scalable, preventive cardiology.
Cardiovascular diseases (CVDs) remain the leading cause of global mortality with coronary artery disease (CAD) representing the most prevalent subtype 1 . Invasive coronary angiography (ICA) represents the gold standard test for CAD diagnosis. It permits the visualization of the coronary arteries and the anatomical evaluation of stenosis severity, such the quantification of diameter stenosis (DS) i.e. the percentage reduction in luminal diameter 2 . However, anatomical narrowing does not always correspond to physiological impairment. Fractional flow reserve (FFR), the gold standard hemodynamic measure of lesion severity 3 , provides complementary information on the functional impact of a stenosis, with current guidelines recommending the revascularisation of lesions with an FFR ≤ 0.8 2 . However, whilst the combination of an angiographic anatomical assessment with FFR permits the risk stratification of coronary stenoses, a significant proportion of such patients with intermediate stenoses with an FFR > 0.8 go on to experience adverse events during long-term follow-up 4,5 . The limitations of this current approach, together with the invasive nature, procedural risk, and cost of FFR measurements, highlight the need for personalized, non-invasive predictive frameworks that capture the complex interplay between coronary anatomy and physiology.
Recently, three-dimensional (3D) coronary artery digital twins, represented as point clouds, have emerged as a promising approach to patient-specific cardiovascular modeling 6 . Unlike conventional ICA-derived features, which are often restricted to local descriptors such as diameter stenosis, 3D digital twins capture detailed patient-specific vascular geometry with full spatial context [7][8][9] . They preserve continuous vessel topology and spatial relationships, enabling the integration of both local In this work, we propose PINS-CAD, a Physics-Informed, Self-supervised learning framework for predictive modeling of Coronary Artery Digital twins. Recognizing that hemodynamic features are vital for assessing lesion severity and cardiovascular risk, yet their estimation through CFD or invasive FFR remains resource-intensive, PINS-CAD integrates cardiovascular flow physics into a self-supervised pretraining stage. This enables models to learn physiologically meaningful hemodynamic representations from large-scale unlabeled coronary geometries without CFD supervision. By coupling physics-informed learning guided by hemodynamic principles with downstream clinical fine-tuning, PINS-CAD bridges the gap between geometric modeling and clinical outcome prediction, offering a scalable and interpretable solution for non-invasive risk assessment in coronary artery disease. Specifically, we reconstruct 3D coronary artery digital twins from paired ICA images acquired from two approximately orthogonal views, which provide sufficient angular separation to resolve depth ambiguity and accurately recover the spatial geometry of the coronary artery(Fig. 1a). To increase the training dataset diversity, we generate 200,000 synthetic coronary artery digital twins using A 3 M, an anatomy-aware augmentation method that recombines centerline and radius profiles from different real anatomies while introducing controlled perturbations of curvature, torsion, and vessel radius to mimic physiological variability (Fig. 1b). These anatomically diverse digital twins provide a foundation for learning hemodynamic behavior directly from geometry, enabling the prediction of key quantities such as pressure and velocities without relying on computationally intensive CFD simulations. To predict pressure and velocities from the digital twins, each digital twin is represented as a graph where boundary points sampled from the artery surface serve as nodes and edges are defined based on spatial proximity to capture local geometric and morphological patterns. A GNN-based 21,28 backbone is trained on these graphs to extract geometric and morphological features to predict pressure and velocities along the artery centerline.
To enable centerline-level predictions, node features extracted from the boundary graph are aggregated into the centerline. Training is guided by physics-based loss functions derived from the 1D Navier-Stokes equations and pressure drop laws (Fig. 1c), ensuring that the learned representations remain consistent with cardiovascular fluid dynamics. The pretrained GNN is then fine-tuned on real digital twins to perform downstream clinical tasks such as cardiovascular event prediction (Fig. 1d). By embedding fundamental physical principles into self-supervised pretraining, PINS-CAD ensures physiological consistency, reduces dependency on labeled CFD data, and transforms anatomical digital twins into predictive tools for preventive and personalized clinical decision-making. A large dataset of 200,000 synthetic digital twins is used to pretrain a graph neural network (GNN). Artery graphs are constructe
This content is AI-processed based on open access ArXiv data.
