Physics-informed self-supervised learning for predictive modeling of coronary artery digital twins

TL;DR

PINS-CAD transforms routine angiography into a simulation-free, physiology-aware framework for scalable, preventive cardiology, and generates spatially resolved pressure and fractional flow reserve curves, providing interpretable biomarkers.

Abstract

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.

Figures (7)

• Figure 1: Overview of the proposed PINS-CAD, physics-informed self-supervised learning framework for predictive modeling of coronary artery digital twins. The framework consists of four stages. a: Image-to-digital twin construction. Paired ICA images from two views are used to reconstruct a 3D anatomical model of the coronary artery, referred to as a real digital twin with pointwise correspondence. b: A3M: Anatomy-aware augmentation module for synthetic digital twin generation. Two real digital twins are randomly selected: one contributes the artery centerline, the other provides the radius profile. The centerline undergoes geometric augmentation to simulate anatomical variation. A new synthetic digital twin is then generated by sweeping the radius along the augmented centerline to define its boundary. c: Physics-informed self-supervised pretraining. A large dataset of 200,000 synthetic digital twins is used to pretrain a graph neural network (GNN). Artery graphs are constructed and the GNN is trained to predict pressure and velocity along the centerline, guided by physical constraints based on the 1D Navier–Stokes equations and pressure drop consistency. d: Fine-tuning for downstream tasks. The pretrained GNN is fine-tuned on real digital twins to predict future cardiovascular events using learned hemodynamic features.
• Figure 2: Model performance and ablation analysis of PINS-CAD. A: Receiver operating characteristic (ROC) curves and AUROC values (mean ± standard deviation) for PINS-CAD and four baseline models: PointNet++, Tabular MLP, AngioGraph, and Centerline-PINS. B: Precision–recall (PR) curves with PR-AUC values for the same models, with shaded areas representing the standard deviation. C: Decision curve analysis showing net benefit across threshold probabilities, including the "Treat None" and "Treat All" strategies as references. D-F: Ablation analyses. All bar plots show mean ± standard deviation. D, Effect of physics-informed pretraining: comparison between PINS-CAD, which uses physics-informed pretraining followed by supervised fine-tuning, and the same backbone trained from scratch without pretraining. E, Effect of physics-informed pretraining under centerline-only input: comparison between a model using physics-informed pretraining and another trained from scratch with supervised learning only. F, Comparison between PINS-CAD and Centerline + Phys, both using physics-informed pretraining but differing in backbone architecture (artery-graph vs. centerline-based), to assess the impact of model structure.
• Figure 3: Model performance of PINS-CAD and baseline methods in the contradictory region of FFR and DS. A: Scatter distribution of FFR and DS values across all lesions, illustrating regions of concordance and discordance based on the diagnostic thresholds (FFR = 0.80, DS = 50%). B: Comparison of overall performance metrics (Accuracy, AUROC, and F1-score) for PINS-CAD and four baseline models: PointNet++, Tabular-MLP, AngioGraph, and Centerline-PINS. C: Receiver operating characteristic (ROC) curves and corresponding AUROC values (mean ± standard deviation) across all methods. D: Precision–recall (PR) curves and PR-AUC values for the same models, with shaded areas indicating the standard deviation.E: Decision curve analysis illustrating net benefit across threshold probabilities, including “Treat None” and “Treat All” strategies as references.
• Figure 4: Comparison of model performance with and without synthetic-data pretraining. PINS-CAD was pretrained on anatomically diverse synthetic digital twins. SysAblation used the same physics-informed pretraining protocol but only on real data. The Backbone model was trained from scratch without any pretraining. Performance was evaluated using accuracy, F1 score, AUROC, and AUPRC.
• Figure 5: The predicted pressure and FFR from physics-informed self-supervised learning serve as hemodynamic biomarkers, enabling physiological explainability of the model’s predictions. A-B. Comparison of mean pressure and pressure drop between lesion and non-lesion segments. Lesion regions show significantly lower pressure and higher pressure drop, consistent with physiological expectations. Statistical significance is assessed using a two-sided t-test. Error bars denote standard deviation. C-F. Four visualization examples of predicted pressure distributions and corresponding FFR curve along the 3D coronary centerline, sampled at 500 evenly spaced points. The left panels in each pair show the pressure distribution mapped onto the artery, while the right panels show the FFR curve computed as the ratio of local pressure to inlet pressure. Red dots indicate lesion regions identified by anatomical labeling. The FFR curves derived from predicted pressure reveal functionally significant drops in flow reserve consistent with lesion severity.