A physics-informed deep learning framework for modeling of coronary in-stent restenosis

Area of Science:

• Biomedical Engineering
• Computational Biology
• Cardiovascular Research

Background:

• Machine learning (ML) shows promise in cardiovascular surgery but struggles with in-stent restenosis (ISR) prediction due to data limitations and measurement variability.
• Accurate modeling of neointimal hyperplasia is crucial for understanding and predicting ISR.
• Existing ML models lack the ability to fully capture the complex biological mechanisms driving ISR.

Purpose of the Study:

• To develop a robust multiphysics surrogate model for in-stent restenosis (ISR) estimation using physics-informed deep learning (DL).
• To incorporate biological constraints and drug elution effects for enhanced prediction accuracy.
• To provide insights into ISR progression factors and aid in diagnosis and treatment planning.

Main Methods:

• Utilized physics-informed neural networks (PINNs), a DL approach integrating physical laws with data.
• Developed a set of coupled advection-reaction-diffusion partial differential equations (PDEs) to model ISR factors (e.g., PDGF, TGF-β, ECM, SMC density, drug concentration).
• Integrated patient-specific data (procedural, clinical, genetic) into the PINN model.

Main Results:

• The PINN model effectively tracks the evolution of key factors influencing ISR.
• The model demonstrates potential for improved prediction accuracy by incorporating biological and physical constraints.
• Patient-specific data integration enhances model performance and aids in risk mitigation.

Conclusions:

• Physics-informed DL offers a powerful framework for modeling complex biological processes like ISR.
• The developed multiphysics surrogate model advances predictive capabilities for ISR.
• This approach holds potential for personalized risk assessment and optimized treatment planning in cardiovascular surgery.