Scholarlink

Efficient modeling of cardiac electrophysiology is essential for advancing personalized medicine and improving treatment strategies. Traditional mathematical models provide accurate simulations but are computationally expensive, limiting their scalability to complex geometries and their generalization to patient-specific applications. In this work, we propose AGATA, an autoregressive graph neural network-based approach for simulating cardiac action potential propagation. AGATA is trained on simple ellipsoid meshes yet demonstrates a remarkable ability to generalize to more complex and realistic geometries. It has been applied to meshes obtained from medical images, including one of the left ventricle endocardium (surface) and one of the two ventricles (volume). AGATA effectively captures spatial propagation in healthy tissue and reproduces patterns in damaged areas. The mean absolute error (MAE) between the values generated by the Finite Element Method and AGATA is \(3.0 \times 10^{-5}\) (dimensionless) for a simple ellipsoid, and 0.007 for more complex geometries. Unlike traditional methods, AGATA’s ability to adapt to previously unseen, intricate geometries without requiring retraining emphasizes its potential for patient-specific modeling. Additionally, AGATA achieves up to a twelve-fold reduction in computational time compared to the Finite Element Method, making it suitable for near-real-time applications while maintaining a good approximation of the electrical signal.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Learning Cardiac Electrophysiology with Graph Neural Networks for Fast Data-Driven Personalised Predictions

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Maëlis Morier,
  • Jairo Rodríguez Padilla,
  • Patrick Gallinari,
  • Maxime Sermesant

摘要

Efficient modeling of cardiac electrophysiology is essential for advancing personalized medicine and improving treatment strategies. Traditional mathematical models provide accurate simulations but are computationally expensive, limiting their scalability to complex geometries and their generalization to patient-specific applications. In this work, we propose AGATA, an autoregressive graph neural network-based approach for simulating cardiac action potential propagation. AGATA is trained on simple ellipsoid meshes yet demonstrates a remarkable ability to generalize to more complex and realistic geometries. It has been applied to meshes obtained from medical images, including one of the left ventricle endocardium (surface) and one of the two ventricles (volume). AGATA effectively captures spatial propagation in healthy tissue and reproduces patterns in damaged areas. The mean absolute error (MAE) between the values generated by the Finite Element Method and AGATA is \(3.0 \times 10^{-5}\) (dimensionless) for a simple ellipsoid, and 0.007 for more complex geometries. Unlike traditional methods, AGATA’s ability to adapt to previously unseen, intricate geometries without requiring retraining emphasizes its potential for patient-specific modeling. Additionally, AGATA achieves up to a twelve-fold reduction in computational time compared to the Finite Element Method, making it suitable for near-real-time applications while maintaining a good approximation of the electrical signal.