Learning cardiac activation and repolarization times with operator learning.

Area of Science:

• Computational biology
• Cardiac electrophysiology
• Machine learning

Background:

• Cardiac electrophysiology simulations using differential equations are computationally intensive.
• High-resolution meshes are needed for accurate heart dynamics, increasing computational cost.
• Clinical applications require efficient and interpretable computational tools.

Purpose of the Study:

• To explore operator learning approaches for cardiac electrophysiology.
• To learn mappings from applied stimulus to activation and repolarization times.
• To evaluate the efficiency and robustness of these data-driven methods.

Main Methods:

• Utilized Fourier Neural Operators (FNO) and Kernel Operator Learning (KOL).
• Learned operator mapping applied stimulus to activation and repolarization time distributions.
• Evaluated methods on synthetic 2D/3D domains and a left ventricle geometry.

Main Results:

• FNO and KOL efficiently learned the stimulus-to-activation and stimulus-to-repolarization time mappings.
• These methods are computationally efficient compared to traditional Monodomain models.
• Both approaches demonstrated robustness to hyperparameter choices.

Conclusions:

• Operator learning offers a computationally efficient alternative to traditional PDE models in cardiac electrophysiology.
• FNO and KOL can accelerate cardiac simulations for potential clinical integration.
• These surrogate operators provide valuable insights, especially for repolarization dynamics where PDE models are lacking.