Differentiable Cardiac Electrophysiology Simulations for Dynamical State and Parameter Estimation

Area of Science:

• Computational Biology
• Cardiac Electrophysiology
• Biophysics

Background:

• Cardiac contractions are driven by action potential waves with complex dynamics.
• Current cardiac electrophysiology simulations use partial differential equations (PDEs), but fitting them to real-world data is difficult.
• Patient-specific models and digital twins are crucial for understanding heart conditions.

Purpose of the Study:

• To introduce differentiable cardiac electrophysiology simulations for automatic fitting to spatio-temporal action potential wave data.
• To enable the creation of patient-specific cardiac models and digital twins.
• To improve the diagnosis of heart rhythm abnormalities.

Main Methods:

• Developed a framework for differentiable cardiac electrophysiology simulations using finite-difference and smoothed particle hydrodynamics methods.
• Utilized gradient-based optimization and backpropagation to minimize a loss function comparing simulated and observed dynamics.
• Applied perceptual loss and generative diffusion models for fitting to noisy or sparse experimental data.

Main Results:

• Successfully fitted simulations to spatio-temporal action potential wave data, even with sparse or noisy observations.
• Located early activation sites in a 3D bi-ventricular geometry and fitted models to spiral wave imaging data.
• Demonstrated the framework's applicability to various data types (pixel-, voxel-, point-based) and geometries (2D/3D slabs, arbitrary shapes).

Conclusions:

• Differentiable simulations offer a powerful method for parameter learning and dynamic recovery in cardiac electrophysiology.
• This approach facilitates the development of personalized cardiac models and digital twins.
• Potential to significantly improve the diagnosis and treatment of cardiac rhythm disorders.