Simultaneous data assimilation and cardiac electrophysiology model correction using differentiable physics and deep...

Area of Science:

• Computational Biology
• Biophysics
• Cardiovascular Research

Background:

• Patient-specific cardiac models, or 'digital twins', are crucial for diagnosing arrhythmias and personalizing treatments.
• Accurate cardiac electrophysiology (EP) models require balancing mathematical complexity, parameterization, and validation.
• Existing EP models are either computationally intensive (biophysical) or less realistic (phenomenological).

Purpose of the Study:

• To develop a hybrid framework leveraging deep learning to augment simplified cardiac models with data-driven components.
• To create a robust biophysical tool for improved cardiac EP modeling and predictions.
• To enable accurate prediction of cardiac transmembrane potential and parameter estimation.

Main Methods:

• A novel hybrid framework decomposing cardiac dynamics into physics-based and data-driven terms.
• Utilizing deep learning to learn from data and estimate model parameters simultaneously.
• Validation using both simulated (in silico) and experimental (ex vivo) optical mapping data of action potentials.

Main Results:

• The framework successfully reproduced complex cardiac transmembrane potential dynamics, even with noisy data.
• Accurate identification of key physical parameters for different anatomical zones was achieved.
• The model effectively reproduced action potential wave characteristics from various pacing locations.

Conclusions:

• The proposed physics-based, data-driven approach offers a robust method for cardiac EP modeling.
• This hybrid framework enhances the accuracy and efficiency of building predictive cardiac models.
• The approach holds potential for improving patient-specific diagnostics and treatment personalization in cardiology.