Fast Posterior Estimation of Cardiac Electrophysiological Model Parameters via Bayesian Active Learning

Area of Science:

• Computational Biology
• Biophysics
• Cardiovascular Research

Background:

• Probabilistic estimation of cardiac electrophysiological model parameters is crucial for personalized medicine and uncertainty quantification.
• Direct Markov Chain Monte Carlo (MCMC) sampling is computationally intensive due to expensive model simulations.
• Surrogate models offer computational efficiency but often lack sufficient accuracy for posterior probability density function (pdf) approximation.

Purpose of the Study:

• To develop a novel Bayesian active learning method for direct approximation of cardiac model parameter posterior pdfs.
• To enhance the accuracy and efficiency of parameter estimation in complex cardiac electrophysiological models.
• To enable high-dimensional parameter inference at the resolution of cardiac mesh.

Main Methods:

• Integration of a generative model within Bayesian active learning to handle high-dimensional parameter spaces.
• Introduction of new acquisition functions designed to prioritize shape approximation over mode-finding for posterior pdfs.
• Intelligent selection of training points to query the simulation model, minimizing sample requirements.

Main Results:

• The proposed method demonstrated superior accuracy in approximating posterior pdfs compared to standard Bayesian active learning.
• Significant reduction in computational cost was achieved relative to traditional and accelerated MCMC sampling techniques.
• Effective estimation of tissue excitability in a 3D cardiac electrophysiological model was validated using synthetic and real data.

Conclusions:

• The developed Bayesian active learning approach offers a computationally efficient and accurate solution for probabilistic parameter estimation in cardiac electrophysiology.
• This method facilitates improved model personalization and uncertainty quantification for cardiac models.
• The novel acquisition functions enhance the learning process for complex posterior probability distributions.