High-dimensional Bayesian Optimization of Personalized Cardiac Model Parameters via an Embedded Generative Model

Jwala Dhamala, Sandesh Ghimire, John L. Sapp, B. Milan Horácek, Linwei WangMICCAI 2018 · DOI · OpenReview

Abstract

The estimation of patient-specific tissue properties in the form of model parameters is important for personalized physiological models. However, these tissue properties are spatially varying across the underlying anatomical model, presenting a significant challenge of high-dimensional (HD) optimization in the presence of limited measurement data. A common solution to reduce the dimension of the parameter space is to explicitly partition the anatomical mesh, either into a fixed small number of segments or a multi-scale hierarchy. This anatomy-based reduction of parameter space presents a fundamental bottleneck to parameter estimation, resulting in solutions that are either too low in resolution to reflect tissue heterogeneity, or too high in dimension to be reliably estimated within feasible computation. In this paper, we present a novel concept that embeds a generative variational auto-encoder (VAE) into the objective function of Bayesian optimization, providing an implicit low-dimensional (LD) search space that represents the generative code of the HD spatially-varying tissue properties. In addition, the VAE-encoded knowledge about the generative code is further used to guide the exploration of the search space. The presented method is applied to estimating tissue excitability in a cardiac electrophysiological model. Synthetic and real-data experiments demonstrate its ability to improve the accuracy of parameter estimation with more than 10x gain in efficiency.

Introduction

Personalized cardiac models are powerful tools for treating heart disease, but they require precise "tuning" to match an individual patient's tissue properties, such as excitability or scarring. This parameter estimation presents two major challenges: the "curse of dimensionality" and high computational costs. Tissue properties vary across the entire heart mesh, creating a high-dimensional (HD) search space that is often ill-posed given sparse clinical data like ECGs. Furthermore, a single cardiac simulation can take hours, making traditional optimization methods that require thousands of iterations clinically unfeasible. Previous solutions often oversimplified the heart into rigid segments or used complex, inflexible multi-scale hierarchies that failed to capture irregular tissue heterogeneity.

Methodology

The authors propose a novel framework, BO-VAE, which embeds a generative Variational Auto-encoder (VAE) into the Bayesian Optimization (BO) loop. This approach operates in two distinct stages:

  1. Learning the Latent Space: A VAE is trained on a large dataset of synthetic tissue patterns. The encoder learns to compress complex HD heart maps into a low-dimensional (LD) latent code ($z$), while the decoder learns to reconstruct the full HD map from this code.
  2. Embedded Bayesian Optimization: Instead of direct optimization in the HD space, BO searches the LD latent space. A Gaussian Process (GP) models the relationship between the latent code and simulation error.

Crucially, the VAE’s learned probability distribution is used to guide the BO's acquisition function. This "informed exploration" forces the optimizer to prioritize regions of the latent space that represent realistic heart tissue patterns, preventing the algorithm from wasting time on physically impossible configurations.

Overview of the BO-VAE framework, illustrating the integration of a VAE with Bayesian Optimization for cardiac parameter estimation.

Experimental Setup

The VAE was trained on 116,000 to 155,000 synthetic heterogeneous tissue infarct configurations across three human heart models. Testing included 27 synthetic experiments with known ground truths and two clinical cases of patients undergoing catheter ablation for ventricular tachycardia. Clinical validation utilized invasive bipolar voltage mapping as the gold standard for comparison.

Results and Discussion

BO-VAE consistently outperformed fixed-segment (FS) and fixed-hierarchy (FH) methods in both accuracy and efficiency. It achieved higher Dice Coefficients (DC) and lower Root Mean Square Error (RMSE), indicating significantly better overlap with actual tissue abnormalities.

Quantitative performance comparison: (a) Dice Coefficient, (b) RMSE, and (c) number of model evaluations for BO-VAE versus baseline methods.

The most significant gain was in computational efficiency. BO-VAE reduced costs by approximately 87.6% compared to the FS method and a staggering 98.7% compared to the FH method. In clinical cases, BO-VAE reached accurate solutions in just 105 evaluations, compared to up to 5,800 for previous methods. This efficiency is driven by the VAE's ability to learn an implicit representation of irregular scar shapes and the "physics-aware" search guided by the latent manifold's geometry, where the search space organizes itself logically by infarct size and location.

Conclusion

By mapping high-dimensional cardiac properties into a manageable latent space, BO-VAE enables accurate and efficient personalized modeling suitable for clinical timeframes. Future work will focus on improving generative expressiveness with 3D imaging data and expanding the framework to estimate multiple interacting physiological parameters simultaneously.

Summary updated

Abstract

The estimation of patient-specific tissue properties in the form of model parameters is important for personalized physiological models. However, these tissue properties are spatially varying across the underlying anatomical model, presenting a significance challenge of high-dimensional (HD) optimization at the presence of limited measurement data. A common solution to reduce the dimension of the parameter space is to explicitly partition the anatomical mesh, either into a fixed small number of segments or a multi-scale hierarchy. This anatomy-based reduction of parameter space presents a fundamental bottleneck to parameter estimation, resulting in solutions that are either too low in resolution to reflect tissue heterogeneity, or too high in dimension to be reliably estimated within feasible computation. In this paper, we present a novel concept that embeds a generative variational auto-encoder (VAE) into the objective function of Bayesian optimization, providing an implicit low-dimensional (LD) search space that represents the generative code of the HD spatially-varying tissue properties. In addition, the VAE-encoded knowledge about the generative code is further used to guide the exploration of the search space. The presented method is applied to estimating tissue excitability in a cardiac electrophysiological model. Synthetic and real-data experiments demonstrate its ability to improve the accuracy of parameter estimation with more than 10x gain in efficiency.
BibTeX
@inproceedings{dhamala2018high,
  author = {Jwala Dhamala and Sandesh Ghimire and John L. Sapp and B. Milan Horácek and Linwei Wang},
  title = {{High-dimensional Bayesian Optimization of Personalized Cardiac Model Parameters via an Embedded Generative Model}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2018},
  doi = {10.1007/978-3-030-00934-2\_56},
}
Figure