Bayesian optimisation for efficient parameter inference in a cardiac mechanics model of the left ventricle

Agnieszka Borowska, Hao Gao, Alan Lazarus, Dirk HusmeierInternational Journal for Numerical Methods in Biomedical Engineering 2022 · DOI

TL;DR

This paper introduces a Bayesian Optimization (BO) framework to efficiently estimate parameters in complex cardiac mechanics models. By leveraging BO's ability to optimize expensive black-box functions, the authors significantly reduce computational time compared to traditional methods, making these models more practical for clinical applications.

What Problem Does This Paper Address?

Understanding the mechanics of the heart is crucial for diagnosing and treating cardiovascular diseases. Cardiac mechanics models, particularly those describing the passive filling of the left ventricle (LV), are essential tools. However, these models are governed by complex equations that lack closed-form solutions, requiring computationally intensive numerical methods (like Finite Element Method - FEM) to solve. This computational cost makes it impractical to use these models for parameter inference (i.e., estimating the model's parameters from patient data) in a clinical setting, where speed and efficiency are paramount. The authors aim to bridge this gap by developing a faster and more efficient parameter inference method.

Methods

The core of the paper is the application and adaptation of Bayesian Optimization (BO) for parameter inference. BO is a global optimization technique designed for expensive-to-evaluate black-box functions. Its key idea is to build a probabilistic surrogate model (typically a Gaussian Process) of the objective function and use an acquisition function to intelligently select the next point to evaluate, balancing exploration of new regions and exploitation of promising ones.

The authors introduce several novelties:

  • Partial Error Surrogates: Instead of building a surrogate for the entire objective function (which is a sum of squared errors), they propose building separate surrogates for each individual error term. This approach, they argue, leads to a closed-form expression for the acquisition function and can improve accuracy.
  • Klotz-curve Penalty Term: To ensure parameter estimates are realistic even at high pressures (which are not directly observable in vivo), they incorporate a penalty term derived from an empirical law (the "Klotz-curve") based on ex vivo data. This helps constrain the model's behavior in unobserved physiological regimes.
  • Constraint Handling: They use a GP classifier to model the probability of "crash points" (parameter sets that cause the simulator to fail), allowing them to weight the acquisition function and avoid these problematic regions.
  • V30 Emulator: To further reduce computational cost, they develop an emulator for predicting the LV volume at high pressure (30 mmHg), avoiding the need to run the expensive simulator for this specific output.

Datasets

The study uses clinical in vivo data from cardiac magnetic resonance (CMR) scans of healthy volunteers. These datasets provide measurements of LV volume and circumferential strains at a physiological pressure (8 mmHg). The paper also references ex vivo data used to derive the Klotz-curve empirical law. No new datasets are introduced; rather, existing clinical data is used for validation.

Key Insights and Ideas

  • BO is a Game Changer for Cardiac Modeling: The paper strongly suggests that BO can revolutionize parameter inference in cardiac mechanics by drastically reducing computational costs, making complex models clinically viable.
  • Partial Error Surrogates Offer Robustness: While not universally superior, the partial error surrogate approach shows promise, especially when dealing with heterogeneous data, offering flexibility in modeling individual error components.
  • Bridging In Vivo and Ex Vivo Data is Crucial: Incorporating knowledge from ex vivo experiments (like the Klotz-curve) is essential for ensuring model predictions are physiologically realistic across a wider range of conditions, including those not directly observable in vivo.
  • Intelligent Exploration is Key: BO's ability to balance exploration and exploitation allows it to navigate complex, potentially multi-modal objective functions more effectively than traditional gradient-based methods.

Key Results

The proposed BO framework consistently outperformed the state-of-the-art HGO algorithm (a multi-step iterative optimization method) in two main aspects:

  • Efficiency: BO achieved lower objective function values (indicating better parameter fits) in significantly fewer invocations of the forward simulator. In some cases, BO achieved comparable or better results with 4-7 times fewer iterations.
  • Accuracy: The parameter estimates obtained through BO led to more realistic stress-stretch curves and LV geometry deformations, particularly when incorporating the Klotz-curve prior. The V30 emulator was also shown to be highly accurate.

Limitations and Future Work

  • Computational Cost Still a Factor: While significantly faster than HGO, BO still requires a substantial number of iterations (hundreds), which might be too slow for real-time clinical applications if starting from scratch for each patient. The authors suggest multi-task BO as a potential solution to leverage prior knowledge from previous optimizations.
  • Surrogate Choice is Case-Specific: The paper notes that the choice between target and partial error surrogates is not universally optimal and can depend on the specific LV geometry and data. They recommend trying both.
  • Future Work: The authors suggest extending the framework to account for:
  • Stiffness heterogeneity within the myocardium.
  • Interactions with active contraction and electrophysiology.
  • More detailed cardiac motion measurements (e.g., from 3D tagged CMR).
  • Incorporating unloaded residual-stressed geometry.
  • Applying the framework to other constitutive laws and inverse problems.

Summary updated

Abstract

We consider parameter inference in cardio-mechanic models of the left ventricle, in particular the one based on the Holtzapfel-Ogden (HO) constitutive law, using clinical in vivo data. The equations underlying these models do not admit closed form solutions and hence need to be solved numerically. These numerical procedures are computationally expensive making computational run times associated with numerical optimisation or sampling excessive for the uptake of the models in the clinical practice. To address this issue, we adopt the framework of Bayesian optimisation (BO), which is an efficient statistical technique of global optimisation. BO seeks the optimum of an unknown black-box function by sequentially training a statistical surrogate-model and using it to select the next query point by leveraging the associated exploration-exploitation trade-off. To guarantee that the estimates based on the in vivo data are realistic also for high-pressures, unobservable in vivo, we include a penalty term based on a previously published empirical law developed using ex vivo data. Two case studies based on real data demonstrate that the proposed BO procedure outperforms the state-of-the-art inference algorithm for the HO constitutive law.
BibTeX
@article{borowska2022bayesian,
  author = {Agnieszka Borowska and Hao Gao and Alan Lazarus and Dirk Husmeier},
  title = {{Bayesian optimisation for efficient parameter inference in a cardiac mechanics model of the left ventricle}},
  journal = {International Journal for Numerical Methods in Biomedical Engineering},
  year = {2022},
  doi = {10.1002/cnm.3593},
}
Figure