Uncertainty and variability in models of the cardiac action potential: Can we build trustworthy models?

Ross H. Johnstone, Eugene T. Y. Chang, Rémi Bardenet, Teun P. de Boer, David J. Gavaghan, Pras Pathmanathan, Richard H. Clayton, Gary R. MiramsJournal of Molecular and Cellular Cardiology 2015 · DOI

TL;DR

This paper tackles the critical issue of uncertainty in cardiac action potential (AP) models, which currently provide single-point predictions despite inherent biological variability. By applying uncertainty quantification (UQ) techniques, the authors demonstrate how to infer probabilistic distributions for model parameters from experimental data and, crucially, how to propagate this uncertainty to predict model outputs. This approach promises to build more trustworthy models that can better reflect real-world biological systems and reveal the underlying causes of cardiac variability.

What Problem Does This Paper Address?

Cardiac electrophysiology models are powerful tools for understanding how heart cells generate electrical signals (action potentials). However, these models have historically been built using single, fixed values for parameters like ion channel conductances. This approach fails to account for the inherent variability observed in real biological systems, where ion channel properties can differ significantly between cells and even from beat to beat within the same cell. This lack of uncertainty representation means that model predictions are often treated as absolute truths, which can be misleading and limit their reliability for critical applications like drug safety testing or personalized medicine. The paper aims to bridge this gap by introducing and applying UQ methods to cardiac AP models, enabling them to represent and propagate uncertainty, leading to more robust and trustworthy predictions.

Methods

The paper presents two key case studies demonstrating UQ in cardiac AP modeling:

  1. Bayesian Inference for Parameter Estimation: This method uses experimental voltage recordings (action potentials) to infer probability distributions for model parameters, such as maximal ion channel conductances. Instead of finding a single "best" value, it quantifies the range of plausible values given the data and the model. This is achieved using Markov Chain Monte Carlo (MCMC) methods, which explore the parameter space to find distributions that best explain the observed data. The authors also investigate how different experimental protocols (e.g., pacing rates, drug application) influence the ability to constrain these parameters.
  2. Gaussian Process (GP) Emulators for Uncertainty Propagation: Once parameters are represented as probability distributions, the next step is to understand how this uncertainty affects model outputs (e.g., action potential duration). Running the full cardiac model many times with different parameter combinations (Monte Carlo) is computationally expensive. GP emulators act as fast, statistical surrogates for the complex cardiac model. They are trained on a limited number of model runs and can then rapidly predict model outputs and, importantly, quantify the uncertainty in these predictions. This allows for efficient sensitivity analysis (identifying which parameters most influence the output) and uncertainty propagation.

Datasets

  • Synthetic Data: Generated from existing cardiac AP models (e.g., Beeler & Reuter, ten Tusscher et al.) with added noise to simulate experimental conditions. This allows for controlled testing of the inference methods.
  • Experimental Canine Cardiomyocyte AP Recordings: Action potentials recorded from isolated canine cardiomyocytes were used to validate the inference methods and to investigate beat-to-beat variability.

Key Insights and Ideas

  • Moving Beyond Point Predictions: The core idea is that "trustworthy" models must acknowledge and quantify uncertainty. Treating model parameters as probability distributions, rather than single values, is essential for this.
  • Experimental Design Matters: The choice of experimental protocol significantly impacts which model parameters can be reliably inferred. Some parameters are more constrained by certain types of data than others.
  • Uncovering Sources of Variability: By inferring parameter distributions from individual AP traces, the authors can identify which ion currents are most likely responsible for beat-to-beat variability within a single cell. This offers a novel way to dissect the mechanisms of biological variability.
  • Emulators as Powerful Tools: GP emulators dramatically reduce the computational cost of exploring uncertainty and performing sensitivity analysis, making UQ more accessible for complex models.
  • UQ is a Foundation for Trust: Quantifying uncertainty is a crucial first step towards building confidence in computational models, especially for safety-critical applications.

Key Results

  • Successful Parameter Inference: The Bayesian inference method, using MCMC, was able to successfully infer many ion channel conductances from both synthetic and experimental data, providing probabilistic distributions rather than single values.
  • Protocol Impact: For simpler models, a single AP recording was sufficient to infer most parameters. For more complex models, longer or more informative protocols (e.g., drug block simulations) were necessary to constrain a larger number of parameters.
  • Beat-to-Beat Variability Insights: Analysis of experimental APs revealed that currents like GKr and GNa showed the widest spread in inferred distributions, suggesting they are key contributors to beat-to-beat variability in canine cardiomyocytes.
  • Efficient Uncertainty Propagation: GP emulators allowed for rapid estimation of output uncertainty (e.g., APD90) based on input parameter uncertainty, demonstrating that reducing uncertainty in sensitive parameters can significantly reduce output uncertainty. For example, reducing the variance of GNa reduced the standard deviation of Max. dVm/dt by over half.
  • Model Fit Improvement: Inferred conductances generally led to a better fit of the model to experimental APs compared to using default parameters.

Limitations and Future Work

  • Model Complexity and Unidentifiability: For highly complex models, even with informative protocols, some parameters may remain unidentifiable (multiple parameter sets fitting the data equally well).
  • Emulator Assumptions: GP emulators assume a smooth output surface, which might not hold for models exhibiting complex behaviors like alternans.
  • Interpreting Variability: While the study identified currents likely responsible for beat-to-beat variability, further work is needed to distinguish between stochastic gating, regulatory mechanisms, and other sources.
  • Spatial and Structural Uncertainty: The paper primarily focuses on parameter and condition uncertainty. Future work should address spatial variability (e.g., across different regions of the heart) and structural uncertainty (differences between the model and reality).
  • Integration of UQ Stages: Combining the parameter inference from Case Study 1 with the emulator-based uncertainty propagation from Case Study 2 is a natural next step, provided the inferred distributions can be approximated by the emulator's assumptions.
  • Broader Applications: The authors emphasize the need for UQ in other sub-models of cardiac APs and for developing more robust models for clinical applications and drug safety.

Summary updated

Abstract

Cardiac electrophysiology models have been developed for over 50years, and now include detailed descriptions of individual ion currents and sub-cellular calcium handling. It is commonly accepted that there are many uncertainties in these systems, with quantities such as ion channel kinetics or expression levels being difficult to measure or variable between samples. Until recently, the original approach of describing model parameters using single values has been retained, and consequently the majority of mathematical models in use today provide point predictions, with no associated uncertainty. In recent years, statistical techniques have been developed and applied in many scientific areas to capture uncertainties in the quantities that determine model behaviour, and to provide a distribution of predictions which accounts for this uncertainty. In this paper we discuss this concept, which is termed uncertainty quantification, and consider how it might be applied to cardiac electrophysiology models. We present two case studies in which probability distributions, instead of individual numbers, are inferred from data to describe quantities such as maximal current densities. Then we show how these probabilistic representations of model parameters enable probabilities to be placed on predicted behaviours. We demonstrate how changes in these probability distributions across data sets offer insight into which currents cause beat-to-beat variability in canine APs. We conclude with a discussion of the challenges that this approach entails, and how it provides opportunities to improve our understanding of electrophysiology.
BibTeX
@article{johnstone2015uncertainty,
  author = {Ross H. Johnstone and Eugene T. Y. Chang and Rémi Bardenet and Teun P. de Boer and David J. Gavaghan and Pras Pathmanathan and Richard H. Clayton and Gary R. Mirams},
  title = {{Uncertainty and variability in models of the cardiac action potential: Can we build trustworthy models?}},
  journal = {Journal of Molecular and Cellular Cardiology},
  year = {2015},
  doi = {10.1016/j.yjmcc.2015.11.018},
}
Figure