A Parameter Optimization to Solve the Inverse Problem in Electrocardiography

Gwladys Ravon, Rémi Dubois, Yves Coudière, Mark PotseFimh 2017 · DOI

TL;DR

This paper introduces a novel parameter optimization approach to reconstruct both the activation and repolarization phases of the heart's electrical activity from body surface electrocardiography (ECG) signals. By modeling the action potential as a parameterized function, they bypass traditional regularization methods and achieve more accurate and continuous maps of electrical events, particularly for repolarization.

What Problem Does This Paper Address?

Electrocardiography (ECG) aims to understand the heart's electrical activity by measuring potentials on the body's surface. A major challenge is the "inverse problem": inferring the electrical activity on the heart's surface (epicardium) from these body surface measurements. Traditional methods often focus on reconstructing activation maps, which describe when different parts of the heart depolarize (activate). However, accurately capturing the repolarization phase (when the heart resets electrically) and obtaining detailed information about the shape of individual cardiac action potentials (APs) from body surface ECGs remains difficult. Existing methods often struggle with signal amplitude, introduce artificial discontinuities in the electrical maps, or are not designed to reconstruct repolarization information.

Methods

The core of this paper is a parameter identification (PI) problem formulation. Instead of directly solving the inverse problem with regularization, the authors model the cardiac action potential (AP) using a parameterized function with four key parameters: amplitude (A), activation time ($\tau$), plateau phase duration ($T_R$), and repolarization slope ($\alpha_R$). These parameters can vary across the epicardium.

The process involves:

  1. Forward Model: Using the parameterized APs, they compute the extracellular potentials on the torso surface by solving the Laplace equation.
  2. Parameter Identification: The inverse problem is then reframed as finding the set of parameters (A, $\tau$, $T_R$, $\alpha_R$) that best matches the computed torso potentials to the measured body surface potentials (BSPM). This is achieved by minimizing a least-squares error cost function.
  3. Optimization Strategy: To ensure convergence and avoid overfitting, they employ a gradient descent method. Crucially, they split the identification into two phases:
  4. Depolarization Phase: Focuses on identifying the amplitude (A) and activation times ($\tau$) using a time window covering the QRS complex. They use a standard Method of Fundamental Solutions (MFS) with regularization to get initial estimates for ATs.
  5. Repolarization Phase: Focuses on identifying the plateau phase duration ($T_R$) and repolarization slope ($\alpha_R$) using a time window covering the T wave.
  6. Simplifications: For better convergence, they initially assume constant amplitude and repolarization slopes across the entire epicardium.

Datasets

The authors evaluated their method using two types of data:

  • In Silico Data: Generated from a realistic 3D anatomical model of the torso with a monodomain reaction-diffusion model and an anisotropic heart model. This provided ground truth epicardial potentials and activation/repolarization times.
  • Experimental Data: Obtained from a Langendorff-perfused pig heart, with simultaneous epicardial and body surface potential recordings.

Key Insights and Ideas

  • Parameterization is Key: By parameterizing the AP shape, the method inherently incorporates prior knowledge about cardiac electrophysiology, reducing the need for explicit regularization and allowing for the reconstruction of both depolarization and repolarization.
  • Continuous Maps: The parameter optimization approach naturally leads to smoother, more continuous activation and repolarization maps compared to some traditional MFS methods, which can exhibit artificial discontinuities.
  • Repolarization Focus: The method is specifically designed to provide better information on the repolarization phase, which is often a challenge for standard inverse ECG problems.
  • Action Potential Properties: The parameterization allows for the retrieval of local AP properties like APD90 (Action Potential Duration at 90% repolarization), offering deeper insights into regional cardiac function.

Key Results

  • Activation Times (ATs): The PI method achieved satisfactory ATs, with correlation coefficients (CC) comparable to or slightly better than MFS (e.g., CC of 0.85 vs. 0.82 for LV pacing in experimental data). Crucially, it avoided artificial discontinuities observed in MFS maps.
  • Repolarization Times: The PI method showed good accuracy for repolarization times, with CCs of 0.9 for LV pacing and 0.71 for RV pacing in silico data. While the reconstructed repolarization phase was sometimes shorter than the reference, the distribution of points was smoother than with MFS.
  • APD90: The method provided a larger range of APD90 values, reflecting more detailed regional differences, though some differences in APD90 repartition between LV and RV pacing were not perfectly captured.
  • Torso Potentials: Reconstructed torso potentials were close to measured ones, with the optimized amplitude A* effectively fitting BSPM amplitudes. Both depolarization and repolarization phases were well identified on torso electrodes.

Limitations and Future Work

  • Constant Parameters: The current method assumes constant amplitude and repolarization slopes across the entire epicardium. The authors acknowledge that this is a simplification and suggest future work to consider piecewise constant or spatially varying parameters.
  • Plateau Phase Accuracy: While repolarization times were fitted accurately, an error in activation time ($\tau$) could lead to an error in plateau phase duration ($T_R$) to compensate and fit the repolarization phase.
  • Geometry: To improve results, the authors plan to incorporate more detailed geometry, such as the septum or endocardium.
  • Pacing Site Localization: Similar to MFS, precisely localizing pacing sites remained a challenge.
  • T Wave Inversions: While not observed in their tested cases, the paper doesn't explicitly detail how the method would handle T wave inversions.

The authors propose future developments including incorporating more detailed anatomical models and exploring more flexible parameterizations for amplitude and repolarization slopes.

Summary updated

BibTeX
@inproceedings{ravon2017parameter,
  author = {Gwladys Ravon and Rémi Dubois and Yves Coudière and Mark Potse},
  title = {{A Parameter Optimization to Solve the Inverse Problem in Electrocardiography}},
  booktitle = {Fimh},
  year = {2017},
  doi = {10.1007/978-3-319-59448-4\_21},
}

Loading related content...

Figure