Abstract
Noninvasive reconstruction of cardiac electrical activity from surface electrocardiograms (ECG) involves solving an ill-posed inverse problem. Cardiac electrophysiological (EP) models have been used as important a priori knowledge to constrain this inverse problem. However, the reconstruction suffer from inaccuracy and uncertainty of the prior model itself which could be mitigated by estimating a priori model error. Unfortunately, due to the need to handle an additional large number of unknowns in a problem that already suffers from ill-posedness, model error estimation remains an unresolved challenge. In this paper, we address this issue by modeling and estimating the a priori model error in a low dimensional space using a novel sparse prior based on the variational approximation of L0 norm. This prior is used in a posterior regularized Bayesian formulation to quantify the error in a priori EP model during the reconstruction of transmural action potential from ECG data. Through synthetic and real-data experiments, we demonstrate the ability of the presented method to timely capture a priori model error and thus to improve reconstruction accuracy compared to approaches without model error correction.
Introduction
Noninvasive cardiac electrophysiological (EP) imaging aims to reconstruct the 3D electrical activity of the heart using only surface ECG signals. This task is a notoriously "ill-posed" inverse problem, as many different internal electrical states can produce the same surface ECG readings. To address this, researchers typically employ biophysical EP models as a "prior" to constrain the potential solutions.
However, these models are often idealized and may not reflect the specific pathology of a patient. For instance, a model might assume healthy myocardial tissue when the patient actually possesses scar tissue. This mismatch creates significant "model error." Estimating this error is traditionally difficult because it introduces a massive number of new unknowns into an already mathematically unstable problem. This paper introduces a novel Bayesian framework to improve noninvasive cardiac imaging by explicitly accounting for these errors.
Methodology
The authors propose a posterior regularized Bayesian approach that focuses on modeling the error ($\eta$) of the EP model. The key innovation lies in the representation of this error through three main components:
- Sparse Gradient Domain: The authors argue that because cardiac wavefronts are localized, the errors in predicting them should also be localized (sparse) in the spatial gradient domain. This allows for a reduction in the dimensionality of the error estimation problem.
- $L_0$ Norm Approximation: To enforce this sparsity, they utilize a generalized Gaussian distribution where the shape parameter $p$ is near zero, effectively approximating an $L_0$ norm (which counts the number of non-zero elements).
- Variational Inference: Since the resulting equations are analytically intractable, the authors derive a variational lower bound to make the optimization manageable. They use an Expectation-Maximization (EM) procedure to iteratively estimate the model error's covariance while simultaneously reconstructing the transmural action potential.
Experimental Results
The study evaluated the method using both synthetic simulations and real-world clinical data.
Synthetic Experiments
Synthetic tests were conducted using 17 different heart-torso models based on the AHA 17-segment model of the left ventricle. These simulations included various scar patterns and were corrupted with 20 dB of Gaussian noise to mimic real-world conditions. The proposed method significantly outperformed the baseline (which lacked error correction). While the baseline failed to capture late activation due to scars in 50% of cases, the new method consistently identified them. Quantitatively, the method showed statistically significant improvements ($p < 0.005$) in scar detection accuracy, achieving a Dice coefficient of 0.4 (vs. 0.28) and reducing the physical distance between predicted and actual scar centers to 9.4 mm (vs. 17.14 mm).
Clinical Validation
Two clinical case studies were performed on patients undergoing catheter ablation for scar-related ventricular arrhythmia. The noninvasive reconstructions were compared against in-vivo bipolar voltage maps obtained via invasive catheter mapping (the gold standard). In these cases, the method successfully identified scar areas at both the anterior and inferior base of the heart, matching the invasive mapping results. Notably, the standard approach missed the anterior scar entirely in Case 1.
Discussion
A major insight from this work is that model inaccuracy should not be treated as random Gaussian noise, but rather as a structured signal. By recognizing that errors occur primarily at the "wavefront" of electrical activity, the dimensionality of the problem can be effectively reduced. Furthermore, the algorithm produces a "spatial map of variance." High variance regions effectively point to where the EP model is failing, which in this study correlated strongly with the location of physical scars, making variance a powerful diagnostic tool.
The results also highlight the importance of "Timely Correction." The proposed method corrects the model error early in the heart's activation cycle, preventing errors from compounding as the simulated electrical wave propagates through the heart.
Conclusion
The authors have demonstrated a robust method for estimating a priori model error in cardiac EP imaging. While the current approach significantly improves reconstruction accuracy, future work will focus on "closing the loop." The next logical step is to use the estimated error variance to actively update and correct the parameters of the EP model itself in real-time, moving toward self-correcting physiological models that adapt to specific patient anatomies during the imaging process.
Summary updated