A Neural Latent Dynamics Approach for Solving Inverse Problems in Cardiac Electrophysiology

Technical Summary: A Neural Latent Dynamics Approach for Solving Inverse Problems in Cardiac Electrophysiology

1. Problem Statement

The paper addresses the challenge of solving inverse problems in cardiac electrophysiology, specifically the non-invasive identification of cardiac electrical abnormalities (such as ectopic activation sites and ischemic regions) from surface electrocardiogram (ECG) measurements. This task is mathematically formulated as a PDE-constrained optimization problem where unknown physiological parameters must be inferred from indirect, spatially aggregated data.

The core difficulty lies in the severe ill-posedness of the problem, driven by the smoothing nature of the forward operator, limited measurement locations, and the intrinsic nonlinearity of cardiac models. Furthermore, standard approaches relying on high-fidelity (HF) computational models (e.g., Bidomain or Monodomain equations coupled with ionic models) are computationally prohibitive for iterative inversion. HF simulations often require minutes to hours per run, making them unsuitable for multi-query contexts like uncertainty quantification, parameter calibration, or real-time clinical decision support.

2. Methodology

The authors propose a data-driven framework that replaces the repeated evaluation of high-fidelity models with an efficient surrogate model based on Latent Dynamics Networks (LDNets). The methodology consists of three main components:

A. High-Fidelity Forward Model (Ground Truth Generation)
To generate training data, the authors solve the Monodomain equations coupled with the Ten Tusscher ionic model.

• Governing Equations: The system describes the transmembrane potential $v$, recovery-gating variables $w$, and ionic concentrations $c$.
• Observation Operator: Pseudo-ECG signals are computed using an infinite-volume conductor approximation, integrating the intracellular current density over the cardiac domain to determine potentials at body-surface lead locations.
• Numerical Solution: The forward model is solved using the Finite Element Method (FEM) with $Q1$ elements and a first-order implicit-explicit (IMEX) time-stepping scheme.

B. Surrogate Forward Architecture (LDNet)
The core innovation is a neural surrogate operator $\mathcal{S} \approx \mathcal{G}$ that maps low-dimensional parameters (e.g., stimulus coordinates or ischemic region descriptors) directly to pseudo-ECG signals.

• Latent Dynamics: The architecture learns a low-dimensional latent state $s(t; p) \in \mathbb{R}^{n_s}$ whose evolution is governed by a neural ordinary differential equation (Neural ODE): $\frac{ds}{dt} = \text{NN}_{\text{dyn}}(s, p)$.
• Reconstruction: A second neural network, $\text{NN}_{\text{rec}}$, maps the latent state to the observable space (ECG signals at $N_{\text{leads}}$).
• Training: The networks are trained offline on a dataset of HF simulations to minimize the mean squared error (MSE) between predicted and ground-truth ECGs. For the variable-radius ischemia case, a frequency-domain regularization term (based on the Discrete Fourier Transform) is added to the loss function to capture spectral features.

C. Inverse Problem Strategy
The inverse problem is solved by minimizing a discrepancy functional $J(p) = \text{MSE}(\mathcal{S}(p), \text{pECG}_{\text{obs}})$ using the trained surrogate.

• Initialization: To mitigate the non-convexity of the optimization landscape, a multi-start strategy is employed. The domain is partitioned into subdomains; the surrogate is evaluated at the center of each subdomain to find the candidate minimizing the initial data misfit.
• Optimization: The selected initial guess is refined using a gradient-based hybrid method combining Adam iterations and second-order quasi-Newton (BFGS) updates. For 3D cases, a projected gradient descent is used to ensure parameters remain within physical bounds.

3. Key Contributions

• LDNet Integration: The paper introduces the application of Latent Dynamics Networks to cardiac inverse problems, successfully decoupling the learning of temporal dynamics from the reconstruction of observable quantities.
• Computational Efficiency: The framework drastically reduces the computational cost of the forward operator evaluation during the inversion loop, enabling near real-time parameter estimation.
• Robust Initialization: The authors propose a specific multi-start initialization strategy tailored to the non-convex nature of the ECG inverse problem, demonstrating its necessity for 3D scenarios where identifiability issues are prevalent.
• Frequency-Domain Regularization: For the variable-radius ischemia problem, the inclusion of a spectral loss term improves the identifiability of parameters by capturing subtle oscillations in the ECG signals.

4. Numerical Results

The framework was validated on synthetic data across four test cases: 2D ectopic activation, 2D ischemia (fixed radius), 2D ischemia (variable radius), and 3D ectopic activation.

• Forward Surrogate Accuracy:
  • 2D Cases: The surrogate achieved normalized RMSE values on the order of $10^{-3}$ to $10^{-4}$ with Pearson dissimilarity below $4 \times 10^{-3}$. The optimal latent dimension was found to be 20–24 states.
  • 3D Case: Due to increased geometric complexity, errors were slightly higher (Normalized RMSE $\approx 10^{-2}$), but the model remained accurate with 20 latent states.
• Inverse Problem Performance:
  • Convergence: The inverse optimization consistently achieved minimum losses below $10^{-6}$ (2D) and $10^{-6}$ (3D).
  • Accuracy: The method successfully reconstructed stimulus locations and ischemic region centroids with high precision. In the 3D case, the multi-start strategy reduced failure rates associated with local minima.
  • Speed: The average optimization time per instance was approximately 80 seconds for 2D and 640 seconds for 3D cases. This represents a significant reduction compared to the hours required for HF-based inversion, making the approach compatible with near real-time applications.
  • Resource Usage: Memory usage remained within standard CPU/GPU limits (e.g., ~6–12 GB), and training times were manageable (10–110 hours depending on complexity).

5. Significance and Claims

The paper claims that this work represents a significant step toward practical, non-invasive, and data-driven solutions for cardiac inverse problems. By leveraging LDNets, the authors demonstrate that it is possible to:

1. Circumvent Computational Bottlenecks: Replace expensive PDE solves with fast neural evaluations without sacrificing the accuracy required for clinical parameter estimation.
2. Enable Multi-Query Applications: The reduced computational overhead makes the framework suitable for contexts requiring repeated evaluations, such as uncertainty quantification and optimal control.
3. Support Clinical Decision Making: The ability to perform robust parameter reconstruction in near real-time paves the way for applications in arrhythmia source localization and ischemia characterization.

The authors remain modest regarding the current scope, noting that the results are based on in-silico synthetic data. They state that future work will focus on extending the framework to realistic anatomical geometries, incorporating patient-specific data, and improving robustness against noise and model uncertainties before clinical deployment.