Physics-Informed Neural Operators for Cardiac Electrophysiology

1. Problem Statement

Simulating cardiac electrophysiology (EP) is critical for understanding arrhythmias and planning treatments like catheter ablation. However, current methods face significant trade-offs:

• Traditional Numerical Solvers: (e.g., Finite Element Methods) are computationally expensive, sensitive to mesh resolution, and slow for real-time applications.
• Standard Deep Learning (PINNs): While Physics-Informed Neural Networks (PINNs) incorporate physical laws to reduce data requirements, they are limited by:
  • Mesh Dependency: They are trained on fixed mesh resolutions and struggle to generalize to different discretizations.
  • Instance Specificity: They typically learn a single PDE instance, requiring retraining for new initial conditions or parameters.
  • Long-term Stability: They often fail to maintain accuracy over long time horizons, especially in chaotic systems.

The authors aim to overcome these limitations by developing a method that is resolution-invariant, generalizable across initial conditions, and capable of efficient long-term prediction for cardiac EP.

2. Methodology

The proposed solution is a Physics-Informed Neural Operator (PINO) architecture, specifically utilizing a Fourier Neural Operator (FNO) backbone.

A. Mathematical Formulation

The cardiac EP system is modeled using the Aliev-Panfilov cell model, a reaction-diffusion system described by two coupled PDEs governing the transmembrane voltage ($V$) and a recovery variable ($W$):
$$\frac{\partial V}{\partial t} = \nabla \cdot (D\nabla V) - kV(V-a)(V-1) - VW$$
$$\frac{\partial W}{\partial t} = \left(\epsilon + \frac{\mu_1 W}{V + \mu_2}\right)[-W - kV(V-a-1)]$$

B. Architecture: Fourier Neural Operator (FNO)

Instead of learning a function approximation for a specific grid, the FNO learns a mapping (operator) between infinite-dimensional function spaces (Initial Conditions $\to$ Solution).

• Lifting & Projection: The input function is lifted to a higher dimension, processed through layers, and projected back.
• Fourier Kernel: The core of the FNO uses a Fourier convolution operator ($K_l$) defined in the frequency domain:
  $$(Kv_l)(x) = \mathcal{F}^{-1}(R \cdot (\mathcal{F}v_l))(x)$$
  This allows the model to capture global features and learn dynamics independent of the spatial discretization (mesh resolution).

C. Physics-Informed Learning

To enforce physical consistency without relying solely on massive datasets, the authors combine a data-driven loss with a physics-based loss:

1. Data Loss ($L_{data}$): Minimizes the difference between the predicted solution and ground truth (where available).
2. Physics Loss ($L_{phys}$): Penalizes deviations from the PDE residuals, boundary conditions (Neumann/No-flux), and initial conditions.
   • The PDE residual is approximated using central difference methods for temporal derivatives.
   • The total loss is a weighted sum: $L = \lambda_{data}L_{data} + \lambda_{phys}L_{phys}$.
   • The authors experimented with Fixed, Soft Adaptation, and Relative Aggregation strategies for weighting these losses.

D. Training Strategy

• Two-Stage Training:
  1. Stage 1: Train using only data loss to learn the general dynamics.
  2. Stage 2: Fine-tune using both data and physics losses.
• Input Formulations: The model was tested in Single Frame (predicting $n$ steps ahead from one frame) and Multi-Frame (predicting trajectories from a sequence of frames) modes.

3. Key Contributions

1. First PINO for Cardiac EP: This is the first application of Physics-Informed Neural Operators to cardiac electrophysiology, moving beyond the limitations of PINNs.
2. Resolution Invariance: The model can be trained on a coarse mesh and infer on significantly finer meshes (up to 10× the training resolution) with minimal accuracy loss (<8% degradation).
3. Zero-Shot Generalization: The model can predict complex arrhythmic scenarios (e.g., spiral waves) when trained only on simple scenarios (e.g., planar waves), demonstrating the ability to learn underlying system dynamics rather than just pattern matching.
4. Long-Horizon Stability: The PINO model maintains high predictive quality in recursive roll-out scenarios (where predictions are fed back as inputs), a task where standard PINNs and pure data-driven models typically fail or diverge.
5. Efficiency: The inference time for a full 2500ms trajectory is approximately 1 second, compared to ~40 seconds for traditional solvers on similar hardware.

4. Experimental Results

The authors evaluated the model on four propagation scenarios: Planar, Centrifugal, Spiral, and Spiral-Breakup (chaotic).

• Accuracy vs. Baselines:
  • PINO vs. FNO (Data-only): PINO significantly outperformed the pure data-driven FNO, especially in Roll-Out (sequential prediction) scenarios. For example, in the Spiral scenario, PINO reduced RMSE from 0.1682 (FNO) to 0.0173.
  • PINO vs. PINN: PINO models achieved lower error rates than PINNs with comparable or smaller training budgets, highlighting the efficiency of operator learning.
• Mesh Resolution Scaling:
  • Training on a mesh downsampled by 10× and testing on the original high-resolution mesh resulted in an accuracy drop of less than 8%, proving the model's ability to generalize across discretizations.
• Zero-Shot Transfer:
  • Models trained exclusively on "Planar" or "Centrifugal" data could accurately predict the qualitative dynamics of "Spiral" waves. While voltage amplitude scaling required adjustment, the propagation patterns were faithfully reproduced.
• Long-Term Prediction:
  • The model successfully simulated trajectories up to 2500ms (2.5x the training horizon) with stable performance, whereas chaotic systems often cause standard neural networks to diverge.

5. Significance and Future Work

Significance:
This work demonstrates that PINOs are a viable, efficient alternative to traditional numerical solvers for cardiac modeling. Their ability to handle variable resolutions, unseen initial conditions, and long-term recursive prediction makes them ideal for:

• Digital Twins: Creating patient-specific, real-time cardiac models for treatment planning.
• Medical Devices: Enabling onboard, low-latency simulation for Implantable Cardioverter Defibrillators (ICDs).
• Clinical Workflow: Drastically reducing the time required for in-silico trials and arrhythmia analysis.

Future Directions:
The authors plan to extend the framework to:

• 3D and Irregular Meshes: To reflect realistic cardiac anatomy using geometry-aware operator layers.
• Broader Cell Models: Training a general operator across multiple cardiac cell models to handle a wider range of pathologies.
• Stability Enhancements: Improving long-horizon stability for highly chaotic regimes through advanced recurrent operator layers.

In conclusion, the paper establishes PINOs as a powerful tool for bridging the gap between the accuracy of physics-based simulations and the speed of deep learning, specifically tailored for the complex, chaotic dynamics of cardiac electrophysiology.