Non-Invasive Reconstruction of Cardiac Activation Dynamics Using Physics-Informed Neural Networks

This paper presents a physics-informed neural network framework that accurately reconstructs three-dimensional cardiac activation dynamics, active tension, deformation, and pressure from measurable deformation data by integrating nonlinear constitutive modeling and physical constraints, offering a promising pathway for non-invasive, patient-specific arrhythmia assessment.

The Gist

Here is an explanation of the paper, translated into everyday language with some creative analogies.

The Big Picture: Seeing the Invisible Heartbeat

Imagine your heart is a complex, rhythmic drum. When you have an arrhythmia (an irregular heartbeat), it's like the drummer is hitting the drum in the wrong order or at the wrong time. To fix this, doctors usually need to stick a probe inside the heart to "listen" to the rhythm. This is invasive, risky, and uncomfortable.

The goal of this research is to build a non-invasive "X-ray vision" for the heart's electrical rhythm. We want to see the electrical waves traveling through the heart muscle just by looking at how the heart squishes and stretches on the outside.

The Problem: The "Shadow" vs. The "Source"

Think of the heart's electrical signal as a light switch being flipped on. The light itself (the electricity) is invisible from the outside. However, when the light turns on, it causes a lampshade (the heart muscle) to vibrate and move.

• What doctors can see: The movement of the lampshade (mechanical deformation).
• What doctors need to know: Where and when the light switch was flipped (electrical activation).

The challenge is that the movement is a messy, delayed reaction to the electricity. It's like trying to guess exactly when a drummer hit the snare drum just by watching the dust motes dancing in the air caused by the sound waves. It's a difficult "inverse problem."

The Solution: The "Physics-Savvy" AI

The authors created a new type of Artificial Intelligence called a Physics-Informed Neural Network (PINN).

Usually, AI is like a student who memorizes thousands of flashcards. If you show it a new picture that isn't on a flashcard, it might guess wrong.

• Standard AI: "I've seen this pattern before, so I'll guess this."
• Physics-Informed AI (PINN): "I've seen this pattern, AND I know the laws of physics say this pattern must happen this way because of how materials stretch and squeeze."

In this study, the AI wasn't just guessing; it was forced to obey the laws of physics (specifically, how soft tissue stretches and how pressure works) while it learned.

How They Did It (The Recipe)

1. The Training Ground (The Simulation):
   Since they couldn't test this on real patients yet, they built a perfect, digital 3D model of a heart (an ellipsoid shape). They simulated a heart attack or arrhythmia in the computer, knowing exactly where the electrical wave started and how the heart moved. This was their "answer key."

2. The "Delta-PINN" Trick:
   They used a special version of AI called Delta-PINN.

   • Analogy: Imagine trying to describe the shape of a mountain to someone who has never seen one. If you just say "it's big," that's vague. But if you describe it using a specific set of musical notes (eigenfunctions) that naturally fit the shape of a mountain, the description becomes much clearer. The AI used these "musical notes" of the heart's shape to understand the geometry much better than standard AI.

3. The "Weak" Math:
   Instead of trying to solve complex equations perfectly at every single point (which is hard and slow), they used a "weak formulation."

   • Analogy: Instead of checking if every single brick in a wall is perfectly straight, you check the overall stability of the wall. If the wall doesn't wobble, the bricks are probably fine. This made the AI much faster and more stable.

The Results: How Well Did It Work?

They tested the AI under three difficult conditions:

1. Perfect Data: The AI worked like a charm. It could look at the squishy movement of the heart and perfectly reconstruct the invisible electrical wave, even figuring out the pressure inside the heart.
2. Noisy Data: They added "static" (noise) to the data, like a bad radio signal.
   • Result: The AI was very robust. Even with a lot of noise, it still found the main path of the electrical wave. It was like hearing a song through a stormy window; you might miss the high notes, but you still know the melody.
3. Blurry Data: They reduced the resolution, making the data look like a low-pixel, blurry photo.
   • Result: The AI could still see the big picture (where the wave started and where it went), but it missed some tiny details. It's like looking at a map from a plane: you can see the highway, but you can't see the individual cars.

Why This Matters

This is a huge step toward patient-specific, non-invasive heart care.

• No more probes: In the future, a doctor might just use an ultrasound machine (which is safe, cheap, and radiation-free) to record the heart's movement.
• The AI does the rest: The AI would instantly calculate the electrical map, showing exactly where the arrhythmia is coming from.
• Trustworthy: Because the AI is forced to follow the laws of physics, it won't "hallucinate" impossible results. It's a trustworthy tool.

The Bottom Line

The authors built a smart computer program that acts like a detective. It looks at the "footprints" (heart movement) left behind by the "criminal" (electrical arrhythmia) and uses the laws of physics to deduce exactly where the criminal was and what they did. While it still needs more testing on real humans, this technology promises to make diagnosing dangerous heart rhythms safer, cheaper, and easier for everyone.

Technical Summary

Here is a detailed technical summary of the paper "Non-Invasive Reconstruction of Cardiac Activation Dynamics Using Physics-Informed Neural Networks."

1. Problem Statement

Cardiac arrhythmias are driven by complex electrical disturbances that are difficult to visualize non-invasively. Current clinical standards rely on invasive catheter mapping, which is limited to surface measurements and cannot capture the full 3D nature of electrical activation. While non-invasive techniques like 3D echocardiography can measure mechanical deformation (strain), there is a gap in reconstructing the source of this deformation: the active tension generated by myofibers, which serves as a proxy for the underlying electrical activation wave.

The core challenge is an inverse problem: inferring the spatiotemporal distribution of active tension ($T_a$) and hydrostatic pressure ($p$) from noisy, low-resolution, and sparse mechanical deformation data ($U$). Traditional data assimilation methods are computationally expensive, while purely data-driven machine learning models lack physical interpretability and may violate physical laws.

2. Methodology

The authors propose a Physics-Informed Neural Network (PINN) framework, specifically utilizing the Delta-PINN architecture, to solve this inverse problem.

A. Physics and Forward Modeling

• Governing Equations: The system is modeled using non-linear solid mechanics for an incompressible, hyperelastic material. The equilibrium equation is $\nabla \cdot P = 0$, subject to the incompressibility constraint $J=1$.
• Constitutive Model:
  • Passive Stress: Modeled using the Guccione model, an anisotropic hyperelastic energy density function dependent on fiber orientation.
  • Active Stress: Modeled as a force acting only along the local fiber direction ($f_0$), proportional to the active tension $T_a$.
  • Incompressibility: Enforced via a Lagrange multiplier field $p$ (identified as hydrostatic pressure).
• Data Generation: Synthetic datasets were generated using the simcardems software (FEniCS-based). A simplified 3D left ventricle (LV) ellipsoid geometry was used with rule-based fiber orientations (varying from $+60^\circ$ to $-60^\circ$ transmurally). The electrical wave was simulated using the monodomain formulation coupled with the O'Hara-Rudy and Land cell models.

B. The Delta-PINN Framework

Unlike standard PINNs that use raw spatial coordinates, this framework employs Laplacian eigenfunctions as the spatial input space.

• Input Representation: The neural networks (NNs) take the first 200 Laplacian eigenfunctions of the domain and the time coordinate $t$ as inputs. This spectral basis helps capture complex spatial patterns and facilitates learning in 3D domains.
• Output Fields: Three separate fully connected NNs predict:
  1. Deformation vector $U(r, t)$.
  2. Active tension $T_a(r, t)$ (constrained to be positive via sigmoid).
  3. Hydrostatic pressure $p(r, t)$.
• Loss Function: The total loss $L = L_{data} + \alpha L_{phys}$ consists of:
  • Data Loss ($L_{data}$): Mean squared error between predicted and observed deformation data.
  • Physics Loss ($L_{phys}$): Based on the weak formulation (Galerkin method) of the momentum balance equation. Instead of enforcing PDEs at collocation points via automatic differentiation, the residual is calculated using Finite Element (FE) integration over the mesh. This allows for:
    • Hard enforcement of Dirichlet boundary conditions.
    • Use of Gaussian quadrature to handle the non-linearities (exponential Guccione term, cubic active stress).
    • Optimization of nodal values rather than continuous functions, reducing the search space.

C. Training Strategy

• Two-Phase Optimization: The network is first trained for 2000 iterations using only the data loss ($\alpha=0$) to prevent initial physics loss explosions. Then, both terms are optimized for 60,000 iterations using the Adam optimizer.
• Robustness Testing: The method was tested on datasets corrupted with Gaussian noise (5% and 10% of max deformation) and reduced spatial resolution (50% and 25% of data points via smoothing and subsampling).

3. Key Contributions

1. Extension to 3D Non-Linear Mechanics: Moving beyond previous 2D linear studies, this work applies PINNs to a full 3D, non-linear, anisotropic cardiac mechanics problem.
2. Delta-PINN for Cardiac Inverse Problems: Demonstrates the efficacy of Laplacian eigenfunction inputs combined with FE-based weak formulations for reconstructing activation waves in complex geometries.
3. Simultaneous Estimation: The framework successfully recovers not just the active tension (activation proxy) but also the hydrostatic pressure field and deformation, satisfying the incompressibility constraint.
4. Robustness to Imperfect Data: The method is shown to be resilient to significant measurement noise and reduced spatial resolution, which are common in clinical ultrasound imaging.

4. Results

• Reference Case: The PINN accurately reconstructed the active tension wave propagation and Local Activation Time (LAT) maps, matching the ground truth with a Root Mean Square Error (RMSE) of 0.333 kPa for $T_a$. The LAT maps correctly identified the source and propagation direction.
• Noise Robustness: With 5% and 10% Gaussian noise, the global propagation patterns and LAT maps remained accurate. The RMSE for $T_a$ increased slightly (to ~0.388 kPa at 10% noise), but the method preserved the wavefront structure.
• Resolution Robustness:
  • At 50% resolution, accuracy remained high.
  • At 25% resolution, the global propagation patterns were still preserved, though fine spatial details (e.g., transmural variations) degraded. The RMSE for $T_a$ rose to 0.491 kPa, and pressure estimation became less accurate due to the loss of transmural information.
• Limitations: The reconstruction of transmural variations (endo- vs. epicardial differences) was partially limited by the coarse mesh (only 2–3 elements across the wall thickness) and the difficulty in satisfying the inf-sup condition for mixed FE formulations in the pressure field.

5. Significance and Future Outlook

• Clinical Potential: This approach offers a pathway toward patient-specific, non-invasive reconstruction of cardiac electrical activation using standard ultrasound data, potentially replacing invasive mapping for certain diagnostics.
• Trustworthy AI: By embedding physical laws (mechanics, incompressibility) directly into the loss function, the method avoids "hallucinations" common in purely data-driven models, ensuring solutions are physically plausible.
• Future Directions:
  • Parameter Estimation: Simultaneously estimating material parameters (e.g., stiffness) alongside activation.
  • Direct Potential Reconstruction: Moving from active tension proxies to direct transmembrane potential reconstruction.
  • Real-World Validation: Testing on synthetic data generated from full pipelines including echocardiographic noise and on actual patient data (ultrasound + MRI validation).
  • Temporal Regularization: Incorporating fixed temporal schemes to further constrain the solution space.

In conclusion, the paper establishes a robust computational framework that bridges the gap between non-invasive mechanical imaging and the underlying electrical dynamics of the heart, paving the way for advanced digital phenotyping and arrhythmia assessment.