Physics-informed residual learning with spatiotemporal local support for inverse ECG reconstruction

Area of Science:

• Computational physics
• Biomedical engineering
• Machine learning

Background:

• Physics-informed neural networks (PINNs) integrate physical laws into neural networks for complex system modeling.
• The inverse electrocardiographic imaging (ECGI) problem reconstructs cardiac electrical activity from body surface potentials.
• Existing PINN models for ECGI struggle with overfitting, training instability, and scalability for spatiotemporal data.

Purpose of the Study:

• To develop an advanced PINN framework for robust and accurate ECGI reconstruction.
• To overcome limitations of current PINN approaches in handling high-dimensional spatiotemporal data.
• To enhance the modeling of complex dynamical systems with data constraints.

Main Methods:

• Proposed a novel physics-informed residual learning framework with spatiotemporal local support.
• Introduced a numerical differentiation scheme using local neighborhood information for derivative approximation.
• Developed an adaptive residual network architecture with trainable skip connections for stable optimization and expressiveness.

Main Results:

• The proposed method significantly outperforms traditional regularization methods and prior PINN models in ECGI reconstruction.
• Achieved higher accuracy and improved robustness against sensor noise in simulated body-heart geometries.
• Demonstrated effective spatiotemporal constraint enforcement and enhanced model expressiveness.

Conclusions:

• The novel framework advances PINN capabilities for the ECGI problem.
• Offers a more stable and scalable solution for high-dimensional spatiotemporal modeling.
• Provides a methodological foundation with broader implications for data-constrained modeling in complex systems.