Physics-constrained deep active learning for spatiotemporal modeling of cardiac electrodynamics.

Area of Science:

• Computational Biology
• Biomedical Engineering
• Machine Learning

Background:

• Electrophysiological models using partial differential equations (PDEs) solved by finite element method (FEM) lack clinical data integration.
• Sparse cardiac signal measurements from catheterization challenge traditional machine learning for predictive modeling.
• Existing methods struggle to incorporate real-world clinical data for accurate cardiac disease diagnosis and treatment.

Purpose of the Study:

• To develop a physics-constrained deep active learning (P-DAL) framework for modeling spatiotemporal cardiac electrodynamics.
• To integrate physical laws of cardiac electrical wave propagation with deep learning for robust prediction from sparse sensor data.
• To introduce a novel active learning strategy for optimizing data collection locations on the heart surface.

Main Methods:

• Adapted a physics-constrained deep learning (P-DL) framework to integrate cardiac electrophysiology PDEs with deep learning.
• Developed an active learning criterion combining P-DL prediction uncertainty and space-filling design for informative data acquisition.
• Evaluated the framework on healthy and diseased heart systems using sparse sensor measurements.

Main Results:

• The P-DL approach significantly outperformed traditional spatiotemporal Gaussian process (STGP) models, reducing relative error by up to 48.3% (healthy) and 28.0% (diseased).
• The proposed P-DAL framework demonstrated superior performance over P-DL with space-filling design (P-DSL) and random data sampling (P-DRL).
• P-DAL achieved relative error reductions of 16.3% and 28.0% (healthy) and 11.1% and 21.2% (diseased) compared to P-DSL and P-DRL, respectively.

Conclusions:

• The physics-constrained deep learning approach offers a robust method for modeling cardiac electrical behavior from sparse clinical data.
• The novel active learning strategy effectively identifies informative locations for data collection, enhancing predictive model accuracy.
• This P-DAL framework holds significant potential for improving cardiac disease diagnosis and treatment design through accurate computational modeling.