Whole-heart electromechanical simulations using Latent Neural Ordinary Differential Equations.

Area of Science:

• Computational biology
• Biomedical engineering
• Artificial intelligence in medicine

Background:

• High-fidelity cardiac digital twins are crucial for personalized medicine but computationally expensive.
• Current multi-scale cardiac models present a barrier to widespread clinical adoption due to significant computational demands.

Purpose of the Study:

• To develop a computationally efficient method for creating accurate whole-heart digital twins.
• To enable faster simulations for personalized cardiac modeling using artificial intelligence.

Main Methods:

• Utilized Latent Neural Ordinary Differential Equations (LNODEs) to model cardiac pressure-volume dynamics.
• Trained a surrogate model using 400 simulations, incorporating 43 parameters of cardiac electromechanics and hemodynamics.
• Developed a compact Artificial Neural Network representation with 3 hidden layers and 13 neurons per layer.

Main Results:

• Achieved a compact latent space representation of a 3D-0D cardiac model.
• Enabled numerical simulations of cardiac function on a single processor, drastically reducing computational time.
• Successfully performed global sensitivity analysis and parameter estimation with uncertainty quantification within 3 hours on a single processor.

Conclusions:

• Latent Neural Ordinary Differential Equations offer a feasible approach to creating fast and accurate cardiac digital twins.
• This AI-driven method overcomes the computational barriers associated with traditional high-fidelity models.
• Facilitates personalized medicine applications, including efficient parameter estimation and sensitivity analysis for heart failure patients.