A self-taught artificial agent for multi-physics computational model personalization.

Area of Science:

• Computational modeling
• Artificial intelligence
• Biomedical engineering

Background:

• Personalizing multi-physics computational models to patient data is crucial for clinical applications.
• Current personalization methods are often manual, time-consuming, and model/data-specific.

Purpose of the Study:

• To develop a generalizable and automated approach for model personalization using artificial intelligence.
• To create an AI agent capable of learning optimal personalization strategies.

Main Methods:

• Reformulated personalization as a reinforcement learning problem.
• Developed 'Vito,' a self-taught artificial agent that learns a decision process model through computational model exploration.
• Tested the model-independent algorithm on synthetic data, cardiac electrophysiology, and a whole-body circulation model.

Main Results:

• Vito learned to optimize cost functions generically in synthetic scenarios.
• Achieved comparable or better goodness of fit than standard methods in cardiac and circulation models.
• Demonstrated increased robustness (up to 11% higher success rates) and faster convergence (up to seven times).

Conclusions:

• The AI-driven approach makes personalization algorithms generalizable and self-adaptable.
• This method can be applied to any patient and any computational model with minimal adjustments.
• Vito offers a promising solution for efficient and effective clinical model personalization.