LEADS Framework Creates Personalized Cardiac Digital Twins

This research presents LEADS (Learning Cardiac Electrophysiology Digital Twins Through Agentic Discovery of Hybrid Structure), a novel framework for creating personalized digital twins of cardiac electrophysiology. Unlike traditional methods that require manual expert design of hybrid physics-neural architectures, or other LLM-based approaches that may lack stability, LEADS employs an LLM agent to autonomously discover and refine these complex models. The framework encodes cardiac electrophysiology domain knowledge into a structured action space, enabling the LLM agent to iteratively select, combine, and refine hybrid model components. Parameter fitting is handled by gradient descent, ensuring that each candidate model is physically grounded, interpretable, and numerically stable. Validation on both synthetic and real cardiac data demonstrates that LEADS surpasses both human-designed hybrid models and other LLM-based modeling techniques. This advancement promises more accurate and personalized simulations for cardiac conditions.