LEADS: Learning Cardiac Electrophysiology Digital Twins

LEADS, a framework for building cardiac electrophysiology digital twins, was introduced in an arXiv paper (arXiv:2606.18154) submitted on 16 Jun 2026 by Ziqi Zhou, Yubo Ye, Sumeet Atul Vadhavka, Linwei Wang and Zhiqiang Tao. The 10-page paper with four figures frames cardiac domain knowledge as a structured action space and uses a large language model agent to discover hybrid physics-neural model structure, while gradient descent fits model parameters.

What is LEADS and how does it work?

LEADS formulates cardiac electrophysiology domain knowledge as a structured action space and leverages an LLM agent to search that space, selecting, combining and refining hybrid models; parameter fitting is handled by gradient descent. The agent follows an "iterative reasoning-and-action loop" to design candidate architectures that the authors constrain to be physically grounded, interpretable and numerically stable. In practice LEADS separates structure discovery (agent-driven) from parameter optimization (gradient descent), enabling open-ended architectural discovery while keeping numerical stability central to every candidate model.

How was LEADS evaluated and what were the results?

The authors validated LEADS on both synthetic and real cardiac electrophysiology data, showing it outperformed human-designed hybrid models and other LLM-based hybrid modeling. Specifically, the paper reports validation on synthetic datasets built from three ground-truth reaction models and on real cardiac EP data. The submission notes the method's comparative advantage without listing numerical metrics in the abstract; the full manuscript contains four figures that illustrate those experiments across the 10-page document.

Why does framing domain knowledge as a structured action space matter?

Framing domain knowledge as an action space gives the LLM agent explicit, composable choices to build hybrid models, which the authors argue is necessary because prior LLM-based methods lacked structural priors needed for stable cardiac simulations. This approach aims to reduce dependence on human experts to prescribe hybrid physics-neural architectures and to improve transfer across patients by enabling the agent to discover patient-appropriate structures rather than only fitting parameters to a fixed architecture.

Why it matters

Cardiac electrophysiology simulations require numerical stability and physical grounding to be useful in clinical or research settings. By making structure discovery an agentic search over a constrained action space, LEADS attempts to bridge large language model generalization with the structural priors that pragmatic simulations need. If the framework scales beyond the reported synthetic and real-data tests, it could shift how teams build personalized EP digital twins: less manual architecture design, more automated model-structure discovery grounded in physics.

What to watch

Look for the paper's full experimental details and figures to appear in the PDF and HTML versions on arXiv, and for code or datasets linked from the submission. The next concrete signals will be published numerical metrics in the manuscript figures and any released implementations that show how the agent's choices map to simulation stability on additional patient datasets.

Bibliographic note: the paper appears on arXiv as arXiv:2606.18154, submitted 16 Jun 2026, authored by Ziqi Zhou, Yubo Ye, Sumeet Atul Vadhavka, Linwei Wang and Zhiqiang Tao. The submission lists 10 pages and four figures.