Back to AI Research

AI Research

Learning Cardiac Electrophysiology Digital Twins Th... | AI Research

Key Takeaways

  • Building personalized cardiac electrophysiology (EP) digital twins—computational models that simulate a patient’s unique heart electrical activity—is essenti...
  • Building personalized cardiac electrophysiology (EP) digital twins requires identifying the appropriate model structure for each patient, not merely fitting parameters.
  • Traditional methods rely on experts to manually prescribe hybrid physics-neural architectures, which requires deep domain expertise and does not transfer across patients.
  • Recent works have applied large language models (LLMs) to generate or act as hybrid models.
  • However, despite their promising generalization capacity, these LLM-based methods lack the structural priors needed for stable cardiac simulations.
Paper AbstractExpand

Building personalized cardiac electrophysiology (EP) digital twins requires identifying the appropriate model structure for each patient, not merely fitting parameters. Traditional methods rely on experts to manually prescribe hybrid physics-neural architectures, which requires deep domain expertise and does not transfer across patients. Recent works have applied large language models (LLMs) to generate or act as hybrid models. However, despite their promising generalization capacity, these LLM-based methods lack the structural priors needed for stable cardiac simulations. Hence, we propose LEADS, a framework that formulates cardiac EP domain knowledge as a structured action space and utilizes an LLM agent to discover hybrid models. The agent follows an iterative reasoning-and-action loop to select, combine, and refine hybrid models, whilst gradient descent handles parameter fitting. The proposed LEADS designs every candidate model towards physically grounded, interpretable, and numerically stable, while allowing open-ended architectural discovery. We validate LEADS on synthetic data with three ground-truth reaction models and on real cardiac EP data, demonstrating that it outperforms both human-designed hybrid models and other LLM-based hybrid modeling.

Building personalized cardiac electrophysiology (EP) digital twins—computational models that simulate a patient’s unique heart electrical activity—is essential for clinical tasks like guiding heart surgery and assessing disease risk. However, creating these models is difficult because cardiac electrical propagation varies significantly between patients. Traditional methods either rely on experts to manually design complex models or use unconstrained AI that lacks the physical stability required for medical accuracy. This paper introduces LEADS, a framework that uses an AI agent to automatically discover and build these personalized hybrid models, combining the flexibility of neural networks with the reliability of established physics.

The Challenge of Personalized Modeling

Cardiac electrical activity is typically modeled as a reaction-diffusion system, where diffusion describes how electrical signals spread across heart tissue and reaction describes local cell activity. Because these processes differ from person to person, a model built for one patient often fails for another. While human experts can manually adjust these models, the process is labor-intensive and does not scale. Recent attempts to use Large Language Models (LLMs) to automate this process have struggled because unconstrained AI often generates models that violate physical laws, such as failing to maintain stable electrical signals or producing physiologically impossible results.

How LEADS Works

LEADS (Learning Electrophysiology hybrid model through Agentic Discovery of Structure) bridges the gap between human expertise and automated AI. Instead of letting an AI generate code from scratch, LEADS provides the LLM agent with a "structured action space." This space acts as a catalog of proven building blocks:

  • Reaction Catalog: A selection of established, physics-based ionic models.

  • Diffusion Catalog: A range of neural network architectures designed to handle signal propagation.
    The agent follows an iterative "Observe-Think-Act" loop. It reviews the performance of previous model candidates, reasons about how to improve them (such as fixing overfitting or increasing model capacity), and then takes specific actions—like swapping a reaction model, refining a neural layer, or simplifying the architecture. Throughout this process, the agent manages the structural design, while standard gradient descent handles the fine-tuning of model parameters.

Performance and Results

The researchers tested LEADS against both human-designed models and other LLM-based approaches using both synthetic data and real clinical recordings. In synthetic tests, LEADS successfully identified the correct underlying physics and outperformed human-designed hybrid models. When applied to real clinical data, LEADS produced smoother, more accurate activation maps that aligned better with ground-truth patterns than existing hybrid methods. These results suggest that by constraining an AI agent with domain-specific knowledge, it is possible to automate the creation of reliable, interpretable digital twins without requiring manual intervention from a domain expert.

Future Directions

While LEADS demonstrates significant promise in automating the design of cardiac digital twins, the current study is limited to specific heart mesh geometries and a fixed catalog of components. Future research aims to expand the framework to handle more diverse heart geometries, incorporate a wider variety of physiological components into the catalog, and explore the application of this agentic discovery approach to other areas of medical physiology.

Comments (0)

No comments yet

Be the first to share your thoughts!