M.Sc. Stephanie Appel

Motivation
Computer models of cardiac excitation propagation help to understand and treat heart rhythm disorders. The conduction velocity (CV) of these waves depends on local conditions such as the tissue recovery state, wavefront shape, fiber direction, and neighboring cell states.
The monodomain model is a system of partial differential equations describing the propagation of excitation in the heart, which captures all of these dependencies through its underlying physics, but is computationally expensive. The DREAM (Diffusion Reaction Eikonal Alternant Model) is significantly faster, but to achieve this, it needs the local CV as an explicit input rather than computing it from first principles. Currently, DREAM predicts CV using an analytic formula that only considers the local recovery state. Spatial factors, such as wavefront curvature, neighbor recovery, fiber alignment, and tissue heterogeneity are ignored, even though they can significantly affect propagation.
An accurate CV prediction that accounts for both, temporal and spatial context, could substantially improve DREAM's fidelity. Ideally, without sacrificing its speed advantage. This project explores whether a data-driven model, trained on detailed reference simulations, can learn such a prediction.

Student Project
In this project we will use machine learning to predict conduction velocity from local spatiotemporal features. The work covers the full pipeline: setting up and running monodomain reference simulations to generate ground-truth data, extracting local features and ground-truth conduction velocities from the simulation output, developing and training a prediction model, as well as optimizing its architecture and hyperparameters. Once a suitable model is found, it will be integrated into the DREAM solver within our C++ codebase, replacing the current analytic CV formula. Finally, the new approach will be validated on unseen scenarios, comparing accuracy and runtime against both the original DREAM and full monodomain simulations.