M.Sc. Stephanie Appel

Jan André Möthrath

Simulations of the atria are used to study how arrhythmias are sustained and how they might respond to treatment. What such a simulation predicts depends strongly on how fast and in which direction excitation spreads through the tissue. However, neither of these can be measured in a patient. Both have to be reconstructed from clinically acquired activation maps, which usually cover the surface only sparsely and carry noise. Several methods have been developed for this reconstruction, recently physics-informed neural networks, which build the underlying propagation physics into the training of the network.
Such methods are almost always assessed by comparing the reconstructed values against the values they were supposed to recover. That answers a narrower question than the one that matters in practice. A reconstruction can be close to the truth on average and still produce a model that behaves differently, because quite different tissue properties can give rise to almost the same activation pattern. Whether a simulation built on reconstructed properties still predicts the same arrhythmia behaviour as one built on the true properties is a separate question, one that a pointwise comparison of reconstructed and true values does not answer.
This project addresses that question directly. On an atrial geometry for which the true properties are known, tissue properties are reconstructed from activation maps, activation is simulated again from the reconstruction and compared against the map the network was given, and a stimulation protocol is used to test whether reentry can be induced. Running that test on the reconstructed and on the true properties side by side shows whether the remaining error is small enough to leave the arrhythmia prediction unchanged, or whether it changes whether and where reentry occurs.