Problem
Doctors treating persistent atrial fibrillation (AF) must pick which heart tissue to isolate by catheter ablation. Traditional treatment of persistent AF has about 21% long-term success. Guidelines prescribe pulmonary vein isolation (PVI) but do not say what else to isolate. A digital twin built from the patient's MRI can test candidate strategies by simulating the arrhythmia. In the US startup approach the simulation runs for a week or more on a supercomputer, so the patient waits at home.
Example from the article: for 1000 virtual patients, each of four ablation strategies (PVI, PVI+LA, PVI+RA, PVI+LA+RA) needs its own simulation to learn whether the arrhythmia stops.
Solution in brief
Simulate one AF episode per virtual patient, compute anatomical and physiological feature maps from it, and train a classifier that predicts whether a given ablation mask stops the arrhythmia. A new patient then needs one simulated episode instead of one per strategy. Treat the two atria as separate inputs of a Siamese network.
Questions this case answers
- How to replace repeated slow physical simulations with a neural classifier that predicts the outcome of each candidate intervention for an individual
- How to feed two related anatomical objects into one CNN without mixing unrelated regions
- How to replace slow physical simulation with a neural network that predicts an outcome for each candidate action
- How to predict the outcome of atrial fibrillation ablation for a specific patient
Tools and methods
- Biophysical AF simulation — produces features and the "arrhythmia terminated" label for 1000 virtual patients
- 3D to 2D atlas mapping — makes feature maps of different anatomies comparable
- Siamese DenseNet with multi-modal fusion — one branch per atrium
- Diffusion model for fibrosis maps — adds synthetic training cases
- ROC-AUC — comparison of feature sets and training data
Google or a one-time email code. No password.
