The Fibrotic Kernel Signature: Simulation-Free Prediction of Atrial Fibrillation

Abstract

We propose a fast classifier that is able to predict atrial fibrillation inducibility in patient-specific cardiac models. Our classifier is general and it does not require re-training for new anatomies, fibrosis patterns, and ablation lines. This is achieved by training the classifier on a variant of the Heat Kernel Signature (HKS). Here, we introduce the “fibrotic kernel signature” (FKS), which extends the HKS by incorporating fibrosis information. The FKS is fast to compute, when compared to standard cardiac models like the monodomain equation. We tested the classifier on 9 combinations of ablation lines and fibrosis patterns. We achieved maximum balanced accuracies with the classifiers ranging from 75.8% to 95.8%, when tested on single points. The classifier is also able to predict very well the overall inducibility of the model. We think that our classifier can speed up the calculation of inducibility maps in a way that is crucial to create better personalized ablation treatments within the time constraints of the clinical setting.