Augmentation of Cardiac Ischemic Geometry for Improving Machine Learning Performance in Arrhythmic Risk Stratification

AbstractVentricular arrhythmias frequently occur as a complication of myocardial infarction (MI), due to significant changes in the heart’s structure and electrophysiology. If left untreated, these alterations may lead to sudden cardiac death (SCD). It is therefore critical to evaluate risk prediction accurately in post-infarction patients to enable early intervention and improve patient outcomes. This work introduces a novel approach to improve arrhythmia risk assessment in post-infarction patients. We propose a new pipeline to build physiologically realistic image-based models of patient hearts, producing more realistic meshes compared to publicly available pipelines. We generate a library of 90 cardiac geometries of MI patients and use these cardiac models to estimate likelihood of reentry using electrophysiological (EP) simulations. However, due to the computationally expensive nature of this approach, we also introduce a data augmentation pipeline to train a machine learning (ML) model for risk stratification, enabling accurate and real-time prediction of the simulation outcomes. Our trained ML model achieved an accuracy of 88.0% and F1 score of 48%, with a prediction time of 0.01 seconds per case (compare with approximately 5 hours per case for EP simulations). In conclusion, the work presented here improved the accuracy of personalised biventricular geometries, introduced a novel data augmentation approach for scar distribution, and decreased prediction time of risk of arrhythmias post-MI by more than five orders of magnitude.