Investigating Potentially Arrhythmogenic Atrial Substrate using Computational Simulations and Explainable Machine Learning

Abstract

This dissertation discusses applications of computational simulations and explainable machine learning (xML) on the mechanistic understanding of arrhythmogenic atrial substrate. We review the current understanding of atrial arrhythmia mechanisms in Chapter 1. Then, we explore how computational simulations, in highly detailed patient-specific atrial models derived from late gadolinium enhanced (LGE) magnetic resonance imaging (MRI) scans, and machine learning give rise to new frontiers of clinical treatment in Chapter 2. In Chapter 3, we aim to understand the absence of arrhythmia in patients who suffered an embolic stroke of undetermined source (ESUS) despite the presence of putatively pro-arrhythmic fibrosis. We reconstructed patient-specific atrial models from ESUS and atrial fibrillation (AFib) patients then assessed the arrhythmogenic capacity of the fibrotic substrate. Analyzing the reentrant drivers in each subpopulation revealed that the intrinsic pro-arrhythmic substrate properties of fibrosis were indistinguishable between ESUS and AFib patients. In our second aim, we investigate the synergy between ablation-induced scar and native fibrosis in atrial models of persistent AFib patients. We reconstructed pre- and post-ablation models to assess AFib inducibility. We classified simulated arrhythmia episodes and determined that pre-ablation models were more likely to harbor rotor-like activity, while post-ablation models were more likely to harbor anchored reentry around ablation lesions, veins, or valves. We then developed an xML algorithm to predict and quantify the spatial properties of arrhythmogenic scar regions (Chapter 4). Our third aim is to use xML techniques to distinguish populations of patients at risk for recurrent arrhythmias from those who will remain arrhythmia free following ablation. We identified the key population-level and patient-specific level risk factors for post-ablation recurrence. The explainable aspect of our approach allowed us to understand why a particular patient can have large prediction weights for some risk factors without tipping the balance towards an incorrect prediction. We thus present a comprehensive clinical tool to explain patient recurrence risk by combining patient-specific clinical profiles and left atrial pre/post-ablation substrate patterns via an explainable classifier (Chapter 5). In conclusion, this dissertation showcases the potential of computational simulations and xML in advancing our understanding of arrhythmogenic atrial substrate and its implications for clinical practice. The findings highlight the importance of patient-specific modeling, the role of fibrosis in arrhythmia susceptibility, the impact of ablation on arrhythmia inducibility, and the predictive capabilities of xML techniques in assessing catheter ablation outcomes. By integrating these approaches, this research offers valuable insights for risk assessment, personalized treatment strategies, and the development of comprehensive clinical tools for the management of atrial arrhythmias.