Introduction: There are numerous approaches to treating Atrial Fibrillation (AF) using catheter ablation therapy. These approaches still require optimisation, with 41% of persistent AF patients reverting to AF 18 months after receiving pulmonary vein isolation (PVI). We aimed to use a virtual cohort of patient-specific left atrial models to compare ablation techniques targeting the anatomical, structural and electrical substrate of AF. We also wanted to investigate which factors are important for predicting simulated ablation response.
Methods: Atrial anatomy was segmented from contrast enhanced magnetic resonance angiogram (CE-MRA) images of 50 patients (20 with paroxysmal and 30 with persistent AF) and used to generate computational meshes. Human atrial ex-vivo DTMRI atlas endocardial and epicardial fibres were mapped to each anatomical mesh using the Universal Atrial Coordinate system. Fibrosis was incorporated into the models by registering late-gadolinium enhancement magnetic resonance imaging (LGE-MRI) with CE-MRA data. AF simulations were then run on the patient-specific models using Cardiac Arrhythmia Research Package (CARP) software which compute cellular transmembrane potentials across the epicardium and endocardium of the left atria during AF. Simulations were post-processed to generate endocardial phase singularity (PS) density maps indicating potential driver site locations (hotspots) over 15 seconds. 6 different ablation approaches were evaluated: a) PVI alone; or PVI and: b) box ablation; c) all driver hotspots; d) all fibrosis areas; e) single driver hotspot (largest driver site, identified as high PS density); f) single fibrosis area (largest fibrotic area, identified as high LGE-MRI intensity). Ablation lesion maps and transmembrane potential maps were generated for the resulting post-ablation models. Figure 1 provides a visual representation of this model construction, AF simulation and ablation process. A machine learning random forest classifier was trained to predict simulated ablation outcome.
Results: Acute response to ablation was classified as AF (non-responders) and AT (atrial tachycardia) or termination (responders). The optimal ablation approach resulting in termination, or if not possible AT, varied among the virtual patient cohort: 20% PVI alone, 6% box ablation, 46% all driver hotspots, 4% all fibrosis areas, 2% single driver hotspot and 2% single fibrosis area. 20% of the virtual patients were non-responders to all 6 ablation approaches. The accuracy of classifier predictions improved from 0.73 to 0.83 by incorporating patient-specific and ablation pattern specific metrics, rather than simply including ablation type.
Conclusion: Patient anatomy, patient-specific fibrosis properties and driver site locations are all important in determining individual AF mechanisms. The improvement in random forest classifier accuracy with the incorporation of patient-specific and ablation pattern specific metrics highlights the importance of these factors in predicting ablation success. When planning the approach for catheter ablation therapy and predicting acute response, these factors and the distribution of lesions must be considered.
