Radiofrequency ablation planning for cardiac arrhythmia treatment using modeling and machine learning approaches

Cardiac arrhythmias are heart rhythm disruptions which can lead to sudden cardiac death. They require a deeper understanding for appropriate treatment planning. In this thesis, we integrate personalized structural and functional data into a 3D tetrahedral mesh of the biventricular myocardium. Next, the Mitchell-Schaeffer (MS) simplified biophysical model is used to study the spatial heterogeneity of electrophysiological (EP) tissue properties and their role in arrhythmogenesis. Radiofrequency ablation (RFA) with the elimination of local abnormal ventricular activities (LAVA) has recently arisen as a potentially curative treatment for ventricular tachycardia but the EP studies required to locate LAVA are lengthy and invasive. LAVA are commonly found within the heterogeneous scar, which can be imaged non-invasively with 3D delayed enhanced magnetic resonance imaging (DE-MRI). We evaluate the use of advanced image features in a random forest machine learning framework to identify areas of LAVA-inducing tissue. Furthermore, we detail the dataset’s inherent error sources and their formal integration in the training process. Finally, we construct MRI-based structural patient-specific heart models and couple them with the MS model. We model a recording catheter using a dipole approach and generate distinct normal and LAVA-like electrograms at locations where they have been found in clinics. This enriches our predictions of the locations of LAVA-inducing tissue obtained through image-based learning. Confidence maps can be generated and analyzed prior to RFA to guide the intervention. These contributions have led to promising results and proofs of concepts.