Machine Learning for Ventricular Arrhythmias

Project Summary Ventricular tachycardia (VT) and fibrillation are leading causes of cardiac arrest, dizziness, syncope and hospitalization in the United States and worldwide. However, the management of patients at risk for VT remains suboptimal despite scientific discoveries from basic to population science. In particular, there is no framework to estimate which patients with VT are likely to respond to anti- arrhythmic medications or ablation. Therapy is thus empirical. There is great excitement to use analysis of “big data” to personalize VT therapy, but this has not yet improved outcomes. This project develops a novel computational approach to personalize VT therapy that combines machine learning in large registries with computational models. Machine learning will be applied to data across biological scales that span bedside, laboratory and non-invasive imaging, to predict which patients are likely to respond to therapy. Computer models will be used to estimate if a given patient's heart is likely to support VT before versus after therapy. We will validate results in large external registries from different Institutions. We have 3 specific aims: (1) To develop a computational pipeline to predict response to VT ablation using bedside, laboratory and non-invasive imaging; (2) To use machine learning of clinical data and non-invasive imaging to identify which patients with VT will respond to anti- arrhythmic medications in a large database; (3) To combine computational approaches to estimate the relative likelihood that a given patient will respond to various forms of therapy. Results from each Aim will be tested in independent external registries. We will probe computational models to identify clinical phenotypes that could be applied at the bedside. This project will provide immediate clinical impact for patients with VT. We will combine machine learning with physics-based computer models in large registries at Stanford and External centers. We will reduce computational bias using FAIR methods (Findable, Accessible, Interoperable, and Reusable), and make tools freely available per the 2018 NIH Strategic Plan for Data Science. Our team comprises experts in clinical and basic electrophysiology, imaging, machine learning, bioengineering and statistics. The project is very feasible.