Machine Learning in Atrial Fibrillation

点击翻译按钮获取中文摘要

Project Summary Atrial fibrillation (AF) is the most common heart rhythm disorder, affecting 2 million Americans in whom it may cause skipped heart beats, dizziness or stroke. Unfortunately, therapy for AF has limited success, likely because AF represents heterogenous and poorly characterized disease entities between individuals. A central challenge is that it is not clear why a specific therapy works in a given AF patient. This uncertainty makes it challenging to develop a patient-specific approach to tailor therapy for personalized medicine. The premise of this project is that mechanistic data is increasingly available in AF patients at scales spanning tissue, whole heart and patient levels, yet rarely integrated. We set out to use machine learning (ML), a powerful approach proven to classify complex datasets, to integrate data to address 3 clinical unmet needs. First, electrograms are rarely used to guide therapy in AF, unlike organized rhythms, because they are difficult to interpret. Second, it is difficult to understand how arrhythmia is affected by any specific ablation strategy in AF, unlike organized rhythms. This makes it difficult to improve therapy. Third, it is difficult to identify whether an individual patient will or will not have success from AF ablation. We applied machine learning and novel objective analyses to these questions to develop strategies for personalized AF therapy. We have 3 specific aims: (1) To identify components of AF electrograms using ML trained to monophasic action potentials (MAP); (2) To identify electrical and structural features of the acute response of AF to ablation near and remote from PVs; (3) To identify patients in whom ablation is unsuccessful or successful long-term, who are poorly separated at present. Each Aim will compare ML to traditional biostatistics, and use objective explainability analysis of ML to provide mechanistic insights. This study has potential to deliver immediate clinical and translational impact. We will apply specific ML approaches, biostatistics, and computer modeling to our rich multiscale registry. We will develop practical and shareable tools, which we will prospectively test clinically, to deliver meaningful outcomes at tissue, whole heart and patient scales. Our team is experienced in electrophysiology, computer science, signal processing and biological physics. This project is likely to reveal novel multiscale AF phenotypes to enable personalized therapy.