Predicting Short- and Long-term Future Occurrence of Atrial Fibrillation from Single-Lead ECG in Normal Sinus Rhythm with an Explainable Deep Learning Model.

点击翻译按钮获取中文摘要

Project Summary/Abstract More than 30 million individuals worldwide are diagnosed with atrial fibrillation (AF), however, another 13% of individuals with AF are left undiagnosed. People with AF have a five-fold increased risk of stroke with up to one-third of all strokes shown to be related to AF. Timely administration of appropriate preventative therapies, especially anticoagulants, can significantly decrease the complications of AF, including strokes, by 65% and mortality by 30%. Digital health technologies offer new approaches to identify individuals with undiagnosed AF, in particular paroxysmal AF (PAF), characterized by occasional episodes of limited duration, for whom a 10-second 12-lead electrocardiography (ECG) performed in the clinical setting is unlikely to overlap with an AF event. Continuous monitoring is promising, but still costly and burdensome for elderly individuals, who are at higher risk. To maximize the diagnostic yield of these technologies, we propose novel methods to predict the future occurrence of AF from a single-lead ECG during normal sinus rhythm. Only recently it was shown that it is possible to predict the future occurrence of AF from 12-lead ECGs in normal sinus rhythm collected in a clinical setting. Here, we propose to predict the occurrence of AF with commercially available single- lead ECG devices, which will enable a scalable alternative for early detection in a non-clinical setting. To achieve this goal, we will analyze retrospectively the raw single-lead ECG data of 10,000+ individuals with PAF over 14 days of monitoring. Validation work will then be carried out in a unique set of 1,718 asymptomatic individuals who participated in the prospective mSToPS clinical trial of AF screening (mean age 73), with full clinical information and co-morbidities. The three aims of this project are: 1. Compute the probability of a future AF event in the short-term for an individual in normal sinus rhythm using classic single-lead ECG features and representation learning based features. 2. Develop a method for long-term prediction of AF onset by evaluating individuals with AF detected in 1, 3, 6 and 12 months from the initial monitored period of normal sinus rhythm and by validating the algorithms using the mSToPS dataset with 3 years of clinical follow-up and annotated co-morbidities. 3. Develop a technique to provide a preliminary interpretation of representation learning features for time-series data applied to the short- and long-term prediction. This retrospective study will develop and optimize new predictive techniques from single-lead ECGs, available through consumer devices, with the goal of identifying individuals at high risk of developing AF. A future direction to build on from this study's results would include a prospective study of AF prediction using consumer single-lead ECG to improve clinical outcomes.