Novel Algorithm and Data Strategies to detect and Predict atrial fibrillation for post-stroke patients (NADSP) - Project Summary Atrial fibrillation (AF) is the most common arrhythmia, affecting 33.5 million people globally with a growing prevalence. AF is associated with significant morbidity and mortality, including 20% of all strokes, 33% of hospitalizations related to cardiac arrhythmias, and a two-fold increase in risk of death. To reduce AF-associated risks such as stroke, it is important to be able to diagnose AF early in the AF trajectory when it is asymptomatic and paroxysmal in order to initiate effective stroke prevention interventions including anticoagulation. Unfortunately, it is estimated that 700,000 people in the USA may have previously unknown AF, and newly detected AF at the time of stroke was found among 18% of AF-associated stroke incidents. Plethysmography (PPG) measures pulsatile blood volume changes and is available in up to 71% of consumer wearables. Because of this unmatched availability, PPG-based AF detection is ideally poised to enable low-cost, long-term, and continuous AF monitoring at scale. However, modest performance of PPG-based AF detection when PPG signals do not have perfect signal quality remains a critical impediment to fully realize its potential as an AF-monitoring tool at scale. The proposed study aims to overcome this challenge by pursuing the following aims: 1) design, develop, and validate a novel deep neural network (DNN) architecture that integrate PPG signal quality assessment with AF detection to accurately detect AF even for signals with imperfect signal quality; 2) validate and test further personalization of the proposed DNN using prospective data from post stroke patients to be collected in ambulatory settings; 3) develop and validate interpretable EHR-data driven machine learning approaches to identify patients with elevated risk of AF for whom PPG-based AF monitoring can be most likely beneficial.
