M.Sc Thesis

M.Sc StudentChocron Armand
SubjectRemote Diagnosis and Phenotyping of Atrial Fibrillation
using Machine Learning
DepartmentDepartment of Electrical and Computer Engineering
Supervisors PROFESSOR EMERITUS Yehoshua Zeevi
DR. Joachim Behar
Full Thesis textFull thesis text - English Version


Atrial fibrillation (AF) is the most common arrhythmia, with an estimated prevalence of 3% in adults aged 20 years or older. It is associated with quivering or irregular heartbeat, that can lead to blood clots, stroke, heart failure and other heart-related complications. Treatments for AF include cardioversion, cardiac ablation and drugs that can help control the heart rate. The currently accepted convention for AF diagnosis is the presence of an episode lasting at least 30 seconds on the electrocardiogram (ECG). Many individuals with AF go undiagnosed and thus untreated because they are often asymptomatic or have paroxysmal AF (PAF) i.e. episodes of AF that occur occasionally. AF is diagnosed based on ECG analysis including 12-lead ECG examination at the hospital or Holter ECG typically worn for 24 hours. The 12-lead ECG will naturally miss a significant proportion of the PAF while the Holter ECG necessitates specialized equipment to be set-up and is hence limited in quantity thus precluding wider diagnosis and long-term monitoring. In addition, algorithms for AF detection used in software of both 12-lead and Holter ECG have limited performance and thus diagnosis necessitates a specialist to manually interpret the recordings. 

This thesis makes two contributions to the field. First, it presents a novel paradigm for AF diagnosis and monitoring, namely: performing diagnosis of AF during sleep, relying on statistical measures derived from the heart rate. We evaluate the feasibility of AF detection during sleep on a polysomnographic database. We consider in this database 2,890 patients for the study, among which 70 suffer from AF. Model prediction on the database shows an overall Se = 0.97, Sp = 0.99, NPV = 0.99 and PPV = 0.67 in classifying individuals with or without prominent AF. Over 22% of correctly identified prominent AF rhythm cases were not previously documented as AF in the database. Second, this research introduces an innovative solution for AF detection: the usage of a single-channel ECG portable sensor as a way to detect AF remotely and accurately. The proposed algorithm leverages recent advances in machine learning, for the purpose of AF diagnosis and phenotyping through the estimation of the AF burden (the percentage of time spent in AF). The model was developed and evaluated on a large database of p=2,891 patients, totalling t=68,800 hours of continuous electrocardiography (ECG) recordings acquired at the University of Virginia heart station. Specifically, 24h beat-to-beat time series were obtained from a single portable ECG channel. The absolute AF burden estimation error |EAF |(%) median and interquartile, on the test set, was 1.2 (0.1-6.7) for ArNet and 3.1 (0.0-11.7) for XGB for AF individuals. This demonstrates the feasibility of AFB estimation from 24h beat-to- beat interval time series utilizing deep recurrent neural networks. The thesis demonstrates the value of the novel paradigms for AF diagnosis and the superior performance of the novel algorithms over state-of-the-art AF detectors. Algorithms were developed on an overall dataset of 5,979 patients totalling 95,713 hours of continuous recording.