Abstracts Engineering

Deep Learning-Based Risk Prediction of Atrial Fibrillation Using the 12-lead ECG

by Theogene Habineza

Atrial fibrillation is one of the most common cardiac arrhythmias that affects millions of people each year worldwide. As studies have shown, there exists a close link between atrial fibrillation (AF) and increased risk of cardiovascular diseases such as stroke and heartfailure. One way to diagnose AF is to analyse an electrocardiogram (ECG), since cardiac arrhythmias provoke changes in the normal ECG pattern. The aim of this study was to predict the imminent riskof developing AF, and how long it might take for the event to occur using Deep Neural Network (DNN). The objectives were (i) to investigate the ability of a DNN to distinguish an ECG with AF against a normal ECG; (ii) using DNN to predict if a person with normal ECG will develop AF; and (iii) in case a patient might develop AF, predict the time to the occurrence of AF event from the date an ECG exam is recorded. For model development and testing, the study used a collection of ECG recordings and annotations made by a public telehealth system, the Telehealth Network of Minas Gerais, in Brazil. A convolutional neural network based on a residual network architecture was implemented to produce class probabilities and make predictions. A Cox proportional hazards model and a Kaplan-Meier model were developed to carry out survival analysis. The DNN model proved to have the ability of detecting AF condition with an area under the receiver operating characteristic curve (AUC score) of 0.992. It showed the capability of identifying a patient with the risk of developing AF with an AUC score of 0.845. Patients within the high risk group are 50% more likely to develop AF within 40 weeks, while patients belonging to the minimal risk group have more that 85% chance of remaining AF free up until after seven years. If applied in clinical practice, the models possess the potential of providing valuable and useful information in decision-making and patient management processes.