A Comparison of Deep Learning Methods for Identifying Anomalous Heartbeats in Electrocardiogram Data

• Activity Number: 341 - Contributed Poster Presentations: Section on Statistical Learning and Data Science
• Type: Contributed
• Date/Time: Tuesday, August 9, 2022 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Thomas Dunn* and Tyler Cook and Emily Hendryx
• Companies: University of Central Oklahoma and University of Central Oklahoma and University of Central Oklahoma
• Keywords: deep learning; neural networks; electrocardiogram; anomaly detection

Abstract

Electrocardiogram (ECG) data can provide a wealth of information about a patient’s health. While physicians are able to use ECG data to identify anomalous heartbeats and diagnose illnesses, the low frequency of anomalous beats and sheer volume of patient data generated in the clinical setting make automated tools for ECG screening appealing. To address this, we compare the performance of several popular deep learning models including vanilla autoencoders, convolutional autoencoders, recurrent neural networks, and variational autoencoders in identifying anomalous beats within ECG data due to arrhythmias. In some cases, we couple layers from these models with other classifiers to identify normal and anomalous beats. Each method is evaluated on the MIT-BIH Arrhythmia Database, and a variety of classification metrics are calculated for comparison with existing results in the literature. The results of this work can offer insight toward choosing network architectures best suited for identifying anomalous beats within ECG data, in turn offering automated support for improved patient care.