Anomaly Detection in Electrocardiogram Using Fully Convolutional Neural Network for Cardiac Disease Diagnosis

Abstract

The overall aim of the research is to design a healthcare monitoring system for cardiac patients encompassing an efficient sensor network for collecting and transmitting physiological data captured from multiple health care devices to a cloud repository and a reliable deep learning based physiological data processing algorithms for detecting anomaly. Anomaly detection (aka outlier analysis) is a process to identify data points in a sequence of observations that deviate from a waveform's normal behavior. This paper presents a detailed review and experimental results with analysis of the anomaly detection process using fully convolutional neural networks. The potential of Encoder-Decoder architecture AutoEncoder (AE) and its variants in detecting anomalies in Echocardiogram (ECG) data is explored. The problem of detecting anomalies in the ECG data for cardiac disease analysis is solved using unsupervised learning technique. The process of AE-based anomaly detection involves compressing normal data using the AE into a latent space with a lower dimension than the original data, restoring the data, and calculating the difference between the restored and original data. The suitability of the Variational Autoencoder (VAE) and the challenges in fixing an optimal threshold value for categorizing anomaly samples were analyzed in this paper.