Autoencoder-Based Anomaly Detection in ECG Image Time Series Data: A Comparative Evaluation of Three Different Architectures

Abstract

Anomaly detection is an essential component of machine learning that renders the outcomes neutral to any category or class. Due to the wide range of anomalies that might exist in time-series data, it plays a crucial role in time-series modelling. This research paper presents an image reconstruction-based approach for anomaly detection in electrocardiogram (ECG) time series data and image data using three different autoencoders, and concludes with a comparison analysis. Anomaly detection is crucial for accurate diagnosis and treatment of heart diseases, and the proposed approach utilizes three types of autoencoders to learn the normal patterns in image data and identify any deviations from these patterns as anomalies. The approach uses a dataset of normal images and is tested on a dataset containing both normal and anomalous images. To reconstruct unseen data like anomalies, models trained only on normal sets will not be able to do so. The results demonstrate that all three types of autoencoders can effectively reconstruct the images, but the Convolutional Autoencoder was able to reconstruct most of the anomaly images correctly, which may not be ideal for anomaly detection. On the other hand, the Deep Autoencoder and Stacked Autoencoder show the reconstruction of the anomalies similar to the normal images, hence, can be very effective for detecting anomalies in ECG image time-series data and similar problems. The proposed approach demonstrates the effectiveness of autoencoder-based anomaly detection for image data and provides a comparative analysis of different types of autoencoders for this task. This can potentially improve the accuracy and efficiency of anomaly detection in various image datasets. © 2023 IEEE.