Efficient ECG anomaly detection using autoencoder-based dimensionality reduction: A comparative study of machine learning algorithms

Abstract

Electrocardiograms (ECGs) are widely used to diagnose various cardiovascular conditions. While deep learning models have shown high accuracy in ECG classification, challenges such as high dimensionality, inter-individual variability, limited dataset availability, and long training times hinder their practical deployment - especially in wearable or resource-constrained environments. This study investigates dimensionality reduction from 140 to 16 features using an Autoencoder's latent space and evaluates six anomaly detection algorithms on both the original and reduced datasets. Experimental results show that the reduced feature sets can achieve up to 97% accuracy with a computation time as low as 20 milliseconds, highlighting the potential for efficient, lightweight ECG anomaly detection in real-time applications.