Feasibility of ECG Reconstruction from Minimal Lead Sets Using Convolutional Neural Networks

Abstract

Electrocardiography is a commonly applied method of measuring the electrical activity of the heart. The standard 12-lead electrocardiogram (ECG) provides sufficient information to allow various heart conditions to be diagnosed. Despite its relative ease of use, the standard ECG procedure could benefit from a reduction of leads which may allow for continuous monitoring, for example via a wearable device. In this study, we first investigated the use of variational autoencoders (VAEs) to assess what the most representative leads of the standard 12-lead system are. As the VAE learns to compress the ECG data, it focuses on the parts of the input that is important for the reconstruction. This information is then used to assess which leads are the most useful for a reconstruction task in general. Precordial leads V2, V3 and V4 are shown to contain the most information in the 12-lead ECG data. We then investigated the use of a convolutional neural network (CNN) architecture capable of learning patient-specific models to accurately impute 11 missing ECG signals from a single available lead. Our design is unconventional in that it keeps a two-dimensional structure throughout the fully connected layers. We show that this design outperforms the traditional one-dimensional structure and that these architectures can be affected by the presence of symptoms in recorded heart signals.