What is it about?
Electrocardiograms printed on paper are often partial, making them difficult to use. To solve this problem, we have created a tool based on artificial intelligence capable of reconstructing the missing parts. Thanks to this method, these completed ECGs can then be used for other automated medical analyses.
Featured Image
Photo by Joachim Schnürle on Unsplash
Why is it important?
Our work is distinguished by the development of a hybrid 1D/2D U-Net model, whose encoder is based on a dual parallel branch. One branch processes ECG signals as a whole via a 2D multi-lead representation, while the other analyzes them individually, lead by lead, in 1D. This architecture enables inter-lead and intra-lead relationships to be captured simultaneously. In addition, we introduce an original cost function combining Mean Squared Error and Pearson Correlation Coefficient: the former aims to minimize spatial amplitude differences, while the latter values the similarity of temporal dynamics. This dual approach enables us to reconstruct signals that are as close to reality as possible.
Perspectives
ECGrecover is a tool that we have designed with the aim of coupling it to solutions for digitizing ECGs recorded on paper. It can faithfully reconstruct the missing portions of signals, opening the way to the exploitation of entire databases which, until now, have remained unexploitable by digital approaches. Beyond this first application, we can also imagine ECGrecover as a tool capable of correcting certain noisy portions of the signal - for example, when an artifact is caused by electrode movement on a specific lead. Ultimately, we hope that this tool can be integrated into a wide range of practical cases to maximize the reusability and usability of ECGs, even when they are incomplete or altered.
Alex Lence
Institut de recherche pour le développement
Read the Original
This page is a summary of: ECGrecover: A Deep Learning Approach for Electrocardiogram Signal Completion, July 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3690624.3709405.
You can read the full text:
Contributors
The following have contributed to this page
