Exploring the Reconstruction of Electrocardiograms from Photoplethysmograms via Deep Learning

The aim of this study is to explore the reconstruction of electrocardiogram (ECG) signals from photoplethysmography (PPG) signals, using data from healthy and diseased volunteers. To this aim, we review a novel method based in deep learning to reconstruct ECG signals and test it on two public databases: MIMIC-III and BIDMC. In the MIMICIII database, the system's behavior was analyzed in response to different frequency contents of the ECG (0.5 to 20 Hz, 0.5 to 40 Hz and 0.5 to 60 Hz), as well as variations in window width (1 s, 2 s, 3 s and 4 s). Furthermore, comparative tests were conducted between healthy and diseased volunteers. In the second database, the validation of the method was verified. The average correlation values obtained in MIMIC-III and BIDMC were higher than 0.80, confirming the previous results reported in the literature. In the final analysis, the system behavior using data from patients with pre-existing conditions exhibited inferior performance compared to those who were in good health. Clinical Relevance — It is feasible to reconstruct ECG signals from PPG signals under controlled clinical conditions. However, it is still necessary to continue exploring reconstruction methods as the one employed here exhibited lower performance in patients with preexisting heart diseases than in healthy subjects.