CLEP-GAN: An Innovative Approach to Subject-Independent ECG Reconstruction from PPG Signals

TL;DR

This work tackles the problem of reconstructing unseen ECG signals from noninvasive PPG measurements, addressing limited real-world ECG data diversity and noise. It introduces CLEP-GAN, a subject-independent framework that fuses contrastive learning, adversarial training, and attention gating, with an optional VQ-VAE variant CLEP-VQGAN, to improve reconstruction of unobserved ECGs from PPG. A novel ODE-based data generation method creates synthetic ECG-PPG pairs across three common rhythms and varying RR intervals, augmenting diversity and enabling robust training. Extensive experiments on synthetic and real BIDMC/CapnoBase datasets show competitive performance against CardioGAN, QRS-ED, and RDDM, with notable gains in RR interval fidelity, HRV preservation, and T-wave reconstruction in some cases. The study highlights the importance of demographic factors, data diversity, and transfer learning in achieving reliable subject-independent ECG reconstruction from PPG for scalable, non-invasive cardiac monitoring.

Abstract

This study addresses the challenge of reconstructing unseen ECG signals from PPG signals, a critical task for non-invasive cardiac monitoring. While numerous public ECG-PPG datasets are available, they lack the diversity seen in image datasets, and data collection processes often introduce noise, complicating ECG reconstruction from PPG even with advanced machine learning models. To tackle these challenges, we first introduce a novel synthetic ECG-PPG data generation technique using an ODE model to enhance training diversity. Next, we develop a novel subject-independent PPG-to-ECG reconstruction model that integrates contrastive learning, adversarial learning, and attention gating, achieving results comparable to or even surpassing existing approaches for unseen ECG reconstruction. Finally, we examine factors such as sex and age that impact reconstruction accuracy, emphasizing the importance of considering demographic diversity during model training and dataset augmentation.

Figures (22)

• Figure 1: Comparison between raw and filtered signals. ECG signals were filtered between $0.4$ Hz and $45$ Hz, while PPG signals were filtered between $0.3$ Hz and $8$ Hz.
• Figure 2: Examples of three common rhythms generated by our ODE algorithm.
• Figure 3: Comparison of RR interval distributions between synthetic ECGs and actual ECGs. For peak detection, we utilized the peak finding algorithm from NeuroKit2 Makowski2021neurokit. The real ECG signals originate from two different records: record 12 and record 51, within the BIDMC dataset.
• Figure 4: Illustration of RR interval discrepancies between two pairs of synthetic and real ECG signals. The real signals are derived from the BIDMC dataset. Both the black star markers and the vertical black dashed lines indicate the locations of R peaks in the real ECG. For peak detection, we employed the peak finding algorithm from NeuroKit2.
• Figure 5: The architecture of our method.