Abstract
BACKGROUND: Reconstructing ECG signals from PPG measurements is a critical task for non-invasive cardiac monitoring. While several public ECG-PPG datasets exist, they lack the diversity found in image datasets, and the data collection process often introduces noise, making ECG reconstruction from PPG signals challenging even for advanced machine learning models. RESULTS: We propose a novel ODE-based method for generating synthetic ECG-PPG pairs to enhance training diversity. Building on this, we introduce CLEP-GAN, a subject-independent PPG-to-ECG reconstruction framework that integrates contrastive learning, adversarial learning, and attention gating. CLEP-GAN achieves performance that matches or surpasses current state-of-the-art methods, particularly in reconstructing ECG signals from unseen subjects. Evaluation on real-world datasets (BIDMC and CapnoBase) confirms its effectiveness. Additionally, our analysis shows that demographic factors such as sex and age significantly impact reconstruction accuracy, emphasizing the importance of incorporating demographic diversity during model training and data augmentation. CONCLUSIONS: Our method produces synthetic ECG-PPG pairs with RR interval distributions closely aligned with their real counterparts and shows strong potential to simulate diverse rhythms such as regular sinus rhythm (RSR), sinus arrhythmia (SA), and atrial fibrillation (AFib). Furthermore, CLEP-GAN demonstrates robust performance on both synthetic and real datasets, achieving near-perfect reconstruction in synthetic settings and competitive results on real data. These findings highlight CLEP-GAN's promise for reliable, non-invasive ECG monitoring in clinical applications.