Abstract
Cardiovascular diseases (CVDs) remain the leading cause of global mortality, with early detection critical for improving patient outcomes. Photoplethysmography (PPG) is a widely used, non-invasive signal in both clinical monitors and consumer wearables; however, noise susceptibility and the absence of direct electrophysiological markers limit its utility across home and clinical settings. We present CardioPPG, a cross-modal learning framework that augments PPG with ECG-derived representations to improve CVD prediction and enable ECG synthesis for interpretability. Through cross-modal contrastive learning, CardioPPG aligns PPG and ECG features in a shared latent space, followed by an autoregressive generative model that synthesizes high-quality ECG signals. Extensive evaluations show that CardioPPG surpasses a PPG-only self-supervised baseline across multiple CVDs screening-including mitral and aortic valvular disease, atrial fibrillation, cardiomyopathy, and paroxysmal supraventricular tachycardia, among others-with relative AUC gains over the baseline of 41.3%, 12.2%, 8.8%, and 30.3%, respectively. On an external atrial-fibrillation dataset with 86 samples, CardioPPG achieved high AUCs of 99.5% and 98.6% on two PPG channels, confirming the model's generalizability. Furthermore, it generates ECG signals whose distributions closely match those of authentic ECGs, enhancing the interpretability. CardioPPG offers a scalable, real-time, non-invasive solution for CVD monitoring, with significant potential for use in-home settings and resource-limited environments, thus facilitating early detection and timely intervention.