P2E-VQ Augments PPG Signals with ECG Data for Better Predictions

Photoplethysmography (PPG) signals, commonly collected by wearables, are cost-effective but lack the detailed cardiac electrical activity information found in Electrocardiography (ECG), limiting their utility for diagnosing certain cardiac conditions. Existing attempts to reconstruct ECG from PPG are often ill-posed and don't always translate to better predictive performance. This research introduces P2E-VQ, a novel retrieval-augmented framework designed to bridge this gap more effectively. P2E-VQ avoids direct ECG waveform reconstruction. Instead, it converts PPG signal patches into discrete tokens and then retrieves relevant ECG-linked information from a pre-built memory bank, which is constructed solely from training data. This process effectively augments the PPG representations with crucial ECG-derived insights. A key advantage is that P2E-VQ only requires PPG signals during inference, making it practical for real-world wearable applications. Extensive evaluations across five public datasets and six downstream tasks, including clinical endpoint prediction and affective state recognition, demonstrated P2E-VQ's consistent superiority. It outperformed pretrained baselines under a unified frozen-feature linear-probing protocol, highlighting its effectiveness in enhancing PPG's predictive power by leveraging the rich information typically associated with ECG.