Advancing PPG-Based Continuous Blood Pressure Monitoring from a Generative Perspective

Abstract

Cuffless blood pressure (BP) monitoring is a critical task in the cardiovascular diseases (CVDs) domain, commonly based on Photoplethysmography (PPG) and Electrocardiogram (ECG) signals, providing foresight into cardiac health. While ECG often delivers better BP monitoring performance, the acquisition via straps and patches leads to a poor user experience. On the contrary, PPG enables continuous and convenient monitoring, but offers less informative references. A potential approach is to convert PPG signals into ECG signals, ensuring both high convenience and optimal accuracy. Converting PPG into ECG, however, involves a substantial reduction in inherent entropy, necessitating a thorough understanding of the process and specific techniques to guide the ECG generation. In this paper, we present a blood pressure monitoring framework that achieves ECG-level performance using solely the PPG signal. A diffusion model is introduced to conduct a selective ECG-targeted generative process with the condition of PPG. Based on our observation from the experimental investigation, a set of techniques is developed to significantly enhance the model's ability in generating high-quality ECG signals. Specifically, in the forward process, we employ an adaptive search module to adapt the QRS segment within the ECG waveform. In the reverse process, we propose the scale alignment and frequency alignment modules to better guide the generative process. Extensive experiments conducted on two public datasets and one self-collected dataset demonstrate the superior performance of our proposed framework, offering a groundbreaking perspective for PPG-based continuous blood pressure monitoring.