UPR-BP: Unsupervised Photoplethysmography Representation Learning for Noninvasive Blood Pressure Estimation

This study presents UPR-BP, a novel unsupervised representation learning framework utilizing photoplethysmography (PPG) signals for accurate, noninvasive blood pressure (BP) estimation. Leveraging readily available unlabeled PPG data, UPR-BP overcomes the limitations of data-driven models by effectively capturing discriminative BP features without requiring extensive paired measurements. Our framework employs a three-branch architecture with shared weights for joint optimization and incorporates preservation of invariance, variance, and covariance in the PPG temporal encoding, preventing information collapse and generating meaningful deep representations. Additionally, temporal neighborhood coding facilitates the identification of diverse physiological states within the PPG signals. We comprehensively validate UPR-BP on diverse datasets from bedside monitors and wearable wristwatches, encompassing over 4,000 subjects. The proposed approach achieves medical-grade accuracy, demonstrating significant superiority to state-of-the-art techniques with a low estimation error of 0.30 ± 4.68 mmHg for systolic BP and 0.25 ± 2.66 mmHg for diastolic BP. These results highlight the potential of UPR-BP for significantly advancing continuous, noninvasive BP monitoring in clinical settings.