Unsupervised diffusion framework for hypertension detection from photoplethysmography

초록

While blood pressure monitoring is essential for managing hypertension, labeled data from hypertensive patients are extremely scarce and often unreliable due to ethical, clinical, and demographic constraints. This scarcity leads to severe class imbalance in existing datasets and fundamentally limits the scalability of supervised learning approaches that depend on balanced annotations. To address this challenge, we present an unsupervised hypertension detection framework based on Denoising Diffusion Implicit Models (DDIM) with photoplethysmography (PPG) signals. The model is trained exclusively on PPG signals collected from normotensive individuals to learn the distribution of normal physiological waveforms, thereby enabling the identification of hypertensive patterns as deviations from this learned distribution. A temporally conditioned 1D U-Net is employed to estimate noise in the reverse diffusion process, allowing for deterministic signal reconstruction within the DDIM framework. Experiments conducted on three publicly available datasets—BP Assessment, PPG-BP, and PulseDB—demonstrate high F1 scores of 0.9067, 0.9130, and 0.9481, respectively. These findings indicate that the proposed framework achieves consistent anomaly-score separability across multiple public datasets under a validation-calibrated threshold selection protocol, suggesting its potential for label-efficient hypertension screening.

키워드