PPG Heart-Age Gap Tracks Vascular Aging

Wearable photoplethysmography is easy to collect repeatedly, but its relationship to cardiovascular aging has been less established than ECG-, imaging-, or clinic-based assessments. The authors train a smartwatch PPG heart-age model using ECG only during representation pretraining, then test whether the resulting PPG-only heart-age gap carries information beyond chronological age. Across three OPPO cohorts in China, the paper connects that gap to pulse-wave velocity (PWV), a measure of arterial stiffness, and to prevalent hypertension.

Core Contribution

The central contribution is an ECG-guided route to PPG-only inference. Rather than treating PPG as a weaker substitute for ECG, the authors use synchronized single-lead ECG and four-channel PPG recordings to teach the PPG encoder representations aligned with ECG during pretraining. The downstream model predicts chronological age from PPG, and the difference between predicted and chronological age becomes the heart-age gap.

That distinction matters because a low age-prediction error alone would not establish a useful biomarker. The paper tests whether the residual signal is associated with vascular status after chronological-age adjustment. In the PWV cohort, a larger heart-age gap is associated with higher age-adjusted PWV; in the home blood-pressure-monitoring cohort, it is associated with higher odds of already prevalent hypertension. The evidence therefore supports the gap as a cross-sectional risk marker, not yet as a prospective predictor of events.

Technical Approach

The study begins with subject-aware self-supervised learning within each modality and CLIP-style contrastive alignment across synchronized ECG and PPG recordings. Modality-specific 1D ResNet encoders learn representations from the two signals, with the alignment objective bringing recordings from the same synchronized observation closer together. The PPG encoder is then fine-tuned against chronological age. ECG is absent at inference, so the deployed pathway needs only smartwatch PPG.

Figure 1 lays out this separation between multimodal training and PPG-only deployment, then traces the heart-age gap into the vascular and hypertension analyses.

The authors aggregate repeated recordings at the participant level, rather than treating all measurements as independent samples. Their short-term aggregation analysis spans one to 15 recordings acquired within at most three months and assumes that an individual's heart-age phenotype remains approximately stable in that window. This is a practical choice for consumer wearables: repeated measurements can reduce noise, but the reported gain should not be read as validation over longer physiological changes.

Results and Analysis

In external validation, the SA-CLIP-pretrained four-channel PPG model reaches a subject-level MAE of 5.895 years with Pearson's $r=0.819$ in the PWV cohort, and 4.344 years with $r=0.800$ in the home blood-pressure-monitoring cohort. Those figures show that the model recovers chronological age with moderate precision across cohorts, while leaving enough variation in the residual for the cardiovascular analyses.

The more consequential result is the association with arterial stiffness. After adjusting for chronological age, heart-age gap and PWV residuals have a partial correlation of 0.2627 ($P<0.001$); each 1-year increase in heart-age gap is associated with 0.062 m/s higher PWV. Categorically, participants with accelerated heart aging, defined as a gap above 3 years, have 0.91 m/s higher adjusted PWV than the decelerated group, whose gap is below -3 years. Figure 5 makes the group separation tangible after adjustment for age, sex, and BMI. The effect is informative, though correlation magnitude indicates that heart age is one signal among many determinants of vascular stiffness.

For hypertension, each one-standard-deviation increase in adjusted heart-age gap corresponds to an odds ratio of 1.72 (95% CI 1.49–1.99), while the highest quartile has an odds ratio of 4.25 relative to the lowest. Figure 7 reports models adjusted for chronological age, sex, BMI, and family history, alongside sex-stratified analyses. The gradient is substantial for a passive wearable measurement, but the cohort is cross-sectional: it measures odds of existing hypertension, not future incidence or treatment benefit.

Caveats in Practice

The data are large—581,804 participants and 7,452,131 recordings—but all cohorts are drawn from OPPO studies in China, which limits evidence for device, population, and care-setting transfer. The target is chronological age, so the biological interpretation rests on residual associations rather than a direct ground-truth measure of cardiovascular age. Finally, PWV and hypertension analyses establish adjusted associations; they do not show that acting on a PPG-derived gap improves clinical outcomes.