(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Vascular age estimation using a consumer wearable sleep tracker [1]

['Gizem Yilmaz', 'Centre For Sleep', 'Cognition', 'Yong Loo Lin School Of Medicine', 'National University Of Singapore', 'Singapore', 'Shohreh Ghorbani', 'Ju Lynn Ong', 'Hosein Aghayan Golkashani', 'Chen Zhang']

Date: 2026-04 

Vascular aging is traditionally assessed using a combination of clinical markers, blood pressure and arterial stiffness measurement. However, measuring vascular aging with reference equipment is costly and not scalable. Nocturnal photoplethysmography (PPG) from wearable health trackers offer a scalable solution for longitudinal assessment. In this study, we evaluated the ability of a consumer wearable (Oura Ring) to detect age-related differences in PPG waveform, in comparison to a clinical-grade fingertip pulse oximeter. Healthy adults (N = 160; 78 males (49%), median age 31 years (IQR: 23)) underwent overnight polysomnography (PSG) in a sleep laboratory, during which fingertip and wearable ring PPG data were collected simultaneously. Pulse waveforms were extracted from both devices using a custom algorithm and key waveform features were compared across devices. Vascular age was estimated from pulse waveforms using a featureless deep learning model. Prediction performance was compared between the two devices. Age-related waveform changes were most prominent in PPG crest time (CT (samples)) (r = 0.64 and 0.62 for fingertip and wearable devices), while the reflection index (RI) had a weaker correlation with age for the ring sensor (r = 0.22) compared to fingertip (r = 0.58). Despite differences in waveforms between devices, the deep learning model showed comparable prediction performance with mean absolute errors (MAE (SD)) of 6.28 (1.48) and 7.25 (1.29) years, and r (SD) of 0.84 (0.07) and 0.80 (0.10) for clinical-grade and consumer-grade devices, respectively. These findings support the feasibility of using PPG waveforms from wearable devices to assess vascular age.

Vascular aging is an important marker of cardiovascular risk, traditionally assessed using clinical measurements and specialized equipment to evaluate arterial stiffness. The latter is expensive and difficult to scale for long-term or population-level monitoring. Wearable health trackers that collect nocturnal photoplethysmography (PPG) signals offer an alternative. Our study examined whether a consumer wearable device (Oura Ring) can detect age-relevant features in PPG waveforms, compared with a clinical-grade fingertip pulse oximeter. 160 healthy adults (49% male; median age 31 years) underwent overnight polysomnography, during which PPG signals were recorded simultaneously from both devices. Pulse waveforms were extracted and key waveform features were analyzed. Several PPG features showed clear age-related changes, particularly pulse crest time, which correlated strongly with age for both devices. Vascular age was then estimated using a deep-learning model applied to the PPG signals. Although waveforms differed between devices, the deep-learning model predicted vascular age with comparable accuracy using wearable and clinical-grade PPG data. Mean absolute error was approximately 6–7 years for both devices, with strong correlations between predicted and chronological age. These findings demonstrate that PPG data from consumer wearables can capture meaningful age-related vascular information, supporting their potential use for scalable, longitudinal assessment of vascular aging.

Competing interests: I have read the journal’s policy and the authors of this manuscript have the following competing interests: Michael W. L. Chee is a member of the medical advisory board of Oura Health. As of October 2025, Gizem Yilmaz and Ju Lynn Ong are members of the Oura–National University of Singapore (NUS) Joint Lab. This work was conducted independently of Oura Health and the Joint Lab, and was designed, funded, and executed solely by NUS. Oura Health had no role in data analysis, interpretation or in the writing of the manuscript. The manuscript was submitted on August 4, 2025. No other disclosures were reported.

Funding: This work was supported by the Lee Foundation (wearelee.org/lee-foundation), Support funds for the Centre for Sleep and Cognition, Yong Loo Lin School of Medicine (medicine.nus.edu.sg) awarded to Michael W. L. Chee. All grants were awarded to Michael W. L. Chee. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Data Availability: Aggregated participant-level data for Fingertip and Ring devices used to generate tables and figures are available at https://github.com/SCL-NUS/ppg-vascularage-ring . Fingertip PPG waveform data, including all waveforms from each night’s recording as well as the exact subset used for model training and evaluation, can be accessed via OSF repository ( https://doi.org/10.17605/OSF.IO/7TZCG) . The Oura PPG waveform data cannot be publicly available due to contractual agreements between National University of Singapore and Oura Health Oy.

In this work, we evaluated the feasibility of using a ring-based consumer device (Oura Ring) to assess vascular age using PPG waveforms and compared its performance to that of a clinical-grade, fingertip PPG sensor. Specifically, we tested whether ring derived PPG signals can comparably estimate vascular age to fingertip PPG whose signal waveforms are most familiar to clinicians. Critically, while both assess PPG at the finger, waveforms differ when obtained from the fingertip and around the finger. Assessments were conducted during sleep, which is associated with much reduced motion signal artifacts [ 36 ]. We first compared the pulse waveforms from two devices and characterized their differences. Next, we assessed the association strength between age and selected PPG waveform features. Finally, we predicted vascular age from pulse waveforms using deep learning and compared the prediction performance between the two devices.

The contour of the pulse waveform, both arterial and PPG-based, changes with chronological age, as increased arterial stiffness leads to faster pulse wave propagation and enhanced wave reflection from the periphery [ 19 , 20 ]. With increasing age, the PPG waveform exhibits rounder systolic peaks with a lower slope, along with a less pronounced or absent dicrotic notch [ 15 , 16 , 21 – 25 ]. This and other PPG waveform features have been proposed to assess vascular age [ 26 ], and have been utilized in regression-based models for this purpose [ 27 – 31 ]. Such feature-based models rely on accurate fiducial point detection, which may be compromised by distortions in the pulse contour. An alternative is to directly feed pulse waveforms into deep learning models, which can overcome challenges in feature extraction and have demonstrated superior prediction of vascular aging compared to feature-based models [ 32 – 34 ]. A recent landmark study using a PPG-based deep learning framework reported associations between smartwatch-derived PPG estimates and health-related outcomes, illustrating the potential for real-world application [ 35 ]. However, PPG waveforms differ across recording site and device used and understanding which PPG features contribute to age prediction is helpful in giving insight into the underlying physiological changes that drive vascular age prediction.

Wearable fitness and sleep trackers equipped with high-quality photoplethysmography (PPG) sensors are good candidates for large-scale continuous monitoring of cardiovascular health [ 14 ]. Currently, these devices primarily use PPG signals to measure pulse rate and pulse rate variability. However, high fidelity PPG waveforms can also be used to infer arterial stiffness and its age-related change [ 15 – 17 ]. This represents a major opportunity to longitudinally evaluate vascular health [ 18 ] and how lifestyle changes might affect arterial stiffness.

Vascular age is commonly estimated through a vascular risk profile that incorporates established biomarkers such as blood pressure and serum lipid levels, along with arterial stiffness. Arterial stiffness is most commonly assessed as carotid–femoral pulse wave velocity (cfPWV) using applanation tonometry [ 2 , 12 ]. A gap between vascular and chronological age suggests early or accelerated vascular aging [ 3 ]. More recently, vascular age has been predicted using only tonometer derived arterial pulse waveforms, analyzed with a deep learning model. Vascular age predicted in this manner showed strong association with the risk of adverse cardiovascular events [ 13 ]. While this latter study demonstrated the value of analyzing arterial waveforms, a tonometry-based approach uses costly, specialized equipment and is impractical for long-term, longitudinal monitoring of arterial stiffness at scale. Approaches that enable frequent and longitudinal assessment could enhance the early identification of persons at increased risk of cardiovascular disease [ 4 ].

Modifying lifestyle risk factors is important for the prevention of cardiovascular disease, the leading cause of death in industrialized societies [ 1 ]. Vascular aging refers to age-related changes in the structure and function of the arterial system, and is an important marker of cardiovascular disease (CVD) risk [ 2 – 4 ]. Arterial stiffness and blood pressure, two key determinants of vascular aging, both tend to increase with age [ 5 – 8 ]. Unhealthy lifestyles and/or environmental agents can accelerate vascular aging [ 9 – 11 ]. Conversely, premature vascular aging may be mitigated through improving diet and aerobic exercise [ 9 ]. To reinforce desirable behavioural change, it would be useful to provide individuals with regular, accurate feedback through wearable health trackers, in much the same way that continuous performance monitoring has transformed sport performance.

Model-adjusted SBP increased with increasing ΔAge grouped into tertiles ( Fig 6A ). For fingertip PPG, participants in the third (highest) tertile of ΔAge (T3) had significantly higher SBP than those in the first (lowest) tertile (T1) (mean difference = 6.11 mmHg, SE = 2.37, p = 0.032), while differences between T1–T2 and T2–T3 were not statistically significant. For ring PPG, compared with the first ΔAge tertile (T1), model-adjusted SBP was higher in both the second (T2; mean difference = 5.65 mmHg, p = 0.048) and third tertiles (T3; mean difference = 8.00 mmHg, p = 0.003). T2 and T3 did not differ significantly. Although DBP showed an increasing trend across ΔAge tertiles for both Fingertip and Ring, pairwise comparisons did not show statistically significant differences between groups ( Fig 6B ).

Grad-CAM visualizations of PPG waveforms using Fingertip (left) and Ring (right) devices. Results are displayed for different age groups across rows. Top row: all participants included, second row: ages 20–39, third row: ages 40–59, and bottom row: ages 60 and above. The colour intensity reflects the model’s attention, with warmer colours (yellow) indicating areas of higher relevance in age estimation. Across both devices, the focus areas shift with increasing age, highlighting broader regions of intermediate importance in older age groups.

Grad-CAM visualizations showed that areas of importance vary with age, with a diminished focus near the pulse onset and an expansion of intermediate-level importance regions in older age groups (ages over 60). These trends were observed for both the fingertip and the ring, suggesting that both devices capture sufficient signal features for age prediction ( Fig 5 ).

a. Scatter plot of estimated vascular age versus chronological age across cross-validation folds for all subjects for Fingertip (left column) and Ring (right column) (N = 160). Each color represents a fold, with solid lines showing fold-specific regression fits. The dashed gray line marks perfect estimation (identity line). R: Pearson’s correlation coefficient. b. Bland-Altman plot comparing estimated vascular age to chronological age. The solid grey line shows mean difference, with 95% confidence intervals (dashed grey lines) representing the limits of agreement. Fold-specific regression lines and their slope and 95% CI are shown via different colors. c. Age-adjusted ΔAge values across chronological age after proportional bias removal. Fold-specific regression lines and their slope and 95% CI are shown via different colors.

The agreement between actual and estimated age was high for both devices, with strong correlations across the 10 folds ranging from r = 0.69 to r = 0.97 for the fingertip device, and r = 0.63 to r = 0.89 for the ring ( Fig 4a ). In Bland Altman plots, the mean bias (SD) was small (Fingertip: -0.75 (8.25) years, Ring: -1.31 (9.15) years) ( Fig 4b ). However, a pattern of proportional bias was observed with increasing age, indicating underestimation in older participants and overestimation in younger participants. The bias (raw ΔAge) was significantly associated with chronological age for both devices, with slopes [95% CI] of −0.39 [−0.45, −0.33] for the fingertip and −0.46 [−0.50, −0.40] for the ring across all folds. To visualize age-independent residual variability, age-adjusted ΔAge values are additionally presented in Fig 4c . These adjusted Bland Altman plots are intended for the characterization of variability in age-independent ΔAge values and not to assess absolute agreement.

Vascular age prediction performance of deep learning model (CNN) was similar between the fingertip device and the ring, with the fingertip device showing numerically slightly lower MAE and higher R² and r values ( Table 4 ). The CNN model achieved a mean absolute error (MAE) of 6.28 ± 1.48 years (mean ± SD) for the fingertip device and 7.25 ± 1.29 years for the ring. The root mean square error (RMSE) was slightly lower for the fingertip device at 8.01 ± 1.83 years compared to 9.07 ± 1.91 years for the ring. Both devices demonstrated strong correlations between predicted and actual age, with correlation coefficients (r) of 0.84 ± 0.07 for the fingertip device and 0.80 ± 0.1 for the ring. Similarly, R² values were 0.67 ± 0.11 for the fingertip device and 0.59 ± 0.15 for the ring, indicating robust predictive performance. After adjusting for multiple comparisons, none of the differences were statistically significant between the two devices ( S4 Fig and S5 Table ).

Lastly, in regression models predicting vascular age using CT, dT and RI alone, there was no statistical difference between prediction performance between the fingertip device and the ring ( S4 Table ). The regression model had a mean absolute error (MAE) of 7.76 ± 1.49 years (mean ± SD) for the fingertip device and 8.95 ± 1.54 years for the ring. The root mean square error (RMSE) was slightly lower for the fingertip device at 9.69 ± 2 years compared to 10.8 ± 1.9 years for the ring. Both devices showed moderate correlations between predicted and actual age, with correlation coefficients (r) of 0.77 ± 0.11 for the fingertip device and 0.70 ± 0.14 for the ring. Finally, R² values were 0.60 ± 0.17 for the fingertip device and 0.52 ± 0.2 for the ring.

The linear relationship between chronological age and PPG features was in the expected direction for both devices ( Fig 3 ). The strongest correlation with age was observed for CT with coefficients of 0.64 (p < 0.001) and 0.62 (p < 0.001) for fingertip and ring devices, respectively. Association with age was moderate for dT (r = -0.51, p < 0.001) for both devices. The largest discrepancy between devices was observed for RI, as correlation was 0.58 (p < 0.001) for the fingertip sensor and 0.22 (p < 0.001) for the ring. The difference in correlation strength between the two devices was significant only for RI (Fisher’s z test p < 0.05). The observed associations for PPG features and devices remained significant after correcting for sex, BMI, SBP and DBP except for RI from the ring ( S3 Table ). The current subset of randomly sampled windows was representative of participant means, as the correlation with age was comparable across 100 different window subsets ( S3 Fig ).

To examine the noise inherent in the PPG feature extraction process, we measured the agreement in PPG features between the two devices ( Table 3 ). CT had the highest agreement between devices with correlation coefficient of 0.8, with a positive bias of 6.05 (6.16) samples indicating an overestimation for the ring. The dT and RI had similar (moderate) levels of agreement with a correlation value around 0.5, while ring overestimated RI (bias of 0.23 (0.1); S2 Fig ). The degree of agreement and bias were similar for un-normalized PPG features, CT (seconds) and dT (seconds). CT (seconds) showed the best agreement between devices ( S2 Table ).

Differences in waveforms between the fingertip sensor and ring were quantified using PPG features with values summarized in Table 2 . Values of CT and RI differed between the devices, with the ring showing higher values. In contrast, dT did not differ significantly between the fingertip and the ring. In distribution plots, RI showed the largest difference in magnitude, with a clear separation between devices ( S1 Fig ). Additionally, when PPG features were not-normalized for duration, CT (seconds) was significantly higher in the ring, while values for dT (seconds) were similar ( S1 Table ).

a. Distribution of cross-correlation coefficients between devices for all participants. b. Average waveforms (derived from all participants) from Fingertip (blue) and Ring (red). Shaded areas show standard deviation at each sample point. c. Average pulse waveforms for three age groups: 20-39, 40-59, and 60 and above. Vertical dashed blue lines indicate systolic and diastolic peak locations on Fingertip pulse; horizontal dashed lines show approximate distance between 2 peaks.

The median (IQR) cross-correlation value of PPG waveforms across all participants was 0.97 (0.02) indicating overall high similarity between devices ( Fig 2a ). Average waveforms for the fingertip sensor and ring showed that the timing of systolic and diastolic peaks were mostly aligned, with the ring having a rounder systolic peak ( Fig 2b ). However, the amplitude of the diastolic phase and peak was noticeably higher in the ring compared to the fingertip sensor. With increasing age, this systolic peak showed a rounder morphology with a shift in timing, while the diastolic peak gradually flattened, giving the waveform a more blunted appearance in both devices ( Fig 2c ).

Discussion

We found that photoplethysmography (PPG) signals from a ring-based consumer device can assess vascular age as well as a clinical-grade fingertip sensor. The pulse waveforms from the ring were of comparable quality to those from the fingertip device, despite differences in pulse contours between devices. We observed moderate to high correlations with age for CT, dT, and RI when using the fingertip probe. The strength of the correlations was similar for the ring, except for RI, which had a weaker correlation with age. Finally, we predicted vascular age for both devices using raw pulse waveforms in a CNN model, achieving similar performance for both wearable and fingertip devices suggesting that the wearable ring can assess age-related changes in arterial stiffness, comparable to a fingertip sensor.

Comparing fingertip and ring derived PPG waveforms The observed differences in pulse waveforms between devices were primarily driven by the interdependent effects of measurement site, sensor mode and design. The ring-based wearable device used in this study measures PPG on the palmar side at the base of the finger, an area with fewer capillaries and branching arteries compared to the fingertip [37], which likely reduces sensitivity to blood volume changes and contributes to delayed systolic peak timing (i.e., a greater CT value) and a higher diastolic amplitude (i.e., a greater RI). Despite the short distance between the fingertip and the base of the finger, increased reflection at the base may have also contributed to diastolic amplification and delayed systolic timing, compared to the fingertip [38]. Additionally, the ring’s rigid form may increase contact pressure during sleep when hand swelling occurs [39], causing greater variability in temporal features (including CT and dT) with increasing contact pressure levels [40]. Lastly, the fingertip device used in this study utilized a conventional transmission-type pulse oximeter, whereas the consumer wearable uses multiple infrared LEDs in both reflectance and transmission modes to collect PPG during sleep [41]. However, a comprehensive evaluation of the effect of sensor mode on pulse contour is still needed. With age, increased arterial stiffness leads to faster pulse wave propagation, reduced compliance of smaller peripheral arteries, and increased wave reflection [19], which in turn augments systolic pressure, steepens the upstroke, and flattens or blunts the dicrotic notch. While many features can be derived from PPG pulse waveforms, we focused on CT, dT, and RI to capture age-related changes, as these features are strongly linked to age related vascular stiffening [15,16,21–24,31,42], and are straightforward to interpret. Both the fingertip device and the ring showed a similar degree of linear association between age and CT, consistent with studies using data from the fingertip [15,23]. In contrast, the correlation for RI from the ring was significantly weaker than that from the fingertip, with only the latter consistent with previous reports [22]. The rigid structure of the ring may have caused variations in contact pressure throughout the night, potentially reducing RI’s sensitivity [40].

PPG-based vascular age estimation models Despite the utility of PPG features for evaluating blood vessel elasticity [26], there is currently no consensus or established guidelines on the most informative pulse characteristics for assessing vascular age. This lack of standardization becomes evident in the heterogeneity of the number and combinations of pulse waveform features used in vascular age prediction studies [30,31,33,43]. Feature-based models require precise fiducial point detection, which distortions in the pulse contour can compromise. In individuals with multiple risk factors or established cardiovascular disease, stiff arteries cause the reflected wave to merge with the direct wave, obscuring fiducial points [44]. Extracted PPG features may be even noisier with wearable data collected in real-world settings. Using extracted pulse waveforms directly in a deep learning model may circumvent some inherent limitations of PPG feature extraction. As expected, the CNN model had better prediction performance than the PPG feature-based regression model (S5 Fig), although a more detailed comparison between the two models is beyond the scope of the current work. Nonetheless, both models showed similar performance across the two devices, and to our knowledge, these are the first findings to demonstrate that wearable PPG signals can achieve accuracy comparable to fingertip PPG in vascular age prediction. Previous deep learning studies have predicted vascular age using single-site PPG from the finger [33] and nasal cavity [30,43], multiple-site PPG from the finger, wrist, arm, and ankle [32], as well as brachial, radial, and carotid tonometry waveforms [13]. Our CNN model achieved smaller prediction errors and higher correlations than previous studies using single-site PPG [34,43], though direct comparisons are limited by differences in datasets (e.g., wake vs. sleep, sensor mode) and preprocessing pipelines. While the CNN model eliminated the need for manual feature engineering, Grad-CAM visualizations revealed that both devices relied on similar waveform regions to derive vascular age information, emphasizing consistent feature importance across devices. Notably, these highlighted regions (from onset to the dicrotic notch) align with the vascular aging characteristics of waveforms described earlier. Despite the attenuated age-related association observed for RI in the ring sensor, the CNN model captured vascular age information by leveraging the full waveform characteristics.

Broadening the application of vascular age measurement to health monitoring Tonometer based vascular age inference [2–4,12] is mostly confined to specialized clinics, involving infrequent measurements performed on individuals already at risk or who have cardiovascular disease. In contrast, wearable-based arterial health assessments provide a convenient means for the collection of frequent samples in the field. As these devices are widely accessible and affordable, they may prove invaluable in providing feedback about how lifestyle choices affect arterial stiffness. Overall, our findings suggest that consumer wearables collecting high fidelity PPG data, like the Oura ring, can capture age related changes in the pulse waveform. This has significant implications for assessing vascular health unobtrusively, continuously and at scale. Consumer wearables that utilize PPG-based assessments may be instrumental in testing strategies to delay, or reverse age-associated increases in arterial stiffness by providing timely feedback about the effect of interventions. Furthermore, identifying individuals with early or accelerated vascular aging before overt CVD develops (particularly among younger or middle-aged at-risk adults who do not cross traditional treatment thresholds) could allow for pre-emptive interventions. Sleep tracking wearables in particular, offer significant advantages because sleep period is less prone motion artifacts and age-related changes in PPG waveform are more pronounced during sleep [36]. Nocturnal dipping pattern in arterial stiffness [45,46] may enable more precise characterization of vascular health compared to episodic daytime measurements, but this remains to be tested.

[END]
---
[1] Url: https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001329

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/