Wearable-derived cardiovascular fitness age and its lifestyle correlates in 442 adults

Abstract

Introduction:

Cardiovascular age is a powerful risk-communication tool that translates complex physiological data into an intuitive number, yet traditional estimates require clinical testing. Consumer wearables now estimate cardiorespiratory fitness age from photoplethysmography-derived heart rate data, enabling continuous, passive health monitoring, but whether such estimates capture substantive lifestyle variation has not been examined.

Methods:

We characterized Cardio Age, a wearable-derived cardiorespiratory fitness age estimate, in 442 Ultrahuman Ring users across a 12-month window ending February 2026 (retrospective observational cohort), separating independent lifestyle correlates from direct or indirect algorithmic inputs. Analyses used cross-sectional Spearman rank correlations, extreme-group comparisons between the youngest and oldest cardiovascular ages, and within-person longitudinal trajectories.

Results:

The mean Cardio Age gap (CA gap; mean Cardio Age minus chronological age) was −1.84 ± 2.97 years, with 82.6% of participants exhibiting younger estimated cardiovascular ages. Independent lifestyle metrics with no algorithmic link to Cardio Age showed significant associations: sleep efficiency (r = −0.194, p < 0.001), rapid eye movement (REM) sleep (r = −0.203, p < 0.001), sleep duration (r = −0.200, p < 0.001), and daily steps (r = −0.145, p = 0.003). A monotonic body mass index (BMI) dose-response was observed, with underweight participants showing a mean CA gap of −3.73 years versus −0.52 for obese participants. Extreme-group comparisons revealed that users with the youngest cardiovascular ages slept 37 minutes longer, achieved 22 more minutes of REM sleep, and had 1.8% higher sleep efficiency than those with the oldest cardiovascular ages (all p < 0.05). Sustained improvers over 12 months showed a mean CA reduction of 3.24 years; across the broader longitudinal improver and worsener groups, improving trajectories were accompanied by decreased resting heart rate (−0.8 bpm, p < 0.001) and increased estimated VO₂ max (+1.3 mL/kg/min, p < 0.001).

Discussion:

These independent lifestyle associations indicate that Cardio Age reflects physiological and behavioral variation beyond its algorithmic inputs. The within-person concordance with decreased resting heart rate and increased estimated VO₂ max, both direct algorithm inputs, is consistent with sensitivity to physiological change over time, supporting Cardio Age as a continuous, intuitive cardiovascular-fitness indicator.

1 Introduction

Cardiovascular age translates abstract risk-factor profiles into an intuitive metric that patients and the public can readily understand. The concept builds on the Framingham general cardiovascular risk profile [1], which provided the foundation for subsequent vascular age frameworks that map risk factors onto an equivalent biological age and may improve risk communication relative to percentage-based scores [2, 3]. Heart age has since been adopted in clinical guidelines and public health campaigns as a tool to personalize cardiovascular risk messaging [4, 5]. A complementary approach, fitness age, derives biological age from cardiorespiratory fitness (CRF), typically measured or estimated as maximal oxygen uptake (VO₂ max). Nes et al. [6] developed a non-exercise prediction model for cardiorespiratory fitness in the HUNT Study, and subsequently demonstrated that estimated CRF predicts all-cause and cardiovascular mortality in over 37,000 adults [7]. Large population studies have confirmed that CRF predicts mortality independently of traditional risk factors [8, 9], reinforcing CRF as a clinical vital sign [10].

Traditionally, cardiovascular and fitness age estimation requires clinical exercise testing or laboratory-grade VO₂ max measurement, limiting scalability. Consumer wearable devices primarily estimate VO₂ max passively from photoplethysmography (PPG)-derived heart rate signals [11, 12], enabling daily monitoring without clinic visits. This has expanded the scope of wearable health monitoring beyond step counting to continuous physiological assessment [13]. Several commercial wearable platforms now translate wearable-derived estimated VO₂ max into a fitness age metric [14, 15]. This shift from episodic clinical assessment to continuous passive monitoring enables users to track their cardiovascular fitness age daily, monitoring fundamental health metrics (resting heart rate, heart rate variability, and estimated VO₂ max) through a single age-equivalent rather than abstract physiological values.

Prior validation studies have focused primarily on the accuracy of wearable heart rate and physiological measurements [16, 17], including wearable VO₂ max estimates, where notable accuracy limitations have been reported [18]. Substantial evidence links lifestyle factors such as sleep duration [19, 20], body mass index (BMI) [21], and physical activity [22] to cardiovascular health. Although these factors are established determinants of cardiovascular health, whether they are captured by a wearable-derived fitness age metric, and whether such associations persist across observation windows, has not been systematically examined. To our knowledge, no study has tested whether within-person longitudinal trajectories of wearable cardiovascular age track measurable physiological changes. These gaps matter because wearable fitness age metrics are increasingly available to millions of consumers, and demonstrating that they reflect genuine physiological and lifestyle variation, not just algorithmic artifacts, is essential for establishing their scientific credibility and potential utility. A further analytical challenge is algorithmic circularity: because wearable fitness age is derived from VO₂ max, which itself depends on resting heart rate, correlations between these constituent inputs and the output metric are partly tautological. Any characterization study must explicitly separate constituent from independent metrics to avoid overstating the novelty of circular associations.

Here we examined Cardio Age, a cardiorespiratory fitness age estimate derived from the Ultrahuman Ring, in 442 adult users across a 12-month observation window. Our aims were threefold: (a) characterize the distribution of the Cardio Age gap, defined as estimated cardiovascular age minus chronological age; (b) identify independent lifestyle correlates (metrics with no direct algorithmic link) to establish that Cardio Age captures physiological variation beyond its constituent inputs; and (c) examine within-person longitudinal trajectories to determine whether Cardio Age tracks real physiological improvements over time, supporting its utility as a practical tool for cardiovascular fitness monitoring.

2 Results

2.1 Cardio age gap distribution

The mean CA gap was years (SD 2.97), with a median of (IQR to ). A total of 82.6% of participants had negative CA gaps (younger estimated cardiovascular age), while 16.7% had positive gaps (older estimated cardiovascular age), and the remaining 0.7% (3 participants) had an estimated cardiovascular age equal to their chronological age. The distribution was left-skewed (Figure 1). The preponderance of younger hearts likely reflects self-selection bias, as users who purchase and consistently wear a fitness-oriented ring are plausibly more health-conscious than the general population [23].

Figure 1

2.2 Independent lifestyle correlates

To identify modifiable behaviors associated with cardiovascular fitness age beyond the algorithm’s direct inputs, we examined metrics with no direct algorithmic pathway (all study variables, with units, tier classification, and within-Ultrahuman validation references, are defined in Supplementary Table S1). Stratified by gender, the Cardio Age gap distribution was markedly shifted (Female median vs. Male median years; Mann–Whitney ; Cohen’s ; Figure 1); within each gender the gap was approximately flat across chronological age, so the cohort-level age correlation is largely a gender-composition effect (see Discussion). The Youngest Hearts group was 80.0% female, while the Oldest Hearts group was 86.7% male, consistent with sex-specific differences in the algorithm’s normative CRF-age equations [24]; these extreme groups are defined by the 10th and 90th percentiles of the Cardio Age gap distribution, not by chronological age.

Among independent lifestyle metrics, the Youngest Hearts and Oldest Hearts groups differed significantly on sleep efficiency (90.4% vs. 88.6%, median difference , ), REM sleep (120.5 vs. 98.3 min, , ), sleep duration (454.6 vs. 417.4 min, , ), and daily steps (7221 vs. 5540, , ). Deep sleep duration was not statistically significant (Supplementary Table S2; Figure 2).

Figure 2

Spearman rank correlations across the full cohort confirmed the extreme-group findings (Supplementary Table S3; Figure 3). The strongest independent correlates were REM sleep (, 95% CI to , ), sleep duration (, to , ), sleep efficiency (, to , ), and daily steps (, to , ). All three sleep metrics survived Bonferroni correction for 14 tests (); daily steps also survived (). Deep sleep (, ) was not significant. BMI (, 0.184–0.365, ) and weight (, 0.227–0.393, ) were strongly correlated but are classified as indirect constituents because BMI is a known physiological determinant of VO₂ max (the CA algorithm’s primary input); these associations are therefore partly mediated through the algorithm. Recovery score (, to , ) was also strongly correlated but shares heart rate variability (HRV) and resting heart rate components with the algorithm.

Figure 3

A monotonic dose-response relationship was observed between BMI category and CA gap (Figure 4). Mean CA gap was years for underweight participants (), for normal weight (), for overweight (), and for obese (). Self-reported mobility level was also associated with CA gap: sedentary years, moderate , active . Mean BMI and the Cardio Age gap both rose with chronological age through the 50–59 decade and then declined in the oldest stratum (60+ years; Supplementary Table S4).

Figure 4

2.3 Constituent metric profiles and algorithmic transparency

The Cardio Age algorithm uses three tiers of inputs. Direct constituent inputs (flagged ‡) are the foundational algorithm inputs: nighttime resting heart rate, HRV root mean square of successive differences (RMSSD), and estimated VO₂ max. Indirect constituents (flagged †) are partly mediated through the algorithm: recovery score, stress rhythm score, BMI, and weight. Independent lifestyle correlates have no direct algorithmic link: sleep duration, sleep efficiency, deep sleep, REM sleep, daily steps, active hours, and movement index.

In extreme-group comparisons, the Youngest Hearts group had substantially lower nighttime resting heart rate‡ (47.9 vs. 61.7 bpm, ), higher HRV‡ (RMSSD 55.9 vs. 37.2 ms, ), and higher VO₂ max‡ (48.7 vs. 36.4 mL/kg/min, ). These differences are expected given the algorithmic relationship and are presented for completeness rather than as novel findings; they illustrate how Cardio Age translates abstract physiological metrics into a single age-equivalent. Recovery score† also differed significantly (74.2 vs. 68.1, ), as expected given its shared HRV and resting heart rate components. BMI† (22.8 vs. 29.6 kg/, ) and weight† (69.0 vs. 89.0 kg, ) also differed substantially between groups (Supplementary Table S2).

2.4 Longitudinal trajectories

To examine whether within-person changes in Cardio Age correspond to physiological shifts, we classified users by their trajectory direction over 12 months using 30-day windows at the start and end of the observation period.

Using a threshold of years, 52 participants were classified as sustained large improvers (mean delta years) and 59 as sustained large worseners (mean delta years; Supplementary Table S5; Figure 5). The two groups were similar in age (33.1 vs. 35.4 years) and BMI (26.4 vs. 26.8 kg/). Across the broader longitudinal improver and worsener groups (Supplementary Table S6), improving trajectories were accompanied by decreased resting heart rate‡ ( bpm vs.  bpm, ) and increased VO₂ max‡ ( vs.  mL/kg/min, ). Temporal divergence in constituent metrics between sustained improver and worsener groups became visible by months 3–4 (Supplementary Figure S1). Within-person Cardio Age was highly consistent over 12 months (median SD 1.21 years; 37.3% of users with SD years).

Figure 5

3 Discussion

We characterized the distribution, lifestyle correlates, and longitudinal trajectories of Cardio Age, a wearable-derived cardiorespiratory fitness age estimate, in 442 Ultrahuman Ring users across a 12-month observation window. The principal findings were that the strongest independent correlates of the Cardio Age gap were REM sleep, sleep duration, sleep efficiency, and daily steps, all metrics with no direct algorithmic link to Cardio Age. BMI and weight, classified as indirect constituents, showed associations consistent with prior epidemiological findings linking higher BMI to lower cardiorespiratory fitness [9, 22, 25] and to increased lifetime cardiovascular disease risk [21]. Within-person Cardio Age was highly consistent over time (median SD 1.21 years). Sustained improver and worsener trajectories were resolved over 12 months with bidirectional within-person sensitivity; because resting heart rate and estimated VO₂ max are direct algorithm inputs, this concordance is expected by design, and the circularity considerations and non-tautological features of the trajectory analysis are discussed below.

Although numerous cardiovascular age concepts exist, several features support the relevance of the Ultrahuman Cardio Age metric specifically. Cardiorespiratory fitness, as quantified by VO₂ max, is a well-documented independent predictor of all-cause and cardiovascular mortality [7–9]. An algorithm that estimates VO₂ max from wearable data therefore offers a practical bridge between clinical fitness testing and everyday wellness monitoring. The Ultrahuman CA algorithm draws on established cardiorespiratory fitness estimation frameworks [24, 26], adapting laboratory-grade physiological foundations for a consumer wearable context. The cross-sectional correlates we observe (sleep, BMI as indirect constituent) align with known determinants of cardiorespiratory fitness in the epidemiological literature, providing face validity. Within-person stability (median SD 1.21 years over 12 months) confirms the metric captures genuine individual-level physiological variation rather than measurement noise.

The observed associations are consistent with known physiological relationships. The inverse relationship between BMI and CRF is well documented [9, 22, 25], higher BMI is associated with increased lifetime cardiovascular risk [21], and our finding that higher BMI is associated with older Cardio Age extends these relationships to a consumer wearable context. Sleep duration has been linked to cardiovascular health in large prospective studies; short sleep increases the risk of all-cause mortality and coronary heart disease [19, 20, 27]. The association between REM sleep and younger Cardio Age is consistent with evidence linking sleep architecture to autonomic regulation and cardiovascular health [28]. Recovery score, a composite autonomic measure incorporating resting heart rate and HRV, was classified as an indirect constituent because it shares algorithmic components with Cardio Age. Its strong association with CA gap () aligns with a growing literature on HRV-based recovery monitoring [29, 30].

The fitness age concept underlying our work was established by Nes et al. [6, 7], who demonstrated that estimated cardiorespiratory fitness predicts all-cause mortality in over 37,000 adults. Our findings extend this framework from single-timepoint clinical assessments to continuous daily monitoring, providing evidence that a wearable-derived fitness age metric reflects lifestyle and physiological variation beyond its algorithmic inputs over a 12-month period.

Algorithmic circularity is an important consideration for interpreting these results. Nighttime resting heart rate is a key input to the Cardio Age algorithm, which estimates VO₂ max using a proprietary combination of validated methods [24, 26]. Correlations between these constituent metrics and CA gap are therefore expected, and we present them as descriptive characterization. BMI is an indirect constituent: it is not a direct algorithm input, but BMI is a known physiological determinant of VO₂ max [9], so the BMI-CA gap association is partly mediated through the algorithm’s primary input. After flagging these layers, the remaining independent correlates (sleep duration, REM sleep, sleep efficiency, and daily steps) represent associations with no direct algorithmic link. The fact that these independent lifestyle metrics correlated with CA gap despite having no algorithmic pathway suggests that Cardio Age captures physiological variation beyond its constituent inputs. These associations were also stable across 3-, 6-, and 12-month observation windows, with sleep duration, sleep efficiency, and REM sleep showing consistent correlation magnitudes across all three horizons. The same circularity considerations apply to the within-person trajectory analysis (Section 2.4): directional concordance between Cardio Age and its direct inputs over time is mechanistic rather than independent evidence, while the within-person stability (median SD 1.21 years), bidirectional sensitivity, and the months-3–4 horizon over which sustained change resolves are non-tautological empirical features. A multivariate model including BMI, sleep duration, sleep efficiency, gender, and age explained 30.1% of CA gap variance (Supplementary Table S7); in that joint model sleep efficiency retained a significant residual association with the Cardio Age gap (, ), whereas sleep duration became non-significant (, ); BMI and gender were the dominant predictors, and age was non-significant once gender and BMI were included. REM sleep was not entered as a predictor in this joint model; the sleep metrics differed substantially between female and male participants (Cohen’s 0.47–0.66, including REM sleep ; Supplementary Figure S2), so the cohort-level sleep-duration and REM associations are largely shared with gender and BMI, while sleep efficiency, which shows no meaningful gender difference, survives adjustment. The extreme-group sleep contrasts are therefore partly gender-composition effects. The female–male shift in the Cardio Age gap reflects the algorithm’s sex-specific normative reference [24] rather than an intrinsic cardiovascular difference between male and female participants, and the apparent cohort-level Cardio Age gap to chronological age trend is partly mediated by the age–BMI covariation in this cohort. The within-person longitudinal analyses (Section 2.4) are unaffected by these between-person compositional structures.

A moderate positive correlation was observed between CA gap and chronological age (, ), indicating that older participants tended to have more positive (less favorable) CA gaps. This pattern could reflect genuine age-related cardiovascular fitness decline, or it could indicate calibration differences in the normative CRF-age reference data [24] relative to this cohort’s demographic composition. The two explanations are not mutually exclusive; disentangling self-selection from algorithmic calibration would require external validation against laboratory-measured VO₂ max, which is beyond the scope of this characterization study.

The within-person trajectory analysis over 12 months characterized the metric’s temporal behavior: improvements were accompanied by decreased resting heart rate and increased estimated VO₂ max (directional changes that are mechanistically expected given the algorithmic relationship), with divergence between improver and worsener groups becoming visible by months 3–4. The physiological mechanism most plausibly captured by Cardio Age is autonomic adaptation to aerobic exercise. Regular endurance training increases vagal tone, reducing resting heart rate [31], a key input to the Cardio Age algorithm. This autonomic adaptation pathway is consistent with the trajectory patterns observed and is consistent with Cardio Age functioning as a continuous fitness metric that reflects cardiovascular conditioning. Sleep duration may correlate cross-sectionally with CA gap as a marker of overall health status (individuals who sleep more may have lower allostatic load and more time for exercise) rather than as a direct driver of resting heart rate changes.

The sustained worsener group (, mean delta years) was slightly larger than the improver group () and was 61% female compared with 50% in improvers. Several mechanisms could underlie worsening trajectories: genuine physiological decline (e.g., reduced physical activity, weight gain, or increased allostatic load), regression to the mean for participants who started with unusually favorable CA gap values, or seasonal or behavioral changes over time. The two groups were demographically similar at baseline (age 33.1 vs. 35.4 years; BMI 26.4 vs. 26.8 kg/), suggesting that worsening was not simply a function of older or less fit starting points. This symmetric treatment of worsening and improving trajectories strengthens the interpretation that Cardio Age tracks bidirectional physiological change.

Heart age has been proposed as a motivational tool for health behavior change [3, 4], but evidence has been limited to single-timepoint clinical assessments. Cardio Age extends this concept by providing continuous, daily feedback. The sustained improver cohort achieved a mean 3.24-year reduction in cardiovascular fitness age over 12 months without any structured intervention. We hypothesize that continuous passive feedback translating physiological metrics into an interpretable age-equivalent may support sustained user engagement with cardiovascular health, but this study did not measure engagement directly; prospective designs (such as a randomized comparison of Cardio Age visibility on/off, or matched cohorts with and without access to the metric) would be needed to test this hypothesis. Cardio Age also addresses a gap in how individuals interact with their health data. A decrease in resting heart rate of 0.8 bpm or an increase in estimated VO₂ max of 1.3 mL/kg/min is difficult for the average user to interpret or act upon; by contrast, a three-year decrease in cardiovascular age provides an immediately intuitive signal. The translation of abstract physiological values into a personal, age-based metric is what may make Cardio Age actionable at scale. When users can see that their daily behaviors (sleep habits, physical activity, weight management) are reflected in a single number they can track over time, the perceived feedback loop between behavior and outcome becomes tangible, potentially supporting sustained engagement with personalized health monitoring. The dose-response relationship between BMI and CA gap, the significant associations with sleep metrics and daily steps, and the trajectory data showing sustained improvement all point to a metric that may serve as a practical indicator for everyday cardiovascular health awareness at the population level.

The translation of multi-metric physiological data into a single, interpretable age-equivalent is part of a broader participatory health paradigm in which individuals engage actively with their own physiological data rather than receive episodic clinical assessment [32]. Behavioral-change literature on wearable self-monitoring reports that consumer activity trackers support increases in physical activity through self-monitoring and feedback mechanisms [33], with effects of clinically meaningful magnitude that can be sustained over time [34]. Whether translating multi-metric physiological data into a single age-equivalent metric strengthens or sustains the behavioral-change signal beyond what is achievable with individual-metric feedback is an open empirical question that the present observational design does not resolve.

Several methodological strengths support confidence in these findings: daily longitudinal data from 442 users rather than single-timepoint assessment, transparent separation of constituent and independent metrics, and within-person stability analysis over 12 months. Limitations include: (1) Algorithmic circularity: constituent metrics are direct or indirect algorithm inputs, as fully disclosed above. (2) Sample characteristics: participants are self-selected wearable users (median age 31 years) with self-reported anthropometrics; however, these characteristics are representative of the early-adopter population to whom the metric is delivered. (3) Observational design: associations do not imply causation; however, within-person longitudinal analyses and concordance with prior epidemiological evidence provide convergent support. (4) External validation: no comparison to laboratory-measured VO₂ max or cardiopulmonary exercise testing (CPET)-derived fitness age was performed within the present study; this work should be interpreted as a characterization study establishing that the metric reflects genuine physiological and lifestyle variation. (5) Device-specific validation status of analyzed metrics. The validation status of each analyzed metric is summarized in Supplementary Table S1. Nighttime resting heart rate [35], peripheral skin temperature [36, 37], and the sleep-timing measures (total sleep time, sleep onset, sleep offset) [38] have been independently validated for the Ring AIR against research- or clinical-grade comparators. For estimated VO₂ max and heart rate variability, the present analyses rely on within-person longitudinal trajectories rather than absolute values; external validation of the estimated VO₂ max algorithm against laboratory CPET is the subject of a separate validation study in preparation. Deep and REM sleep durations both feed the Ultrahuman Sleep Score, characterized externally as a behaviorally meaningful composite at population scale [39]. (6) Health status, medications, and shift work. No participant reported a disease at onboarding, but onboarding does not separately record current medications or shift-work patterns, both of which can influence the algorithmic inputs to Cardio Age.

The cohort spans a wide chronological-age range (18–77 years, IQR 26–42) with a median of 31 years, weighting these findings toward the preventive-health context while retaining substantial representation across adulthood. Lifestyle correlates of cardiorespiratory fitness in younger adults are increasingly relevant as cardiovascular and metabolic risk emerge at younger ages, and the behavioral levers identified here (sleep, physical activity, BMI) align with the core constructs of the American Heart Association’s Life’s Essential 8 cardiovascular-health framework [40]. Translating these findings to older populations requires further work, since older adults face additional barriers to consumer-wearable adoption including digital-literacy gaps and different baseline health profiles, and whether continuous wearable feedback supports engagement with cardiovascular health equally well in older cohorts remains an open question. Future work should pursue prospective validation linking Cardio Age to cardiovascular outcomes, algorithm refinement incorporating sleep and recovery directly, cross-device generalizability studies, and examination of intervention effects on Cardio Age trajectories.

4 Methods

4.1 Study design and participants

This was a retrospective observational cohort study of 442 adult users of the Ultrahuman Ring AIR smart ring. Inclusion criteria were age 18 years and at least one day of Cardio Age data between March 1, 2025, and February 28, 2026. One participant aged <18 was excluded from the initial dataset of 443 users. The primary analysis window was 12 months (March 1, 2025, to February 28, 2026; 365 days). Demographic data, including age, sex, height, weight, and self-reported mobility level, were collected while onboarding on the Ultrahuman platform. No participant reported a disease at onboarding, and onboarding does not separately collect medication use or shift-work status.

The study cohort had a median age of 31 years (interquartile range [IQR] 26–42; range 18–77). The sample was 65.4% female (), 32.6% male (), and 2.0% other (). Median BMI was 25.0 kg/ (IQR 22.2–28.9), with 17 participants classified as underweight (3.8%), 207 as normal weight (46.8%), 126 as overweight (28.5%), and 92 as obese (20.8%). Self-reported mobility level was sedentary in 112 participants (25.3%), moderate in 272 (61.5%), and active in 58 (13.1%). The median number of observation days per user was 352.5 (IQR 267–360.8) for the 12-month window. Full cohort characteristics are presented in Table 1.

Self-reported mobility level was selected by users at onboarding from three categories reflecting broad self-assessed habitual physical activity: Sedentary (little to no regular exercise), Moderate (regular light-to-moderate activity), and Active (regular vigorous or structured exercise, or physically demanding occupational activity). Categorization is at the user’s discretion. The categorical self-report is used by the Cardio Age algorithm as an adjustment factor in the VO₂ max to fitness-age translation step (Methods Section 4.4): more active categories shift the estimate toward a younger Cardio Age, all else equal. Mobility level enters as an adjustment factor rather than as a primary input.

4.2 Ethics and data governance

This was a real-world, retrospective, observational study based on data derived from Ultrahuman platform users and adhered to Ultrahuman’s terms of use [41] and privacy policy [42], which allows for analysis of de-identified grouped data for scientific research. Participants consented via the onboarding process on the Ultrahuman platform and continued product use. As the study was non-invasive and involved no dietary, sleep, or exercise interventions, with all reported data de-identified, explicit institutional ethics board approval was not required. A separate set of analysts extracted the data, ran computational approaches, and then reviewed the results to ensure blinding.

4.3 Device

The Ultrahuman Ring AIR is a titanium finger-worn smart ring equipped with a green-light photoplethysmography (PPG) sensor for continuous heart rate monitoring, a tri-axis accelerometer for motion and activity detection, and a skin temperature sensor. The ring form factor enables continuous, unobtrusive physiological data collection during daily activities and sleep without requiring user interaction.

4.4 Cardio age algorithm

Cardio Age is a proprietary cardiorespiratory fitness age estimate. Estimated VO₂ max is derived from a proprietary weighted combination of established cardiorespiratory fitness estimation methods [24, 26], incorporating nighttime resting heart rate as a foundational input. The estimated VO₂ max is then converted to a fitness age using sex-specific normative CRF-age data [24], with temporal smoothing applied to limit day-to-day variation. Cardio Age is estimated daily for each user. Self-reported mobility level influences the estimation as an adjustment factor; see Methods Section 4.1 for category definitions. These relationships create layers of algorithmic circularity that are explicitly addressed in our analytical framework.

4.5 Outcome and predictor variables

The primary outcome was the Cardio Age gap (CA gap), defined as the mean of all available daily Cardio Age values across the observation window minus chronological age. Negative values indicate a cardiovascular system estimated to be younger than chronological age. Per-user values for each wearable-derived metric were computed as the mean of all available daily observations within the observation window.

Constituent metrics (direct algorithm inputs, flagged with ‡ throughout) were nighttime resting heart rate (bpm), HRV RMSSD (ms), and estimated VO₂ max (mL/kg/min). Indirect constituent metrics (flagged with †) were recovery score, stress rhythm score, BMI (kg/), and weight (kg). Recovery score and stress rhythm score share HRV and resting heart rate components with the Cardio Age algorithm. BMI and weight are not direct algorithm inputs but are known physiological determinants of VO₂ max [9], so associations between these metrics and CA gap are partly mediated through the algorithm’s primary input.

Independent lifestyle correlates included sleep duration (min), sleep efficiency (%), deep sleep duration (min), REM sleep duration (min), daily steps, active hours, and movement index.

Sleep efficiency was defined as the ratio of total sleep time to total time in bed, expressed as a percentage. A consolidated definitions table covering every variable used in this study, with unit and within-Ultrahuman validation reference where applicable, is provided as Supplementary Table S1. Ultrahuman-proprietary composite metrics were defined as follows. Recovery score (0–100) is a proprietary daily score from nighttime resting heart rate, heart rate variability, skin temperature, and Stress Rhythm Score, with additive contributions from breathwork, non-sleep deep rest, and nap sessions; higher values reflect greater readiness. Because it shares resting heart rate and heart rate variability components with the Cardio Age algorithm, it was classified as an indirect constituent†; sleep duration, sleep efficiency, deep sleep duration, and REM sleep duration are not inputs to the Recovery Score, and their classification as independent lifestyle correlates is preserved. Stress rhythm score (0–100) is a proprietary daily score from nighttime resting heart rate, heart rate variability, and the daily pattern of heart-rate excursions above resting baseline contextualized by circadian phase. It similarly shares algorithmic inputs with Cardio Age†. Movement index (0–100) is a proprietary daily score from active hours, inactive waking time, daily steps relative to a personalized goal, and weekly workout frequency and volume. Active hours are waking hours with sufficient step activity, excluding sleep, naps, and non-wear; workout periods are auto-classified as active. Skin temperature is measured by the Ring AIR’s peripheral temperature sensor at 5 min intervals and feeds the Recovery Score composite; it is not an analytical predictor of Cardio Age gap in this study. These metrics are platform-derived composites and should not be equated with validated clinical instruments. Device-specific validation references for nighttime resting heart rate [35], peripheral skin temperature [36, 37], and the sleep-timing measures [38] are detailed in Supplementary Table S1; implications of remaining validation gaps for the present interpretation are discussed in Section 3.

Extreme groups were defined using the 10th and 90th percentiles of the 12-month CA gap distribution: Youngest Hearts (CA gap years, ) and Oldest Hearts (CA gap years, ). This analytical choice enables comparison of users with the most divergent cardiovascular fitness ages.

Longitudinal change was quantified as delta CA: last 30-day mean minus first 30-day mean CA over the 12-month window. Thirty-day windows were used to provide robust estimates and mitigate regression to the mean. Improvers were defined as delta years; worseners as delta . These thresholds were chosen to exceed the within-person measurement variability (median SD 1.21 years), ensuring that classified changes exceed typical metric noise.

4.6 Statistical analysis

Continuous variables are reported as median (IQR); categorical variables as (%). Extreme-group comparisons used Mann-Whitney tests (two-tailed) with median difference and 95% confidence interval. Spearman rank correlations were computed with bootstrapped 95% confidence intervals (1,000 iterations). Longitudinal within-person changes were assessed by Wilcoxon signed-rank tests within trajectory groups and Kruskal-Wallis tests between trajectory groups. Recovery score sentinel values () were converted to missing before aggregation. Statistical significance was set at (two-tailed). For the 14 correlations reported in Supplementary Table S3 (constituent, indirect constituent, and independent metrics), a Bonferroni-corrected threshold of was applied; results are reported at both nominal and corrected thresholds. Exact -values are reported throughout. Analyses were performed in Python 3.12 [pandas 2.2 [43], scipy 1.12 [44], statsmodels 0.14 [45], scikit-learn 1.4 [46]].

References

• 1.

  D’AgostinoRBVasanRSPencinaMJWolfPACobainMMassaroJM, et al. General cardiovascular risk profile for use in primary care: the Framingham heart study. Circulation. (2008) 117:743–53. 10.1161/CIRCULATIONAHA.107.699579

  • CrossRef
  • Google Scholar
  • View reference in article

• 2.

  CuendeJI. Vascular age versus cardiovascular risk: clarifying concepts. Rev Esp Cardiol. (2016) 69:243–6. 10.1016/j.rec.2015.10.019

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3.

  GroenewegenKAPasterkampGPolakJFBotsMLPetersSAE. Vascular age to determine cardiovascular disease risk: a systematic review of its concepts, definitions, and clinical applications. Eur J Prev Cardiol. (2016) 23:264–74. 10.1177/2047487314566999

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4.

  BonnerCBatcupCCornellSFajardoMAHawkesALTrevenaL, et al. Interventions using heart age for cardiovascular disease risk communication: systematic review of psychological, behavioral, and clinical effects. JMIR Cardio. (2021) 5:e31056. 10.2196/31056

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5.

  PiepoliMFHoesAWAgewallSAlbusCBrotonsCCatapanoAL, et al. 2016 European guidelines on cardiovascular disease prevention in clinical practice. Eur Heart J. (2016) 37:2315–81. 10.1093/eurheartj/ehw106

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6.

  NesBMJanszkyIVattenLJNilsenTILAspenesSTWisløffU. Estimating VO2peak from a nonexercise prediction model: the HUNT Study, Norway. Med Sci Sports Exerc. (2011) 43:2024–30. 10.1249/MSS.0b013e31821d3f6f

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7.

  NesBMVattenLJNaumanJJanszkyIWisløffU. A simple nonexercise model of cardiorespiratory fitness predicts long-term mortality. Med Sci Sports Exerc. (2014) 46:1159–65. 10.1249/MSS.0000000000000219

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8.

  KodamaSSaitoKTanakaSMakiMYachiYAsumiM, et al. Cardiorespiratory fitness as a quantitative predictor of all-cause mortality and cardiovascular events in healthy men and women: a meta-analysis. JAMA. (2009) 301:2024–35. 10.1001/jama.2009.681

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9.

  RossRBlairSNArenaRChurchTSDesprésJ-PFranklinBA, et al. Importance of assessing cardiorespiratory fitness in clinical practice: a case for fitness as a clinical vital sign. Circulation. (2016) 134:e653–99. 10.1161/CIR.0000000000000461

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10.

  KaminskyLAArenaRBeckieTMBrubakerPHChurchTSFormanDE, et al. The importance of cardiorespiratory fitness in the United States: the need for a national registry. Circulation. (2013) 127:652–62. 10.1161/CIR.0b013e31827ee100

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11.

  AllenJ. Photoplethysmography and its application in clinical physiological measurement. Physiol Meas. (2007) 28:1–39. 10.1088/0967-3334/28/3/R01

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12.

  ShcherbinaAMattssonCMWaggottDStalFMattssonODarberD, et al. Accuracy in wrist-worn, sensor-based measurements of heart rate and energy expenditure in a diverse cohort. J Pers Med. (2017) 7:3. 10.3390/jpm7020003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13.

  DunnJRungeRSnyderM. Wearables and the medical revolution. Per Med. (2018) 15:429–48. 10.2217/pme-2018-0044

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14.

  Apple Inc. Data from: Using Apple Watch to estimate cardio fitness with max. Apple Healthcare Technical Document (2021). Available online at: https://www.apple.com/healthcare/docs/site/Using_Apple_Watch_to_Estimate_Cardio_Fitness_with_VO2_max.pdf (Accessed May 28, 2026).

  • View reference in article

• 15.

  Garmin. Data from: What does fitness age mean in Garmin Connect? Garmin Customer Support (2024). Available online at: https://support.garmin.com/en-US/?faq=CM1YJmMrrNAbEpM9PapJ07 (Accessed May 28, 2026).

  • View reference in article

• 16.

  BentBGoldsteinBAKibbeWADunnJP. Investigating sources of inaccuracy in wearable optical heart rate sensors. NPJ Digit Med. (2020) 3:18. 10.1038/s41746-020-0226-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17.

  FullerDColwellELowJOrychockKTobinMASimangoB, et al. Reliability and validity of commercially available wearable devices for measuring steps, energy expenditure, and heart rate: systematic review. JMIR Mhealth Uhealth. (2020) 8:e18694. 10.2196/18694

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18.

  PasslerSBohrerJBlochingerLSennerV. Validity of wrist-worn activity trackers for estimating VO2max and energy expenditure. Int J Environ Res Public Health. (2019) 16:3037. 10.3390/ijerph16173037

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19.

  CappuccioFPD’EliaLStrazzulloPMillerMA. Sleep duration and all-cause mortality: a systematic review and meta-analysis of prospective studies. Sleep. (2010) 33:585–92. 10.1093/sleep/33.5.585

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20.

  St-OngeM-PGrandnerMABrownDConroyMBJean-LouisGCoonsM, et al. Sleep duration and quality: impact on lifestyle behaviors and cardiometabolic health: a scientific statement from the American Heart Association. Circulation. (2016) 134:e367–86. 10.1161/CIR.0000000000000444

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21.

  KhanSSNingHWilkinsJTAllenNCarnethonMBerryJD, et al. Association of body mass index with lifetime risk of cardiovascular disease and compression of morbidity. JAMA Cardiol. (2018) 3:280–7. 10.1001/jamacardio.2018.0022

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22.

  LavieCJOzemekCCarboneSKatzmarzykPTBlairSN. Sedentary behavior, exercise, and cardiovascular health. Circ Res. (2019) 124:799–815. 10.1161/CIRCRESAHA.118.312669

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23.

  PiwekLEllisDAAndrewsSJoinsonA. The rise of consumer health wearables: promises and barriers. PLoS Med. (2016) 13:e1001953. 10.1371/journal.pmed.1001953

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24.

  JacksonASSuiXHébertJRChurchTSBlairSN. Role of lifestyle and aging on the longitudinal change in cardiorespiratory fitness. Arch Intern Med. (2009) 169:1781–7. 10.1001/archinternmed.2009.312

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25.

  WeiMKampertJBBarlowCENichamanMZGibbonsLWPaffenbargerRS, et al. Relationship between low cardiorespiratory fitness and mortality in normal-weight, overweight, and obese men. JAMA. (1999) 282:1547–53. 10.1001/jama.282.16.1547

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26.

  TanakaHMonahanKDSealsDR. Age-predicted maximal heart rate revisited. J Am Coll Cardiol. (2001) 37:153–6. 10.1016/S0735-1097(00)01054-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27.

  LaoXQLiuXDengH-BChanT-CHoKFWangF, et al. Sleep quality, sleep duration, and the risk of coronary heart disease: a prospective cohort study with 60,586 adults. J Clin Sleep Med. (2018) 14:109–17. 10.5664/jcsm.6894

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28.

  TobaldiniECostantinoGSolbiatiMCogliatiCKaraTNobiliL, et al. Sleep, sleep deprivation, autonomic nervous system and cardiovascular diseases. Neurosci Biobehav Rev. (2017) 74:321–9. 10.1016/j.neubiorev.2016.07.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29.

  PlewsDJLaursenPBStanleyJKildingAEBuchheitM. Training adaptation and heart rate variability in elite endurance athletes: opening the door to effective monitoring. Sports Med. (2013) 43:773–81. 10.1007/s40279-013-0071-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 30.

  ShafferFGinsbergJP. An overview of heart rate variability metrics and norms. Front Public Health. (2017) 5:258. 10.3389/fpubh.2017.00258

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31.

  CarterJBBanisterEWBlaberAP. Effect of endurance exercise on autonomic control of heart rate. Sports Med. (2003) 33:33–46. 10.2165/00007256-200333010-00003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 32.

  HoodLAuffrayC. Participatory medicine: a driving force for revolutionizing healthcare. Genome Med. (2013) 5:110. 10.1186/gm514

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33.

  BrickwoodK-JWatsonGO’BrienJWilliamsAD. Consumer-based wearable activity trackers increase physical activity participation: systematic review and meta-analysis. JMIR Mhealth Uhealth. (2019) 7:e11819. 10.2196/11819

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34.

  FergusonTOldsTCurtisRBlakeHCrozierAJDankiwK, et al. Effectiveness of wearable activity trackers to increase physical activity and improve health: a systematic review of systematic reviews and meta-analyses. Lancet Digit Health. (2022) 4:e615–26. 10.1016/S2589-7500(22)00111-X

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 35.

  KrishnanNNageshRSinghPAsudaniVSinghalVSrinivasanB, et al. Data from: Sleep heart rate sensing by Ultrahuman Ring AIR demonstrates high overlap with FDA-approved device and consumer-grade wearable. Ultrahuman Science Whitepaper (2024). Available online at: https://ultrahuman.com/science/studies/sleep-heart-rate-sensing (Accessed May 28, 2026).

  • Google Scholar
  • View reference in article

• 36.

  GuptaKPirrieAShanmugamADaviesKHendiAMilnesR, et al. Data from: Listening in: introducing the ultrahuman cycle tracking & ovulation pro (C&O Pro) feature powered by OvuSense technology. Ultrahuman Science Whitepaper (2025). (Accessed May 29, 2026). Sensor equivalence Table 1: Ring AIR skin temperature vs OvuSense SWS (CE Class IIa); bias C, MAE C. Available online at: https://ultrahuman.com/science/studies/cycle-tracking-pro-accuracy

  • Google Scholar
  • View reference in article

• 37.

  ShanmugamAPrabhakaranPAsudaniVJoshiASurajAPrasathA, et al. Data from: Comparison of Ultrahuman ring AIR temperature sensing at device and hand level reveals superior overlap with gold standards. Ultrahuman Science Whitepaper (2025). Available online at: https://ultrahuman.com/science/studies/ultrahuman-ring-air-temperature-sensing (Accessed May 28, 2026).

  • Google Scholar
  • View reference in article

• 38.

  DrillerMBodnerMEFenutaAStevensonSSuppiahH. How does the sleep regularity questionnaire relate to wearable- and diary-derived sleep regularity?J Circadian Rhythms. (2026) 24(1):2. 10.5334/jcr.265

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 39.

  DhawaleNGandhiDShanmugamAReddyAKubisH-PDrillerM, et al. Data from: Sleep consistency is a low-cost reliable indicator of nocturnal glycemic control. medRxiv [Preprint]. (2026). 10.64898/2026.03.04.26347496

  • CrossRef
  • Google Scholar
  • View reference in article

• 40.

  Lloyd-JonesDMAllenNBAndersonCAMBlackTBrewerLCForakerRE, et al. Life’s essential 8: updating and enhancing the American Heart Association’s construct of cardiovascular health: a presidential advisory from the American Heart Association. Circulation. (2022) 146:e18–43. 10.1161/CIR.0000000000001078

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 41.

  Ultrahuman. Data from: Terms of use (2024b). Available online at: https://www.ultrahuman.com/us/termsAndCondition/ (Accessed March 20, 2026).

  • View reference in article

• 42.

  Ultrahuman. Data from: Privacy policy (2024a). Available online at: https://www.ultrahuman.com/us/privacyPolicy/ (Accessed March 20, 2026).

  • View reference in article

• 43.

  McKinneyW. Data structures for statistical computing in Python. In: Proceedings of the 9th Python in Science Conference. (2010). p. 56–61. 10.25080/Majora-92bf1922-00a

  • CrossRef
  • Google Scholar
  • View reference in article

• 44.

  VirtanenPGommersROliphantTEHaberlandMReddyTCournapeauD, et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat Methods. (2020) 17:261–72. 10.1038/s41592-019-0686-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 45.

  SeaboldSPerktoldJ. statsmodels: econometric and statistical modeling with Python. In: Proceedings of the 9th Python in Science Conference. (2010). p. 92–6. 10.25080/Majora-92bf1922-011

  • CrossRef
  • Google Scholar
  • View reference in article

• 46.

  PedregosaFVaroquauxGGramfortAMichelVThirionBGriselO, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. (2011) 12:2825–30.

  • Google Scholar
  • View reference in article