A wrist sensor produced an aging clock in 213,593 people. The sensor has a known bias

Aging clocks — models that estimate a biological age and compare it against the calendar — have until now required something drawn, scanned or measured in a clinic. Blood panels, vital signs, imaging. A study published on 20 October builds one from the optical sensor on the back of a consumer smartwatch [s1].

The scale is the striking part. The model, called PpgAge, was developed using the Apple Heart & Movement Study, with 213,593 participants and more than 149 million participant-days of data [s1].

What the model does

PpgAge uses photoplethysmography at the wrist — the technique that shines light into the skin and measures what comes back, inferring blood volume changes and from those, heart rate and pulse waveform characteristics [s1].

The authors report that this non-invasive, passively collected clock accurately predicts chronological age and captures signs of healthy aging [s1]. The clinically interesting quantity is not the predicted age itself but the gap: the difference between predicted and actual age.

Participants with an elevated PpgAge gap — predicted age greater than chronological age — had significantly higher diagnosis rates of heart disease, heart failure and diabetes [s1]. An elevated gap was also a significant predictor of incident heart disease events and new diagnoses when controlling for relevant risk factors [s1].

The gap tracked behaviour too, associating with smoking, exercise and sleep [s1]. And longitudinally, PpgAge exhibited a sharp increase during pregnancy and concurrent with certain types of cardiac events [s1].

The pregnancy result is worth pausing on. Pregnancy involves large, well-characterised cardiovascular changes — increased cardiac output, expanded plasma volume, reduced peripheral resistance. A model that reads those as accelerated aging is picking up genuine haemodynamic signal, but it is also demonstrating that the "age" label is a convenient framing for a cardiovascular state rather than a measure of biological age in any deeper sense.

The sensor underneath

Everything PpgAge knows, it knows through green-light photoplethysmography. A narrative review published on 7 October assembles what is documented about how well that method works across skin tones [s2].

The physical concern is straightforward: melanin absorbs green light, which is the wavelength most consumer wrist sensors use [s2]. Less light returning to the detector means a weaker signal and a worse estimate.

The review searched PubMed, Google Scholar and American College of Cardiology databases for English-language studies published between May 2017 and May 2025, yielding 50 records of which 23 met inclusion criteria [s2]. Its findings describe significant variability across devices [s2]:

• One validation study of 60 participants reported that a major smartwatch brand showed mean heart rate differences of less than 5 bpm across skin tones (95% CI −3.2 to +4.1) [s2].
• Other brands underestimated heart rate by 10–15 bpm at rest, and by more than 20% during vigorous activity, in darker-skinned users (n = 75, p < 0.01) [s2].
• Devices using alternate operating systems displayed inconsistent calibration, with error rates reaching up to 12% [s2].
• Large-scale studies involving over 400,000 participants reported greater than 95% sensitivity for atrial fibrillation detection — but were not stratified by skin tone [s2].

That last line is the one that connects the two papers. Very large PPG studies exist, they report strong performance, and they have not reported it by skin tone.

Why this matters for an aging clock specifically

A heart rate error is a bounded problem: the number is wrong by some amount, and a user can notice. An aging clock is different. It is a derived, unfamiliar quantity with no external reference the user can check. If the input waveform is systematically degraded for some users, the output gap is systematically biased for them — and there is no way to tell from the number itself.

Whether PpgAge is affected is not established by either paper. The aging clock study does not report performance stratified by skin tone [s1], and the bias review does not examine aging clocks [s2]. The concern is inherited from the sensor modality, not demonstrated in this model.

The review's own recommendations are for standardised validation protocols stratified by Fitzpatrick skin type and activity level, alongside regulatory mandates for inclusive testing [s2]. It notes that skin-tone-related bias in smartwatch PPG poses risks extending beyond fitness tracking to arrhythmia detection, hypertension monitoring and telehealth reliability [s2].

The limits on both sides

The aging clock study is observational [s1]. Associations between an elevated PpgAge gap and heart disease diagnoses do not establish that the gap causes anything, or that reducing it would change risk. The cohort is also self-selected in a specific way: people who own an Apple Watch and enrol in a research study are not a random sample of any population, and they skew toward higher income, better health engagement and, in most such cohorts, lighter skin tones.

The bias review carries a different set of caveats. It is a narrative review, not a meta-analysis, published in Cureus — a journal with a lighter peer-review process than the specialty cardiology or engineering venues where device validation usually appears. Its included studies are small: the two validation studies it reports numbers from had 60 and 75 participants respectively [s2]. Those are underpowered for the question, and the review does not pool them.

Neither paper should be read as settling how PPG performs across skin tones. The honest state of the literature is that a plausible physical mechanism for bias exists, some small studies have found large device-dependent errors, and the very large studies that could resolve it have not reported the relevant stratification [s2].

What to watch

Whether the next generation of large PPG-derived models — aging clocks, arrhythmia detectors, blood pressure estimators — report performance stratified by skin tone as a matter of course. The datasets to do it exist. A study of 213,593 participants [s1] has the statistical power to answer the question that a 60-person validation cannot.

Sources

• [s1] A wearable-based aging clock associates with disease and behavior — Nature Communications, published 20 October 2025. https://doi.org/10.1038/s41467-025-64275-4
• [s2] Photoplethysmography in Diverse Skin Tones: Evaluating Bias in Smartwatch Health Monitoring — Cureus, 7 October 2025. Narrative review. https://doi.org/10.7759/cureus.94074