BGPT: Paper Review: A wearable-based aging clock associates with disease and behavior

Fuel Your Discoveries

Quick Explanation Copied

Paper in one breath

The paper introduces PpgAge, a wearable wrist-PPG (Apple Watch) deep-learning aging clock, trained on “healthy” participants and evaluated in a much larger general cohort. It predicts chronological age (MAE ~2.4–2.5 years in the healthy cohort; ~3.1–3.3 years in the general cohort) and its “age gap” (predicted minus chronological age) associates with prevalent diagnoses and incident cardiometabolic events, and correlates with behavior (smoking, exercise, sleep), with observable longitudinal shifts around pregnancy and major cardiac events.

Long Explanation

Paper Review: A wearable-based aging clock associates with disease and behavior

DOI: 10.1038/s41467-025-64275-4

Authors / affiliations / funding: Apple Inc (multiple authors) and Princeton University (one author); study funded by Apple Inc, American Heart Association, Brigham and Women’s Hospital.

1) What the authors built (mechanistic core)

Signal: Apple Watch consumer wrist PPG; 60-s segments, green channel, 4-channel optical configuration; sampled at 64 Hz or 256 Hz.
Representation learning: self-supervised “foundation” encoder trained contrastively on ~19,993,427 segments (172,318 participants) to map each 60-s segment → a 256-d feature vector.
Normative aging model: a ridge-regularized linear model trained on a “healthy” subset using average representation(s) over the first 30 days; outputs PpgAge (predicted chronological age).
“Age gap” statistic: PpgAge − chronological age, then age-dependence is removed by regressing raw PpgAge gap on chronological age using a spline and retaining residuals (adjusted gap).

2) Data scale & evaluation design (what’s strong vs what’s fragile)

Scale: AHMS has 213,593 participants and >149 million participant-days; PpgAge is trained on a healthy subset (n=6,728 initially; train=5,355/test=1,373) and then evaluated both on held-out healthy test participants and a much larger “general” held-out cohort (n=120,235).

Evaluation fragility to check: this is observational and uses survey-reported diagnoses/behaviors; the “healthy” training definition is restrictive and may induce distribution shift in older or nonrepresentative users; PpgAge is trained on “healthy” physiological patterns and may partly proxy cardiometabolic fitness rather than a universal multi-system aging process.

3) Visualizations (derived from reported summary numbers)

Source: MAE values for healthy and general cohorts are reported in the Results section (by sex).

Source: Cox model hazard ratios for +6-year adjusted PpgAge gap are explicitly stated for ASCVD events and hypertension.

Source: Pregnancy longitudinal analysis reports a median ~3.56-year increase (IQR 1.65–5.65) in the ~270 days preceding birth in the analyzed subset (n=165).

4) Evidence quality checklist (skeptical, evidence-weighted)

Known / directly measured vs inferred:

Directly supported: PpgAge predicts chronological age with reported MAEs and CIs in both cohorts.
Directly supported (association): Adjusted PpgAge gap stratifies diagnosis rates for multiple conditions and predicts incident ASCVD/hypertension in Cox models adjusting for specified risk factors.
Behavior links: Age gap correlates with smoking status/frequency, exercise minutes quintiles, and sleep stage metrics (REM latency, total sleep duration, deep sleep duration, sleep efficiency).
Longitudinal physiological sensitivity: Pregnancy and specific adverse cardiac events show detectable increases or changes in participant-level PpgAge time series.

Uncertain / not established by this paper:

Causality: observational cohort + self-report outcomes cannot establish that “age gap” is causal or that behavior changes “drive” biological aging as opposed to reflecting underlying health.
Mechanism / attribution: the paper argues PpgAge uses morphological features beyond HR/HRV and relates to vascular aging, but exact physiological mapping from learned waveform features → mechanisms remains unresolved in the presented text.

5) Potential biases / blind spots (critical but fair)

Selection & representativeness: AHMS participants are Apple Watch users and may not generalize; “healthy” training cohort may be biased toward younger/healthier patterns (and the paper acknowledges distribution shift and reduced “healthy” representation above age 75).
Information bias: diagnosis rates and behaviors rely on self-reported surveys; missing/under-reported diagnoses would tend to weaken observed associations (directional uncertainty), while differential reporting could also bias strata.
Confounding & reverse pathways: PpgAge gap could reflect existing subclinical disease or cardiovascular fitness rather than “aging per se”; the paper controls for several risk factors in incident analyses, but residual confounding and pathway overlap are still possible.
Company / device ecosystem incentives: multiple authors are Apple employees, and AHMS is funded by Apple. This does not invalidate results, but it raises the need for transparent external validation independent of the device ecosystem.

6) What would most strengthen (or falsify) this line of work?

Highest-impact follow-ups:

External replication with independent wearables + data access: demonstrate similar age prediction and disease/incident associations across other sensor models and populations; the paper itself notes data cannot be fully public due to privacy and contractual constraints.
Prospective intervention tests: randomized behavior or clinical control trials assessing whether changes in PpgAge gap track mechanistic improvements (not just associations).
Mechanistic interpretability: connect learned waveform morphology (e.g., notch/diastolic peak changes noted by the paper) to vascular physiology; at minimum, validate whether perturbations known to affect arterial stiffness alter PpgAge morphology signatures.

Simple falsification targets (grounded in their claims):

Across held-out cohorts, PpgAge should not predict chronological age worse than chance; similarly, the PpgAge gap should show diminished disease association when using a null surrogate that preserves chronological-age dependence only. (The paper includes internal HR/HRV ablation showing worse performance with HR/HRV-only models.)

7) Author-review deep dives (bespoke links)

Feedback:

Updated: April 01, 2026