Measuring 'Biological Age' with Wearables — Guillermo Sapiro's CHIL 2025 Keynote

Read this article on Note (Japanese)

In June 2025, at the Conference on Health, Inference and Learning (CHIL) held at UC Berkeley, Guillermo Sapiro, a Princeton University professor and Apple Distinguished Engineer, delivered a noteworthy keynote titled "A Wearable-Based Aging Clock Associates with Disease and Behavior."

The presentation introduced large-scale research that exemplifies modern health technology: estimating "biological age" from PPG (photoplethysmography) signals routinely captured by smartwatches, and using the difference from chronological age to identify connections with disease risk and lifestyle habits.

About Guillermo Sapiro

Guillermo Sapiro is a world-renowned researcher in image processing, computer vision, and machine learning. As of January 2025, he joined Princeton University as a professor in the School of Engineering (Electrical and Computer Engineering, Augustine Family Professor). After 13 years at Duke University, he currently leads medical AI tools at Apple as an engineer.

During his time at HP, he developed image compression algorithms for Mars rovers, with applications extending to Adobe and FDA-approved neurosurgical technologies. Recently, he has also developed apps enabling early autism diagnosis, creating support technologies capable of diagnosing children as young as 18 months.

Why Measure Aging?

Aging is directly linked to numerous disease risks. However, "age = calendar years" fails to explain individual differences. This is where "biological age" comes into focus—quantifying the actual degree of aging within the body to inform health status and preventive medicine.

Traditionally, invasive and costly methods such as DNA methylation and blood tests have been mainstream. However, Sapiro's team recognized significant potential in PPG waveforms automatically captured by smartwatches worn on the wrist during daily life.

Photoplethysmography (PPG) is a device that obtains pulse wave information by measuring changes in blood volume in arteries and capillaries corresponding to heart rate changes. PPG sensors detect heart rate changes by emitting infrared light (typically using light-emitting diodes).

Model Development and Accuracy

This research constructed a model using PPG data from approximately 180,000 participants in the Apple Heart & Movement Study (AHMS), a large-scale study led by Apple. Over 20 million 20-second waveform segments were collected, and self-supervised machine learning was used to learn vector representations capturing "vascular individuality." Subsequently, associations with specific diseases were investigated.

Using 30 days of data from approximately 9,600 healthy individuals, a linear model for age estimation was constructed. The results showed that age could be predicted with an error (MAE) of ±4 years.

What is Wearable Age Gap?

A new metric called "Wearable Age Gap (WAG)"—the difference between estimated age and chronological age—was used to identify high-risk populations. When hospital examinations were encouraged for groups with high WAG, clear correlations with health status and lifestyle habits were demonstrated.

For example:

• Men with high WAG had 1.5-2.5 times higher cardiovascular disease prevalence
• Low WAG groups showed protective effects (0.4-0.6 times)
• Smokers showed +4-6 years, physical inactivity +3-4 years, and sleep deprivation +1 year WAG increase
• Late pregnancy showed a temporary increase of +3.5 years, which recovered postpartum

These results demonstrate that WAG functions not merely as a theoretical proposal but as an effective biomarker linked to actual behaviors and diseases.

What is Next?

Sapiro emphasized the fact that "PPG-derived age metrics can objectively identify high-risk populations" and the prospect that "collaboration with intervention studies is a future challenge." The main focus is on complementing traditional examinations and enhancing the accuracy of early stratification and intervention evaluation.

Key Points

• "Large, daily living digital health" → Massive PPG data collected during daily life serves as the analysis target → Rather than hospital examinations, continuous vital information from daily living joins medical practice

• WAG associations with disease/behavior → Higher WAG correlates with 1.5-2.5 times higher cardiovascular and metabolic disease risk → Enables early stratification of high-risk groups → Can be used to prioritize screening and interventions

• Transient WAG increase and recovery with pregnancy → WAG temporarily increases in late pregnancy and recovers postpartum → Captures short-term physiological fluctuations → Effective for real-time monitoring

• "What is next? – Biological age × interventional study?" → Next step is combination with intervention trials → Quantifying WAG changes through lifestyle improvements or drug administration → Enables quantitative evaluation of intervention effectiveness

In Closing

Guillermo Sapiro, originally from Uruguay, has conducted research at the forefront of NASA, Adobe, and Apple for nearly 40 years. After this presentation, during a brief conversation, he mentioned "a friend's case where discovery was delayed despite obvious signs," and matter-of-factly stated, "The data appears there as numbers."

As a researcher, he maintains his stance of strictly accumulating scientific evidence, never speaking beyond about "how medicine should be." That sense of distance seemed to epitomize his research style.

---

Bellth Inc. supports corporate productivity improvement and innovative business creation through standardization of AI-driven business operations.