Digital Biomarkers: How Wearables Are Becoming Medical Devices

In the Apple Heart Study, published in 2019, roughly 420,000 people wore a consumer smartwatch that watched their pulse in the background for signs of an irregular heart rhythm. A small fraction received a notification suggesting possible atrial fibrillation. Among those who followed up with a clinical-grade ECG patch, a substantial share were confirmed to have it, often people with no symptoms and no reason to have sought a cardiac workup otherwise. That's the moment consumer wearables stopped being step counters and became something closer to a screening tool. It's also the moment that made clear how far the category still had to go: most of the flagged notifications did not end up confirmed as AFib, a tradeoff between catching real cases early and generating false alarms that has shaped how these tools are regulated and used ever since.

What a Digital Biomarker Actually Is

• A digital biomarker is a measurable, objective signal collected from a digital device, a wearable, a smartphone sensor, an implant, that correlates with a specific physiological or health outcome, and that correlation has been established through clinical research rather than assumed from the raw sensor data alone.
• The distinction that matters: a sensor producing a number is not automatically a digital biomarker. Heart rate is a digital biomarker because decades of clinical research established what elevated or depressed resting heart rate actually means for health outcomes. A proprietary "readiness score" that blends several signals into one number is not automatically a validated biomarker unless that specific composite score has itself been studied against real outcomes.
• This distinction, measured signal versus validated health-outcome correlation, is the difference between a wellness feature and something a clinician can actually use.

The Progression

The category has expanded in a fairly consistent order, moving from easy-to-measure, loosely predictive signals toward harder-to-measure, more clinically specific ones: steps, then heart rate, then blood oxygen (SpO2) estimation, then atrial fibrillation detection, then sleep-apnea screening features, then continuous glucose monitoring, then early cuffless blood pressure estimation, with stress and cortisol-adjacent signals now emerging as the newest, least-validated frontier. Each step up that list requires more sophisticated sensing and, critically, more sophisticated algorithmic interpretation of noisy real-world data, not just a better sensor.

The FDA Clearance Landscape

This distinction matters more than most marketing copy suggests. A feature can be FDA-cleared as a medical-grade tool, meaning the manufacturer submitted clinical validation data and the FDA reviewed it against a regulatory standard, or it can be sold as general consumer wellness, which carries no such review requirement. Apple Watch's ECG app and irregular rhythm notification feature, and equivalent AFib-detection algorithms from Fitbit and other manufacturers, are FDA-cleared. Blood-oxygen (SpO2) estimation on most consumer wearables is explicitly marketed as a wellness feature, not a medical-grade pulse oximeter reading, a distinction most users never see spelled out on the box. Continuous glucose monitors have historically required a prescription as Class II medical devices; more recently, over-the-counter CGM products marketed for general wellness rather than diabetes management have started blurring that line further, available without a prescription but without the same clinical-management backing as a prescribed system.

The AI Layer

None of this works from the sensor alone. Raw signal from a wrist-based optical sensor is noisy, affected by motion, skin tone, ambient light, and placement. What turns that noise into a usable health signal is the algorithm sitting between the sensor and the number a user sees, increasingly a machine learning model trained on large labeled datasets to distinguish a real irregular heartbeat from a motion artifact, or genuine sleep-stage transitions from restless movement. This is why two wearables with functionally similar sensors can produce meaningfully different accuracy: the hardware is often a smaller differentiator than the model interpreting it. It's the same underlying dynamic covered in our piece on AI's role across radiology, where the sensor or image is only half the system; the trained model doing the interpretation is the other, increasingly more important, half.

Validated vs. Marketed

• Well-validated: resting heart rate trends, step count as an activity proxy, FDA-cleared AFib detection, total sleep duration.
• Improving but imperfect: sleep-stage classification (light, deep, REM), heart rate variability as a recovery proxy, VO2 max estimation from wrist sensors rather than direct gas-exchange testing.
• Early and less validated: cuffless blood pressure estimation, cortisol or stress biomarkers from wearable sensors, proprietary composite "readiness" or "strain" scores that combine multiple signals into a single number without independent peer-reviewed validation of the composite itself.

What's Coming

Three directions are furthest along in active research: continuous glucose monitoring without a needle-inserted sensor, using optical or other non-invasive methods, still an unsolved problem at consumer accuracy levels as of today; gait- and movement-pattern analysis for early Parkinson's disease detection, using smartphone and wearable motion data to catch subtle changes years before clinical diagnosis; and cardiac event prediction that moves from detecting an arrhythmia as it happens toward flagging elevated risk before an event occurs, using longitudinal trend data rather than a single reading. None of these is a consumer product today. All three are active, well-funded research areas with published early results, the kind of trajectory that connects wearable data to the personalized risk models discussed in our companion piece on AI mammography.

What to Actually Do With Wearable Data

• Bring trend data, not single readings, to a physician: a gradually rising resting heart rate over months is more clinically useful than one day's number.
• Know which of your device's features are FDA-cleared versus wellness-only; this is usually disclosed in the app's fine print or the manufacturer's regulatory documentation, and it changes how much weight the reading deserves.
• Treat an AFib or arrhythmia notification as a reason to seek clinical follow-up, not as a diagnosis; consumer detection algorithms are a real screening signal, but confirmation requires clinical-grade evaluation.
• Be more skeptical of proprietary composite scores (readiness, strain, recovery) than of individual, well-established metrics like heart rate or sleep duration, since composite scores are typically not independently validated the way their underlying inputs often are.