Wearable Health Data: Which Metrics Are Measurements and Which Are Estimates?

Wearables directly capture signals such as motion and reflected light; software turns them into the health metrics shown on screen.

Resting heart rate and ordinary step trends are often the closest to measurements, although fit, movement, activity and placement can affect accuracy.

Sleep stages and calories burned are model-dependent estimates. Sleep trackers do not directly measure brain-defined stages, and no consumer brand is consistently accurate across energy-expenditure conditions.

Interpret wrist-derived HRV under consistent conditions, and treat oxygen saturation cautiously when readings are low, symptoms are present or signal quality is poor.

Accuracy must be judged for each metric, model, population and use condition. Strong performance for one output does not validate every feature on the device.

A wearable directly measures relatively few things. Its sensors record signals such as wrist movement and changes in reflected light. Software then converts those signals into familiar outputs: heart rate, steps, sleep stages, calories burned, heart-rate variability and oxygen saturation. Some outputs remain fairly close to the underlying signal; others depend heavily on assumptions and proprietary algorithms.

That distinction changes how wearable data should be used. Heart rate during quiet rest may be a reasonable measurement for many users, while a nightly “deep sleep” total is an inference about physiology the device did not directly observe. Accuracy also has to be assessed metric by metric. Good heart-rate performance does not validate the same device’s calorie or sleep-stage estimates.

A practical measurement hierarchy

Wearable output	What the sensor detects	What the device reports	Best interpretation
Heart rate	Optical pulse waveform	Estimated beats per minute	Often close to a measurement at rest; more condition-dependent during movement
Steps	Acceleration and movement patterns	Classified step events	A generally useful count that varies with gait, activity and placement
Sleep stages	Movement and peripheral cardiovascular signals	Inferred wake, light, deep and REM periods	An algorithmic classification, not direct observation of brain-defined sleep stages
Calories burned	Movement, profile data and sometimes heart rate	Estimated energy expenditure	A model-dependent approximation
HRV	Pulse intervals or cardiac electrical intervals	Variation between detected beats	Depends on sensor type, artifact handling, metric and recording conditions
Oxygen saturation	Absorption of different wavelengths of light	Estimated SpO2 percentage	An optical estimate affected by physiology, fit and signal quality

Heart rate is measured indirectly—and conditions matter

Most wrist wearables use photoplethysmography, or PPG. The sensor shines light into tissue and detects cyclic changes associated with blood volume. An algorithm identifies pulse peaks and converts their timing into beats per minute. That makes the result an optical estimate of pulse rate rather than a direct recording of the heart’s electrical activity.

Heart-rate accuracy is often strongest during sleep or quiet rest and weaker when wrist movement competes with the pulse signal. A systematic review and meta-analysis of wrist PPG devices found average differences near zero during sleep, rest and treadmill activity, but average underestimation of 4.55 beats per minute during cycling and 7.26 during resistance training. Those averages do not describe every individual reading: a small group-level bias can coexist with substantial person-level errors.

Fit, pressure, wear location, ambient light, temperature, circulation and optical skin properties can all affect the waveform. Evidence concerning skin tone and wrist heart-rate accuracy is mixed, so neither universal reassurance nor a universal error correction is justified.

Steps are classifications, not literal detections of feet

An accelerometer records movement. The device’s algorithm decides which movement patterns count as steps. Step counts are among the better-performing consumer metrics under typical walking conditions, but performance changes with walking speed, gait, wear location, activity and population. Arm movements can resemble steps, while slow or altered walking can be missed.

For personal tracking, consistent placement may matter more than agreement with another brand. A sudden change can reflect behavior, but it can also follow a device switch, software update or change from wrist to pocket. Comparing trends is most defensible when the measurement setup remains stable.

Sleep duration is easier than sleep staging

Clinical sleep staging uses polysomnography, which combines measurements such as brain electrical activity, eye movement and muscle activity. A wrist tracker does not collect that full physiological record. It infers sleep and stages from peripheral signals such as motion and pulse patterns.

Consumer trackers generally distinguish sleep from wake better than they identify individual sleep stages. In one simultaneous polysomnography study, three commercial devices had sleep/wake sensitivity of at least 95%, while sleep-stage sensitivity ranged from 50% to 86%. Devices also differed in which stages they overestimated or underestimated.

Across studies, trackers often overestimate total sleep time and sleep efficiency while underestimating wake after sleep onset. The broader systematic review evidence comparing wearables with polysomnography is too variable to treat consumer sleep stages as stand-alone diagnostic measurements. A nightly stage chart can support trend tracking, but a precise “deep sleep” number should not be interpreted like a direct laboratory measurement.

Calories burned are modeled estimates

Energy-expenditure estimates commonly combine movement, user profile information and sometimes heart rate. The algorithm must translate those inputs into metabolic energy use, and that relationship varies by activity and individual. Systematic-review evidence has found substantial activity-specific error, with no consumer brand consistently accurate across energy-expenditure conditions.

The research reference also depends on the question. Indirect calorimetry can assess expenditure in controlled activities, while doubly labeled water is used to estimate total expenditure under free-living conditions. The appropriate conclusion is not that wearable calorie totals are meaningless, but that they are approximate model outputs rather than calories directly counted by a sensor.

HRV depends on what recorded the intervals

Heart-rate variability describes variation in the intervals between heartbeats. An electrocardiogram records cardiac electrical timing. A wrist PPG sensor detects pulse-wave timing, so its corresponding output is more precisely called pulse-rate variability unless it has been validated against ECG-derived HRV for the particular metric and conditions.

Agreement can be reasonable for selected measures during quiet rest, but motion, missed or extra beat detections and artifact correction can substantially alter results. Different HRV statistics are not interchangeable either. For personal trends, use the same device, metric, posture and time of day where possible; measurement conditions and device algorithms can materially affect HRV.

Oxygen saturation is an optical estimate, not a blood-gas measurement

Pulse oximetry estimates arterial oxygen saturation by comparing the absorption of different wavelengths of light. It does not directly sample arterial blood. The FDA’s pulse-oximeter guidance identifies circulation, skin pigmentation, skin thickness, skin temperature, tobacco use and nail polish among factors that can affect readings. A 2024 systematic review found inaccuracies across pigmentation groups and pooled overestimation, alongside substantial variation among devices and study methods.

A wearable SpO2 display should therefore not be presumed equivalent to a clinical assessment. The FDA advises against relying only on a pulse-oximeter number when concerned about health. Symptoms, low readings or an unexpected change warrant appropriate clinical assessment rather than repeated attempts to obtain a reassuring wearable value.

Consumer wearables and clinical instruments answer different questions

Consumer devices are commonly designed for convenient, repeated tracking. Clinical instruments are selected and validated for defined medical purposes, populations and operating conditions. Regulatory status also applies feature by feature: a device can include a low-risk wellness function alongside a separately regulated medical function. A wellness label does not establish clinical validation, while a medically authorized feature does not automatically validate every other output from the same device.

Even a reference instrument is specific to the variable being assessed. ECG is relevant to cardiac electrical timing, polysomnography to sleep staging, arterial co-oximetry to oxygen-saturation validation, and calorimetry or doubly labeled water to energy expenditure. Validation means comparing the complete device output with an appropriate reference under defined conditions; it is distinct from calibrating a sensor against known values.

How to judge an accuracy claim

A useful validation study should identify the exact model and software version, wear location, population, activity or clinical setting, simultaneous reference method, failed readings and artifact handling. It should report more than average bias. Person-level error and limits of agreement help show how far an individual reading may depart from the reference even when the group average looks excellent.

The most defensible approach is to match confidence to the metric. Treat resting heart rate and ordinary step trends as measurements with known conditions and error. Treat sleep stages and calories as model-dependent estimates. Interpret wrist-derived HRV only within a consistent protocol, and treat oxygen saturation cautiously when readings are low, symptoms are present or signal conditions are poor. A wearable can be valuable without every number being a clinical-grade measurement; the key is knowing which question each number can actually answer.