What a Wearable Actually Measures

A wearable-derived health claim begins before any statistical model sees the data. The specific problem here is that an app presents a finished number while hiding the sensor, device algorithm, and feature choices that produced it. Understanding those hidden steps helps readers ask the right validation question: is a reported value a direct measurement, an estimate, a summary of an estimate, or a pattern whose health meaning has been tested?

From physical signal to device-generated estimate

A wearable does not observe “sleep quality” or “cardiovascular health” as a single physical object. It records signals that are related to those concepts, then applies rules or models to turn the signals into values a person can read. A time series from an ECG sensor is not the same kind of object as a heart-rate number. A dietary photograph is not the same kind of object as a temperature trace. Even when two devices claim to measure the same construct, they may sample, filter, summarize, or classify it differently.

The difference becomes clear in the sleep data used by the All of Us studies. Fitbit supplied daily summary measures, including sleep duration, restless sleep, and sleep stages. The stages were not recorded by watching the brain directly. Fitbit estimated them using heart rate and movement, and it reported them only for sleep periods meeting its own conditions. The researchers then used those device-generated values as inputs to epidemiologic analyses. The resulting number is useful only with its production history attached: it is a Fitbit-derived estimate of a sleep construct, not an observation independent of Fitbit’s algorithm.

The diabetes study makes the same point with a different device. Oura generated nightly measures of heart rate, heart-rate variability, sleep, and sleep stages, while the researchers also derived features from minute-level distal body temperature. Some inputs came through proprietary algorithms; others were calculated from time series using methods described in earlier work. The study’s final feature table therefore combined device-produced summaries with researcher-produced transformations. Both can be informative, but they do not carry the same measurement assumptions.

An ECG illustrates a further distinction. A wearable may record a short electrical time series, but a clinician or model does not interpret every sample independently. The signal may contain motion artifacts, missing segments, or changes caused by poor contact. A platform that receives the record must decide whether it is sufficiently intact to store, display, or analyze. The Xalute platform treats ECG time series and dietary images as different data types, sends them through a common interface, and applies a measurement fault detector before the data are forwarded. A common transport format makes them easier to move through a system; it does not make their measurement processes identical.

The same principle applies to what the device does not reveal. Researchers using the All of Us resource received Fitbit summaries rather than raw accelerometry. That choice can make analyses more consistent within the resource because everyone works with a common product, but it also prevents a researcher from rebuilding the underlying signal or checking whether two device models behaved differently at the raw-sensor level. Standardization and transparency can pull in opposite directions: a fixed vendor metric is convenient to use, while an inspectable signal makes it easier to test alternative definitions and error corrections.

A device-generated estimate can also be fit for one purpose and weak for another. If the goal is to follow a person’s broad change in sleep over time, a stable within-person estimate may be useful even if it is not identical to a laboratory measure on every night. If the goal is to compare small differences in REM sleep between groups, a systematic bias in stage estimation could change the result. The question is not whether an algorithm has an error-free output; it is whether its error is acceptable for the decision the study wants to make and whether that error behaves similarly across the people being compared.

The whole path can be represented as a sequence rather than a single act of “measurement”:

sequenceDiagram participant B as Body or behavior participant S as Wearable sensor participant D as Device algorithm participant F as Study feature participant Q as Reference or quality check participant H as Health claim B->>S: Produce physiological or behavioral signal S->>D: Capture and preprocess signal D->>F: Report metric or estimated construct F->>F: Summarize or derive pattern across observations F->>Q: Compare, filter, or assess quality Q-->>H: Support a bounded association or classifier claim

Each arrow creates a possible failure or interpretation boundary. The sensor may be affected by contact, motion, skin characteristics, or the environment. The device algorithm may be proprietary, so outside researchers cannot inspect every transformation. The study feature may discard timing or variability that matters biologically. A quality check may remove difficult observations and thereby change the population represented in the analysis. A health claim may overreach if it treats an estimate as a direct reference measurement.

This is why validation against a reference device is necessary but limited. Polysomnography, a laboratory sleep assessment, provides a comparison point for consumer sleep estimates. The sleep study reports that cited Fitbit validation work found no significant difference from polysomnography for total sleep time or deep sleep, while the largest cited study underestimated REM sleep by 11.4 minutes. Those results tell us about agreement for particular sleep measures under particular validation studies. They do not show that Fitbit-derived sleep stages predict a disease, that the same agreement holds for another device, or that an association found with the estimate is causal.

The right question is therefore not whether the wearable is “accurate” in the abstract. Accuracy must be attached to a construct, a reference, a population, and a use. A device can be adequate for tracking change within one person while not being interchangeable with another brand in a population study. It can be useful for detecting a broad pattern while being too uncertain for a fine clinical distinction. Measurement validity is a property of the complete measurement arrangement, not a permanent label placed on the hardware.

Feature engineering turns a stream into a biological hypothesis

Once a device has produced a metric, researchers still have to decide what representation will carry the question. The simplest choice is often a primary summary: an average heart rate, total sleep time, or a temperature amplitude. Such summaries reduce a large stream to a manageable value, but reduction can also remove the timing, regularity, or shape of change that makes the signal informative.

The diabetes study explicitly separates primary summaries from secondary features. It analyzed 36 features covering sleep, circadian disruption, heart rate, heart-rate variability, and distal body temperature. The common measures included average resting heart rate, average resting heart-rate variability, total sleep time, and diurnal distal body-temperature amplitude. These are reasonable starting points because they describe recognizable aspects of physiology. They are not neutral choices: each assumes that the average level or amplitude contains information relevant to diabetes-related changes.

Secondary features ask a more specific question about how a signal behaves. The researchers derived measures of temperature complexity and relationships between day and night, rather than looking only at its average. They found that secondary temperature features contributed more to the classifier’s performance than the corresponding primary temperature summaries, and the paper reports a substantial gain in AUROC when distal-temperature features were added. The interpretation is not that complexity itself is diabetes. It is that a representation designed to retain physiological dynamics helped separate the study’s labeled groups under the tested conditions.

Sleep features show the same principle. Sleep duration, restless sleep, light sleep, deep sleep, REM sleep, sleep irregularity, and the timing of sleep onset describe different dimensions. Two people can have the same average duration but very different night-to-night variability. A person can have a similar total time asleep but a different distribution of estimated stages. Calling all of these values “sleep quality” would erase the hypotheses built into the feature definitions.

Feature engineering is therefore a form of biological reasoning. If the hypothesis concerns chronic disruption, a long-run measure of irregularity may be more relevant than one night’s duration. If it concerns circadian regulation, the relation between day and night may matter more than a daily mean. If it concerns an acute rhythm disturbance, a long high-resolution trace may be necessary. The feature should preserve the aspect of physiology that the proposed health relationship needs; otherwise a sophisticated model may be making a decision from a convenient but poorly matched summary.

Model results do not remove this responsibility. In the diabetes analysis, heart-rate and heart-rate-variability features were among the most important individual features, while secondary distal-temperature features improved separation and helped distinguish diabetes from other chronic conditions in the reported tests. Feature importance tells us which inputs contributed to that model’s predictions. It does not prove that one feature is a cause, a specific diagnostic marker, or a complete representation of the disease. Correlated features can share information, and a model can exploit differences between groups that researchers did not intend to measure.

The choice of representation also changes the kinds of error a study can see. A daily average can smooth away a short but meaningful disturbance. A highly detailed feature can preserve fluctuations that are mostly noise or artifacts. A derived measure of day-night amplitude may reflect physiology, but it can also respond to the environment in which a person lives. The diabetes researchers therefore compared feature sets rather than assuming that adding every available stream was automatically better. Their result supports a narrower lesson: features should be added because they represent a plausible and testable aspect of the target physiology, then evaluated for whether they improve separation without creating new false alarms.

This is particularly important when an analysis combines values made by different parties. The device manufacturer may define sleep onset, the researcher may compute a variability measure from those nightly values, and the classifier may combine that result with temperature complexity. If the final prediction is wrong, the error could come from the sensor, the device’s inference, the researcher’s transformation, the labels, or the model. Keeping the layers visible does not make the system simple, but it makes the next validation question identifiable.

The distinction matters especially when a feature is treated as a candidate biomarker. A feature may be useful because it combines several weak signals, because it captures a person’s baseline, or because it reflects a confounder associated with both device use and health. It may perform well in the original cohort and fail when the device, environment, or population changes. The feature’s meaning is not settled by its name or by its position in a ranked importance plot.

A reference comparison and a model evaluation answer different questions. A reference comparison asks whether the device-derived estimate resembles another method for measuring the same construct. A model evaluation asks whether a chosen representation helps distinguish labels or outcomes in a specified dataset. Neither alone establishes that the feature improves patient decisions. That requires a further link between the measured construct, the clinical question, and the consequences of acting on the result.

For the same reason, a researcher should preserve the full provenance of a feature: which device produced the source metric, which algorithm generated it, what observations were excluded, what time window was used, and what transformation was added. Without that context, two studies can appear to measure the same variable while actually using different estimates. The visible label—sleep stage, temperature, heart rate—can conceal the most important part of the evidence.

The next issue follows directly from this measurement chain. A feature is not only defined by how it is calculated; it is also defined by how many observations support it, how they are arranged in time, and what happens when nights or days are missing. Repeated measurements can reveal a pattern that one reading cannot, but they can also introduce new sources of bias and uncertainty.

References