The Signal Is Not Yet the Biomarker

A wearable can record movement, heart rate, sleep, temperature, or an electrocardiogram while someone goes about daily life. The primary subject here is not the watch, ring, or app itself, but the chain that turns those recordings into a claim about health. The question is what a reader should be able to judge: whether a wearable-derived signal is merely available, a useful research measure, a screening aid, or evidence strong enough to support a clinical decision.

That distinction matters because the same stream can be described in several true but unequal ways. A sensor may detect light reflected from skin. A device algorithm may turn that signal into an estimate of heart rate. A researcher may average the estimates, derive a pattern of change, and find that the pattern is associated with a disease. None of those steps, by itself, means that the device has diagnosed the disease or that changing the pattern will improve an outcome.

A wearable records a measurement, not a health conclusion

It helps to separate four things that are often compressed into the phrase “digital biomarker.” First is the signal: a stream of values produced by a sensor under particular conditions. Second is the feature: a summary or transformation of that stream, such as sleep duration, heart-rate variability, a temperature amplitude, or a pattern across several nights. Third is the health relationship: evidence that the feature tracks, predicts, or is associated with a defined state or outcome. Fourth is the use: the decision someone proposes to make from it.

These layers are connected, but they are not interchangeable. A feature can be statistically associated with an outcome without being the cause of that outcome. A classifier can distinguish two groups in a retrospective dataset without telling a clinician what to do with a new person. A measure can be technically reliable while capturing a construct that is too indirect for diagnosis. The practical meaning of a digital biomarker therefore depends on the intended use and on the evidence supporting each link in the chain.

Consider sleep. “The person slept six hours” sounds like a direct observation, but a consumer wearable first decides when sleep began and ended, identifies a main sleep period, and may infer stages from movement and heart rate. A study can then calculate an average, a measure of irregularity, or the percentage of nights that fall within a time window. Each version asks a different question. A single night might describe an immediate state; irregularity across months might describe a behavioral pattern; a relationship with later disease is a further claim still.

The same separation applies to diabetes detection. A model may recognize a combination of heart-rate variability, sleep, and temperature features that occurs more often in people who report diabetes. That can show that the streams contain nonrandom information. It does not mean that the wearable has measured blood glucose, distinguished every form of diabetes, or proven that an alert would improve treatment. The word “biomarker” should clarify the intended biological relationship, not hide the steps between the sensor and the decision.

flowchart LR A["Raw sensor stream"] --> B["Device-derived metric"] B --> C["Feature or pattern"] C --> D["Health-related interpretation"] D --> E["Evidence in defined people and time"] E --> F["Quality, integrity, privacy"] F --> G["Research or monitoring use"] F --> H["Screening aid"] H -. "needs additional clinical validation" .-> I["Diagnosis or treatment decision"]

The supplied studies illustrate why this separation is useful. In a large All of Us resource, Fitbit data were linked with electronic health records, surveys, physical measurements, and genomics. That linkage makes it possible to ask whether a change in activity or sleep occurs around a clinical event. In a study of nearly 6,800 people, long-term Fitbit-derived sleep patterns were associated with later diagnoses across cardiometabolic and psychiatric conditions. In a separate diabetes study, multiple nights of sleep, heart-rate, heart-rate-variability, and distal-temperature features helped classify windows of data from people who self-reported diabetes.

Those findings are important, but their verbs matter. The sleep study reports associations. The diabetes study reports the feasibility of detecting physiology consistent with diabetes. Neither result means that a wearable can replace a diagnostic test, and neither shows that acting on the wearable pattern improves health outcomes. The difference is not a technical footnote. It is the difference between learning that a signal carries information and proving that it can safely guide care.

Scale makes signals useful—and harder to trust

Wearables change the kinds of evidence researchers can collect. A clinic measurement may capture one short visit; a consumer device can record months or years in ordinary settings. The All of Us wearables resource contains data from more than 59,000 participants, with tens of millions of step and sleep observations. Nearly half of the participants with Fitbit data also contributed other health information. This scale supports questions about seasonality, individual change, and what happens to activity after events such as a lower-limb fracture.

The fracture example shows why context matters. In a group of participants with a lower-limb fracture, daily steps fell sharply after the event and gradually approached the pre-injury level over the following months, without fully returning by the end of the observation described. The pattern is not a diagnosis of fracture; the EHR supplies the event, and the wearable supplies a continuous trajectory around it. Together they can reveal recovery dynamics that a single clinic visit would miss. They still cannot tell us, without further design, which intervention caused a particular person’s recovery or whether the trajectory applies to every injury.

But scale adds a second question: whose lives and whose bodies produced the data? The resource combines a bring-your-own-device group with a program that distributed Fitbit devices at no cost. The distribution effort broadened participation, yet the combined dataset remains a convenience sample and is not a miniature version of the United States. The two groups also differ in race and ethnicity, income, education, healthcare access, age, and disability status. A large dataset can make an estimate more precise while leaving it systematically unrepresentative.

The measurement itself also has a history. Researchers using All of Us received Fitbit summary metrics and not raw accelerometry. Fitbit sleep stages are produced by a proprietary algorithm using heart rate and movement. A cited validation study found close agreement with polysomnography for some sleep measures but underestimated REM sleep. The resource paper reports that reliability varies by measure, device, and population; it notes concerns about heart-rate measurement for darker skin tones and step estimation for people with irregular gait. These observations do not show that every wearable metric is unusable. They show that “sleep,” “heart rate,” and “steps” are not device-independent facts.

A signal can also be shaped by whether it is missing. People may remove a device because they feel unwell, lose interest, change devices, or respond differently after receiving a tracker. Those choices can correlate with health and behavior. The absence of a reading is therefore not always random blank space. It may be part of the phenomenon being studied. A study that filters missing nights can improve consistency for its stated question, but the filter can also change which people and periods remain visible.

This is why the central question is not simply whether a wearable collects enough data. It is whether the data preserve a defined construct, in a defined population, for a defined period, with enough context to interpret the result. A device can be useful for discovering patterns while still being unsuitable for a national estimate, a cross-brand comparison, or an individual diagnosis.

The first answer is a chain, not a score

A practical answer to the title question begins with a modest standard. A wearable signal becomes a candidate digital biomarker when its intended health-related meaning is explicit and the feature carries information about that meaning under a stated measurement process. It becomes a more credible biomarker for a particular use only as the rest of the chain is supported: the measurement has known limits, the feature is computed from sufficiently complete data, the relationship holds in an appropriate design, the result travels to relevant people and devices, and the surrounding system preserves the record and makes it usable.

That last condition is easy to overlook. A technically interesting feature may be corrupted after collection, mixed with poor-quality readings, trapped in an incompatible format, or delivered too late to matter. The platform evidence supplied here treats integrity, heterogeneous data, measurement reliability, and scalability as separate requirements. They protect the path from a patient’s device to a stored and reviewable record. They do not, on their own, prove that the signal represents disease or improves decisions.

Nor does a high predictive score finish the job. In the diabetes study, a 21-night model achieved an AUROC of 0.88 in the balanced held-out evaluation and retained useful separation under a more realistic imbalance between people with and without diabetes. That is meaningful evidence for detection feasibility. Yet the labels were self-reported, the cohort was predominantly white, and some windows from people with diabetes looked physiologically similar to negative windows. A screening system would need further testing across populations, devices, settings, thresholds, and clinical workflows before its alerts could be trusted.

The same discipline applies to the sleep findings. Longitudinal associations can reveal that sleep duration, regularity, and stages are related to later disease, and they may inform risk stratification or counseling. They cannot by themselves show that changing one of those patterns prevents disease. The claim must stay matched to the design that produced it. A well-controlled observational study can narrow uncertainty while still leaving causality and benefit unanswered.

For any wearable health claim, the first questions are therefore concrete: What did the sensor measure? What did the device algorithm infer? How much time and how much missing data went into the feature? Who was represented, and who was not? Was the result associated with an outcome, used to classify a label, or tested as part of a decision that improved care? These questions do not reject digital biomarkers. They make the proposed use answerable.

This series follows that chain from the device upward. The next question is the most basic one and often the most neglected: when an app displays a number, what exactly did the wearable measure, what did its algorithm infer, and what did the researcher add before calling the result a health signal?

References