When Does a Wearable Signal Become a Digital Biomarker?
The question is not whether a watch, ring, or phone can collect health-related data. It is whether a particular signal has earned a particular meaning for a particular use. This final synthesis brings together the measurement chain, longitudinal evidence, population limits, transportability, and data infrastructure needed to judge that claim. The result is a conditional standard: a wearable signal becomes a digital biomarker only when its intended meaning and its limits are sufficiently validated for the decision it is meant to support.
A biomarker is a claim with a specified use
The word “biomarker” can sound like a property of a number. In practice, it is a claim about a relationship. The claim might be that a wearable-derived feature reflects a physiological state, tracks change in a person, identifies people who merit follow-up, or helps guide a clinical decision. Those uses demand different evidence. A measure can be valuable for observing a trend without being suitable for diagnosis. A classifier can identify a group in a retrospective study without proving that an alert improves care.
The first requirement is a defined construct. “Sleep quality,” “stress,” or “cardiovascular health” is too broad to validate as a single measurement. A study must say whether it is estimating sleep duration, stage proportions, night-to-night irregularity, heart-rate variability, temperature dynamics, or another defined feature. It must also say how the device produced that estimate. Fitbit-derived sleep stages use heart rate and movement; Oura supplies its own algorithmic summaries; a wearable ECG is a time series that can be disrupted by movement or poor contact. The name of the output does not make it device-independent.
The reference standard matters, but it answers only the question it was designed to answer. A comparison with polysomnography can test how closely a consumer device estimates selected sleep measures under the validation conditions. The cited Fitbit evidence found no significant difference for total sleep time or deep sleep in the largest referenced comparison, but underestimated REM sleep by 11.4 minutes. That supports a qualified statement about measurement agreement. It does not prove that the same estimate predicts disease, that another vendor measures the construct equivalently, or that a resulting association is causal.
The second requirement is a feature that preserves the biology or behavior relevant to the question. A simple average can describe a level while discarding timing and variability. The diabetes study’s secondary temperature features, including measures of complexity and day-night relationships, added information beyond primary summaries in the tested models. The sleep study treated duration, restless sleep, stages, onset timing, and irregularity as distinct features rather than one undifferentiated score. Feature importance can show that an input contributed to a classifier; it cannot by itself show that the input causes disease or is a complete diagnostic marker.
The third requirement is temporal context. A feature computed from one night answers a different question from a feature computed across three weeks or several years. In the diabetes study, AUROC for the XGBClassifier rose from 0.74 with one night to 0.88 with fourteen nights, and fourteen- and twenty-one-night windows performed similarly in the balanced evaluation. The improvement suggests that repeated measurements capture persistent differences that are not visible every night. It does not make twenty-one nights a universal rule. Some windows from people who self-reported diabetes still resembled negative windows, showing that the target physiology was not continuously observable.
A longitudinal health association needs its own temporal logic. The sleep study used baseline periods, monthly time-varying measures, and a 180-day exclusion intended to reduce reverse causation. Its nearly 6.5 million person-nights made it possible to study duration, regularity, stages, and incident diagnoses over time. Yet a later diagnosis can be preceded by an early symptom, and an EHR code may not mark the exact biological onset.