Why Time Turns Readings into Patterns

A wearable reading is a snapshot; a wearable-derived health pattern is a relationship across observations. This chapter focuses on the role of time: how a single night becomes a usable window, how a baseline becomes a trajectory, and how a trajectory can be aligned with a clinical event. The practical judgment is simple but demanding: before accepting a wearable finding, ask how much time supports it, what was missing, and what changed between the first observation and the claimed outcome.

From one night to a pattern

One observation can be useful when the question is about an immediate state. It is a weak foundation for a claim about a persistent condition that varies from day to day. A person’s heart rate, sleep, temperature, and behavior change with illness, activity, stress, environment, and ordinary routines. A feature that is visible on one night may be absent on the next. Repeated measurement can help distinguish a recurring difference from a temporary fluctuation, but it does not make every later average meaningful automatically.

The diabetes study provides a clear example. The researchers compared models using windows of one, three, seven, fourteen, and twenty-one nights. The one-night XGBClassifier had an AUROC of 0.74, while the fourteen- and twenty-one-night models reached 0.88 in the balanced held-out evaluation. The authors interpret the improvement as evidence that longer windows accumulate persistent differences that are not expressed equally every night. The model is not simply seeing more rows; it is being given a better chance to observe a pattern that survives ordinary nightly variation.

That improvement has a limit. The fourteen- and twenty-one-night models performed similarly in the reported balanced evaluation, and the study did not establish that twenty-one nights is the correct window for every disease, device, or person. The authors chose that horizon partly so that enough windows would pass their data-availability rules and partly to reduce complications from longer behavioral cycles. A window is therefore a design choice tied to the biological process, the observation schedule, and the intended use.

The study’s false negatives make the point more strongly than the headline performance. Some windows drawn from people who self-reported diabetes looked more like negative windows than like the model’s detectable positive windows. In other words, the person-level label did not guarantee that the target physiology would be visible in every period. A screening system built on such a signal might need repeated testing rather than treating one negative window as decisive. The time dimension changes not only sensitivity but also the meaning of a result.

There is also a difference between learning about a person and learning about a population. A person’s current value can be compared with their own baseline, while a population model may compare one person with the distribution of other people’s values. Those comparisons answer different questions. A large deviation from an individual’s usual pattern might matter even if the value is ordinary in the population; a population-level difference might be statistically clear while saying little about whether a particular person has changed. The time window must therefore be paired with the comparison frame, not treated as a universal setting.

The diabetes analysis also protects against a common mistake by splitting its machine-learning data by individual rather than treating every night as an independent person. If nights from the same person appeared on both sides of a test, a model might learn that person’s stable signature instead of learning a pattern that transfers to a new individual. Even with an individual-level split, the result remains tied to the study’s device, labels, and sampling rules. Time can improve the signal while also creating opportunities for leakage or overinterpretation if the unit of analysis is unclear.

A similar distinction appears in longitudinal epidemiology. The sleep study established a baseline period, calculated averages, and then entered monthly sleep and activity measures into time-varying models. A baseline describes where a person starts. A rolling average describes how the observed state changes while smoothing day-to-day noise. An event-aligned trajectory asks whether a change in the wearable signal appears before, during, or after a known health event. These are different questions, even when they use the same underlying sleep or activity metric.

The All of Us resource demonstrates event alignment with lower-limb fracture. Among participants with such an event, daily steps dropped sharply relative to the pre-event baseline and gradually approached earlier levels over the following months. The value of the wearable record is not that it independently identifies a fracture; the electronic health record supplies the event. The wearable adds a continuous view of functional change before and after it. This makes recovery visible as a trajectory rather than as a sequence of isolated clinic assessments.

Time can also expose a misleading average. Two people may have the same mean sleep duration, while one sleeps a similar amount each night and the other alternates between very short and long nights. A single average hides the difference; a measure of irregularity preserves it. Conversely, a very detailed time series can preserve fluctuations that are mostly sensor noise or missingness. Choosing the window and summary is therefore part of defining the biomarker, not a neutral step after the “real” measurement has been made.

The changing status of an observation can be described as a lifecycle:

stateDiagram-v2 [*] --> IsolatedReading IsolatedReading --> UsableWindow: enough valid observations IsolatedReading --> MissingOrInvalid: gaps or quality failure MissingOrInvalid --> IsolatedReading: later observation arrives UsableWindow --> IndividualBaseline: repeated measures define reference IndividualBaseline --> LongitudinalPattern: later windows accumulate LongitudinalPattern --> EventLinkedTrajectory: align with clinical event EventLinkedTrajectory --> LongitudinalPattern: follow-up continues

The transitions matter. A missing or invalid period should not silently become a normal value. A baseline can be distorted if it begins after a condition has already changed behavior. An event-linked trajectory can show correlation in time without proving that the event caused every movement in the wearable measure. A time-aware analysis preserves these distinctions; a simple average can erase them.

The sleep study used a 180-day exclusion at the beginning of monitoring to reduce the chance that an already-present disease would be recorded as a new diagnosis after the sleep pattern had changed. That design choice addresses reverse causation, but it cannot eliminate it. Sleep disturbance may be an early sign of a condition, and the available records may not capture the exact onset. More time helps, but it does not automatically reveal which direction the relationship runs.

Scale without representativeness

Longer observation becomes especially valuable when it is paired with a large cohort and other kinds of health information. The All of Us wearables resource combines Fitbit data from more than 59,000 participants, more than 39 million step observations, and more than 31 million sleep observations. Nearly half of the participants with Fitbit data also contributed electronic health records, physical measurements, genomics, or surveys. This makes it possible to study seasonal patterns, individual trajectories, and relationships between wearable measures and clinical events at a scale that short laboratory recordings cannot easily match.

The linkage changes the questions researchers can ask. A sleep or activity pattern can be examined alongside diagnoses, social factors, and other measurements rather than treated as an isolated stream. A researcher can ask whether a drop in steps follows a recorded injury, whether a sleep feature differs across groups, or whether a wearable measure tracks a change that is visible in the medical record. These are stronger questions than “what is the average number in this dataset?” because they use time and context together.

Linkage also creates a new dependency: the clinical record and the wearable stream must be aligned well enough for the timing to mean something. A billing code may appear after a condition began, and a wearable account may contain measurements from before a participant formally enrolled. The sleep study accounted for some of this by defining monitoring and diagnosis windows, but it still acknowledged that EHR codes can be misclassified and that the exact onset of a condition may be uncertain. More data sources create a richer context; they do not guarantee that the context is perfectly synchronized or complete.

Yet the cohort is not a neutral window onto the whole population. The resource combines people who volunteered data from a device they already owned with people invited to receive a Fitbit through the WEAR study. Providing devices broadened participation, but the combined dataset remains a broad convenience sample. It is older, more female, and more highly educated than the general US population. The BYOD and WEAR groups also differ across race and ethnicity, income, education, healthcare access, and disability status. A larger denominator does not cancel those differences.

Device composition adds another layer. The resource includes 41 Fitbit models, and device information was not available for every participant. Different models may use different sensors or algorithms, and some people used multiple devices during their donation period. If one model underestimates a metric or if a device change coincides with a change in behavior, a pattern may reflect the measurement system as well as the person’s physiology. The resource paper notes that there is no consensus method for handling all of this heterogeneity.

Missingness is equally important. In a free-living study, a missing night may mean a dead battery, a removed device, a loss of interest, a change in health, or a deliberate choice not to share. Those possibilities do not carry the same meaning. Filtering participants for sufficient data may make a model easier to estimate, but it can also select people who are healthier, more engaged, or more able to wear the device consistently. The sleep and diabetes studies use different completeness rules because their questions and data structures differ; there is no universal definition of a usable observation.

Measurement can even change the behavior being measured. The All of Us paper notes that participants may temporarily alter their activity or other habits after receiving a tracker. If the first days are unusually active, unusually attentive, or unusually incomplete, including them in a long-run average can shift the apparent baseline. Excluding them might reduce that reactivity but would also remove part of the real-world adoption process. The correct choice depends on whether the study asks about stable physiology, ordinary behavior, or the effect of being monitored.

This is the trade-off at the heart of large wearable datasets. Scale and duration make rare patterns, individual change, and event relationships visible. They also expose the study to more device versions, more missingness, more uncontrolled context, and more opportunities for selection to shape the result. A large cohort strengthens a claim when the added observations address the question and the sampling and measurement processes are understood. It merely makes a biased estimate more precise when the same systematic distortions remain unexamined.

The choice between individual and population context is not a contest with one winner. An individual baseline can make a meaningful change visible even when two people have very different normal values. A population reference can reveal a broad association that an individual record cannot establish. Clinical linkage can show that a trajectory coincides with a recorded event, while repeated self-measurement can show how a person’s pattern persists or fluctuates. The strongest interpretation keeps these frames separate and states which one produced the result.

The time window, baseline, and cohort therefore belong in the claim itself. “Average sleep duration was associated with disease” is incomplete without knowing whose sleep, which device, over what period, with what exclusions, and against which outcome definition. “A multi-night model detected physiology consistent with diabetes” is incomplete without its window length, labels, missingness rules, and test population. These details are not procedural clutter; they define what the evidence means.

Once a wearable pattern has been built from time and context, the next question is what kind of evidence it represents. A longitudinal association between sleep and later disease and a classifier that separates diabetes-labeled windows may both look like successful biomarker results, but they answer different questions and support different conclusions.

References