A Year on Your Wrist: 11 Million Days of Fitbit Data Map How Activity Shapes Disease Risk

Wrist-worn fitness trackers have quietly become one of the largest sources of continuous health data in human history, yet most studies that use them still throw away nearly everything they capture. A typical analysis compresses a week of minute-by-minute movement into a single average step count and calls it a day. A new study published in Nature Communications argues that this convention is leaving enormous amounts of information on the table. Drawing on roughly 11 million participant-days of Fitbit data from 29,303 volunteers in the United States National Institutes of Health’s All of Us Research Program, a team led by researchers at the NIH’s National Human Genome Research Institute and the University of Oxford systematically compared how different ways of summarizing wearable data change what we can learn about health and disease.

The scale of the dataset is what makes the work possible. All of Us participants consented to share both their electronic health records and data from personal wearable devices, creating an unusual opportunity to link long stretches of objective physical activity monitoring to clinical diagnoses. Rather than choosing one summary metric in advance, the researchers derived four distinct daily activity measures from the raw Fitbit streams: total step count, the peak one-minute cadence (a measure of brief bursts of intense movement), the peak 30-minute cadence (sustained moderate-to-vigorous activity), and heart rate per step, which serves as a proxy for cardiovascular efficiency during walking.

The second axis of the study was time. For each of the four metrics, the team computed summaries over five different measurement windows: a single day, one week, one month, six months, and one full year. This produced twenty wearable-derived phenotypes per participant, each capturing a slightly different aspect of habitual behavior. The authors then ran phenome-wide association analyses, testing each phenotype against more than 700 incident disease outcomes, meaning conditions diagnosed after the measurement period, and more than 1,400 prevalent outcomes, meaning conditions the participant already had. After strict Bonferroni correction for the enormous number of statistical tests, the results revealed a clear and consequential pattern.

Although all twenty phenotypes were highly correlated with one another, the length of the measurement window mattered enormously. One-year step count was associated with 389 prevalent and 61 incident disease outcomes, compared with 271 prevalent and 33 incident associations for one-day step count. In other words, simply averaging activity over a longer period produced stronger, more stable, and more numerous links to disease. The likely explanation is statistical as much as biological: a single day of tracking is a noisy snapshot of behavior that fluctuates with weather, illness, work schedules, and chance, whereas a year-long average captures the stable, habitual level of physical activity that actually matters for chronic disease biology.

The incident-versus-prevalent distinction adds a second layer of insight. Associations with prevalent diseases reflect how existing illness shapes activity, since conditions such as arthritis, heart failure, or chronic obstructive pulmonary disease can reduce mobility long before they are formally diagnosed. Associations with incident diseases, by contrast, suggest that activity measured before diagnosis may act as a genuine risk marker or even a modifiable risk factor. The differences between the two patterns imply that wearable metrics can serve a dual role: as early warning signals that subtle declines in movement precede a clinical diagnosis, and as targets for intervention aimed at preventing disease altogether.

Among the incident findings were novel associations that had not previously emerged from shorter, smaller studies, including links between long-term activity patterns and chronic pain syndrome and with SARS-CoV-2 infection. The chronic pain connection is particularly intriguing from a mechanistic standpoint. Reduced movement could plausibly contribute to the deconditioning and sensitization that characterize chronic pain, or conversely, subclinical pain could suppress activity months or years before patients seek care. The SARS-CoV-2 association may reflect relationships between baseline fitness, immune function, and exposure risk, though the authors are careful to frame these as hypotheses for future work rather than settled causal claims. Because the study is observational, association does not establish causation in either direction.

Technically, the phenome-wide approach resembles the genome-wide association studies that transformed genetics over the past two decades, but applied to behavior instead of DNA. Rather than testing a single hypothesis, the researchers scanned the entire diagnostic landscape of the electronic health record, letting the data reveal where activity signals are strongest. This agnostic strategy is powerful for discovery but demands aggressive multiple-testing correction, which the team applied through Bonferroni thresholds. It also demands careful handling of confounding: people who wear Fitbits for years and share their health records differ systematically from the general population, tending to be younger, wealthier, and more health-conscious. The All of Us cohort’s deliberate diversity mitigates some of these biases, but the authors acknowledge that generalizability remains a consideration.

The heart rate per step metric deserves particular attention because it illustrates the kind of signal that conventional step-count studies discard entirely. By dividing average heart rate by cadence, the researchers obtained a measure of the cardiovascular cost of walking, which varies with fitness, autonomic tone, medication use, and illness. A rising heart rate per step over months could flag developing cardiovascular disease, anemia, thyroid dysfunction, or deconditioning long before symptoms prompt a clinic visit. Similarly, peak cadence metrics capture the intensity dimension of physical activity that public health guidelines emphasize but that daily step totals obscure. Two people can accumulate identical step counts while one does virtually all walking at a leisurely pace and the other includes brisk exercise bouts with very different physiological implications.

For the research community, the practical message is methodological: measurement window length is a design choice with real consequences, and the field’s default of seven-day snapshots systematically underestimates the breadth of activity-disease relationships. The authors suggest that future studies should prefer longer windows where data availability permits, and should report both prevalent and incident analyses separately rather than pooling them, since the two answer different questions. For device makers and app developers, the findings hint at a future in which a year of passive wrist data could generate a personalized, continuously updated map of disease risk, flagging concerning trajectories such as declining cadence or climbing heart rate per step well before traditional screening intervals.

There are, of course, important caveats. Wearable data are not error-free: devices misclassify steps, lose data when batteries die or wrists are bare, and measure only wrist movement, missing activities such as cycling or swimming. Wear time itself is not random, and periods of illness often coincide with periods of missing data, which can bias summaries in subtle ways. The study population, while diverse, is still a volunteer cohort within a single national research program, and the Fitbit devices analyzed represent one manufacturer’s algorithms. Nevertheless, the sheer scale of the analysis, 11 million days spanning five time windows, four metrics, and more than 2,100 disease outcomes, makes a compelling case that the era of treating wearables as week-long step counters is ending. As longitudinal device data accumulate across medicine, the humble fitness tracker is being repositioned as a population-scale epidemiological instrument, one capable of revealing how the quiet rhythms of everyday movement shape the diseases we develop years later.

Subject of Research: Phenome-wide association analysis of longitudinal wearable physical activity data linked to incident and prevalent disease outcomes

Article Title: 11 million days of longitudinal Fitbit data reveal novel future health insights

Article References: 11 million days of longitudinal Fitbit data reveal novel future health insights. (n.d.). https://doi.org/10.1038/s41467-026-78529-2

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78529-2

Keywords: wearables, Fitbit, physical activity, All of Us Research Program, phenome-wide association study, step count, cadence, heart rate, electronic health records, chronic disease, risk factors, Nature Communications