From Association to Detection: Two Ways Evidence Grows

Wearable studies can produce impressive results while answering very different health questions. This chapter compares a long-term sleep study that linked Fitbit-derived patterns to later electronic health record diagnoses with a diabetes study that classified multi-night Oura data from people who self-reported diabetes. The goal is not to choose one method as superior. It is to learn how study design determines whether a result supports association, risk stratification, retrospective detection, screening feasibility, or clinical diagnosis.

What the sleep study can—and cannot—say

The sleep study followed 6,785 adults who had linked electronic health record data, at least six months of Fitbit monitoring, and a total of nearly 6.5 million person-nights. It examined several dimensions of sleep separately: average duration, restless sleep, the proportions of estimated REM, light, and deep sleep, night-to-night irregularity, and the proportion of sleep onsets within a defined traditional window. This matters because “sleep” is not one exposure. Duration and regularity can carry different relationships to health, and an algorithmically estimated stage is not the same construct as time asleep.

The researchers used two related designs. A discovery-focused phenome-wide association study, or PheWAS, tested sleep measures against many coded health phenotypes. A PheWAS is useful for looking broadly for patterns, but testing many outcomes creates many opportunities for chance findings. The study used a strict Bonferroni correction across 1,636 phecodes. It then used Cox proportional-hazards models for selected chronic diseases, treating sleep and activity as time-varying measures where possible. In plain language, the second analysis asked whether changes in the observed sleep pattern over time were related to the timing of a later recorded diagnosis.

The results were clinically interesting. Greater sleep irregularity was associated with incident obesity, hyperlipidemia, hypertension, major depressive disorder, and generalized anxiety disorder. REM and deep sleep percentages were inversely associated with some outcomes, including atrial fibrillation, while average daily sleep duration showed J-shaped relationships with hypertension, major depressive disorder, and generalized anxiety disorder. A J-shaped relationship means that both shorter and longer durations were associated with higher risk relative to the middle of the observed range; it does not mean that every individual should be assigned a universal ideal from this study alone.

Several design choices make these findings more informative than a one-time survey. The analysis used device-derived observations over years rather than a single self-reported estimate. It adjusted for factors such as age, sex, body mass index, blood pressure, smoking, alcohol use, education, activity, and selected previous conditions in the time-varying models. It excluded diagnoses during the first 180 days of monitoring to reduce the chance that a disease already present but not yet recorded would be mistaken for a new event. Multiple-comparison control reduced the risk of presenting random associations as discoveries.

Those safeguards narrow uncertainty; they do not turn the study into an experiment. People were not randomly assigned to sleep different amounts or levels of regularity. A sleep disturbance may be an early sign of a disease rather than a cause of it. The electronic health record outcome may be misclassified because a billing code is not a perfect measure of disease onset or severity. Other influences, including occupation and patterns of device use, were not fully available in the dataset.