Activity tracking beyond totals: disease-enriched local patterns split hypertension and diabetes into two phenotypes

The unit of analysis for all-day activity records shifts from the average activity of a whole cluster to local neighborhoods on a similarity network, with disease information entering pattern discovery directly. Earlier work first partitioned people by activity profile using k-means, hierarchical clustering or latent profile analysis, then tested the resulting groups for disease association, so the groups mainly captured the dominant axes of variation. In the UK Biobank hypertension analysis, 1,212 representative seed neighborhoods covered all 80,532 participants, and 67 neighborhoods showed significantly higher prevalence than people outside them before being grouped into clusters of 25 and 42 seeds.

The method first removes activity differences attributable to age, sex and body mass index, then asks at each point on the similarity network whether disease is over-represented in that small patch of people. Adjustment fits each person's expected hourly activity with a generalized additive model, so a residual of zero means activity matching what is expected for someone of the same age, sex and body mass index at that time of day. Nearest neighbors by Euclidean distance on residual profiles build the network, a greedy set-cover step selects representative seeds, logistic or Cox regression tests enrichment within each neighborhood, and a weighted Benjamini–Hochberg procedure controls multiple testing.

One disease can map to more than one activity pattern, and the two patterns separate on clinical measures. In hypertension the sleep-disrupted group had 45.4% prevalence versus 39.6% in the low-activity group, with higher systolic and diastolic blood pressure; in diabetes the two groups had 10.6% and 8.7% prevalence, and the sleep-disrupted group had higher glycated haemoglobin. In both hypertension and diabetes, average acceleration and the proportion of time in moderate-to-vigorous activity did not change monotonically across the control, low-activity and sleep-disrupted groups.

Both patterns recurred in an independent cohort, though their temporal detail was not identical. The NHANES accelerometry substudy is much smaller, used an exploratory significance threshold and additionally adjusted for ethnicity. Cross-cohort correlations were 0.771 for the hypertension low-activity pattern and 0.598 for the sleep-disrupted pattern, and 0.753 and 0.965 respectively for diabetes.

The analysis covers 24-hour accelerometry from UK Biobank and NHANES, with patterns expressed as hourly averaged activity intensity and disease status taken from registry records or self-report; the findings are observational associations, not causal effects.

Statistical power depends on cohort size, disease prevalence or incidence, and the number of cases inside local neighborhoods, so rare diseases or short-follow-up event analyses may miss patterns; the number of nearest neighbors, the neighborhood-overlap threshold and the Louvain resolution shape neighborhood composition and pattern granularity, and optimal settings may vary across cohorts and activity representations; the observational design cannot separate behavioral risk factors from consequences of disease, and residual confounding may remain.