Sleep plays a central role in health, but most large epidemiological studies have relied on self-reported sleep duration, which correlates poorly with objective physiological measures. Polysomnography (PSG) remains the gold standard for measuring sleep architecture (REM, N1, N2, N3) but is impractical for very large cohorts. Wrist-worn accelerometers have become widely used for objective sleep measurement in middle-aged and older populations and, when combined with validated machine-learning approaches, can estimate sleep stages and other real-world sleep metrics at scale.
This study leveraged a self-supervised algorithm (SleepNet) to derive sleep-stage and fragmentation metrics from wrist accelerometer recordings in UK Biobank participants. The authors performed a phenome-wide association analysis to map the relationships between these objectively measured sleep features and incident disease across multiple organ systems.
Population and data sources
The analysis included 95,559 UK Biobank participants who wore wrist accelerometers. Sleep-derived metrics were generated from these accelerometer signals using the SleepNet algorithm validated against laboratory-based polysomnography. Participants were followed from the date of accelerometer wear until the first occurrence of a diagnosed disease, death, or April 1, 2024, with a reported median follow-up of 8.9 years.
Sleep metrics derived
Key sleep variables derived from accelerometer data included estimates of sleep-stage amounts (rapid eye movement REM, N1, N2, N3), total sleep duration, sleep irregularity, and wake after sleep onset (WASO). These provided measures of both sleep architecture and continuity in a naturalistic, real-world setting.
Outcomes and statistical approach
A phenome-wide association study (PheWAS) examined incident diagnoses across 1,049 health outcomes. Cox proportional hazards regression models were used to estimate associations between sleep metrics and incident disease, adjusting for demographic characteristics, lifestyle factors, and environmental exposures. Restricted cubic spline (RCS) analyses were applied to evaluate potential non-linear relationships between sleep duration and disease outcomes. Multiple testing correction (Bonferroni) was applied in the PheWAS.
Overview of associations
Across the examined phenome, variations in derived sleep patterns were associated with 156 incident disease outcomes after correction for multiple comparisons. Distinct sleep features showed differing patterns of association with disease risk.
Sleep-stage associations
Sleep fragmentation and irregularity
Sleep duration and non-linear risk patterns
Restricted cubic spline analyses identified significant non-linear associations between total sleep duration and 86 disease phenotypes (P for nonlinear < 0.05). For 69 of these phenotypes, the duration associated with lowest risk concentrated within the 6–8 hours window. In category-specific analyses, extreme short sleep (<5 hours) related to the most widespread clinical vulnerabilities, accounting for 37 of 41 identified significant adverse associations when compared with the reference group of 6–8 hours.
This large-scale cohort study using accelerometer-derived sleep metrics provides an extensive phenome-wide map of associations between real-world sleep architecture, continuity, duration, and incident disease. The principal findings are that greater REM and deep sleep amounts are generally associated with lower incidence for multiple outcomes, whereas increased sleep irregularity and WASO are linked to elevated risk in several conditions.
The non-linear relationships between sleep duration and disease risk, with minimal-risk durations concentrated in 6–8 hours for most identified phenotypes, align with prior literature that suggests both short and long sleep can be associated with adverse outcomes. The study adds specificity by linking sleep-stage composition and fragmentation metrics to a broad set of disease endpoints.
The authors note key limitations inherent in the study design. First, the observational nature of the cohort precludes causal inference; residual confounding may remain despite adjustment for multiple covariates. Second, although SleepNet has been validated against PSG in prior work, accelerometer-derived sleep staging is an indirect measure compared with laboratory polysomnography and may have measurement error. Third, the source data and analytic choices reported in the article determine the scope of inference; any further subgroup or mechanistic interpretation beyond what was reported would require additional targeted studies.
Using accelerometer data and a validated machine-learning approach in nearly 96,000 participants, this study identifies widespread associations between real-world sleep architecture, continuity, duration, and incident disease across multiple systems. Greater amounts of REM and deep sleep were generally associated with lower disease risk, while greater sleep irregularity and WASO were linked to increased risk. Non-linear analyses indicate that a 6–8 hours sleep duration window corresponds to lowest risk for the majority of implicated phenotypes. The findings emphasize the potential utility of objectively measured sleep features for risk stratification and prevention research, while underscoring that causality cannot be assumed from the observed associations.