Summary
A study of 95,559 UK Biobank participants found that accelerometer-estimated REM sleep was associated with lower incidence of 83 diseases. The observational analysis also found that estimated minimum-risk sleep duration for most of the conditions with non-linear associations fell between 6 and 8 hours.
A prospective study in PLOS Medicine linked greater accelerometer-estimated REM sleep with lower incidence of 83 diseases among 95,559 UK Biobank participants. The researchers also found that the sleep durations associated with the lowest predicted risk for most conditions showing non-linear patterns clustered between 6 and 8 hours. These are observational associations, not evidence that changing sleep duration or REM sleep would prevent disease.
Participants wore wrist accelerometers for seven days. Researchers used the SleepNet algorithm to estimate sleep stages and other patterns, then tracked incident inpatient diagnoses across 1,049 disease outcomes. Median follow-up was 8.9 years. The analysis used Cox regression adjusted for demographic, lifestyle and environmental factors, and applied a strict correction for multiple comparisons.
What the cohort found
An interquartile-range difference of 47.6 minutes in estimated REM sleep was associated with lower incidence of 83 diseases across 12 categories. For example, the hazard ratio for heart failure was 0.74 (95% confidence interval 0.68–0.80), and for dementia it was 0.54 (0.47–0.62). A hazard ratio compares the rate of an outcome over follow-up between groups; it is not an estimate of an individual’s absolute chance of developing a condition.
Greater estimated deep sleep was associated with lower incidence of seven diseases. The researchers also found associations between more wakefulness after sleep onset and six diseases, and between greater sleep irregularity and three. These results describe patterns across the cohort, rather than outcomes predicted for any one person.
Sleep duration did not have a single linear relationship with every outcome. Restricted cubic spline analyses found statistically significant non-linear associations for 86 disease phenotypes. For 69 of those, the modelled minimum-risk duration fell within the 6–8-hour range. In comparisons using that range as the reference, short sleep—particularly less than five hours—accounted for most of the significant higher-risk associations. The modelled minima are estimates from this cohort, not a clinical prescription.
What the measurements can tell us
Polysomnography, which records brain activity and other physiological signals, is the standard method for measuring sleep stages. It is difficult to use at the scale of a large population cohort. Here, SleepNet inferred sleep patterns from wrist movement. The algorithm had been trained using more than 1,100 nights of concurrent accelerometer and laboratory polysomnography data. In the study’s three-class REM/NREM/wake classification, it achieved an F1 score of 0.49 against polysomnography, indicating moderate classification performance. The resulting stage estimates should therefore be understood as movement-based estimates, not direct measurements of brain-defined sleep stages.
The study’s prospective design and large number of participants allowed researchers to map associations across a wide range of later diagnoses. But observational analysis cannot determine whether sleep patterns contribute to disease or reflect other influences, including early disease. The measurements covered one week, and participants were mostly White UK Biobank volunteers; outcomes were identified from inpatient records. Whether interventions that change sleep duration or sleep-stage patterns alter disease incidence requires further study.