Less REM on a wrist monitor preceded 83 diseases. Cause or warning?

In 95,559 UK Biobank volunteers who wore a research wristband for a week, more estimated REM sleep went with lower risk of 83 of the 1,049 diseases tested over the next nine years. The wristband infers REM from movement, and more than a third of the links fell below the study's statistical bar once the first two years were set aside. The study can't tell whether short REM harms health or reflects illness already under way.

12 October 2026 · How we rate evidence

For one week between 2013 and 2015, about 100,000 people in Britain wore a small motion sensor on their dominant wrist, day and night. A decade later, a team in Beijing fed those recordings to an AI model that guesses sleep stages from movement. Then they checked the results against every diagnosis recorded in the volunteers' hospital admissions since.

The claim going around. The paper appeared in PLOS Medicine on 17 September. On 23 September ScienceDaily posted the journal's press release as "More REM sleep linked to lower risk of 83 diseases." NDTV asked "Could Getting More REM Sleep Protect You From Dementia And Heart Disease?" and answered "Study Says Yes." Daily Sabah wrote that more REM "may lower risk of Parkinson's, dementia." The paper's abstract is more careful: it says the design "precludes causal inference." The authors close their paper with advice: "Maintaining a sleep duration of 6-8 hours can effectively reduce the risk of multiple diseases." The question readers will ask is simple: should I be trying to get more REM sleep?

What they did. Jingsong Luo, Shengzhi Sun, Rui Chen and colleagues at Capital Medical University and Peking University used 95,559 of the UK Biobank's wristband recordings. An AI model trained on over 1,100 nights of lab sleep studies estimated each night's REM sleep, the stage when most vivid dreaming happens, along with deep and light sleep, total sleep, time awake in the night, and how much sleep varied from night to night. The team then tested each of those six measures against 1,049 diseases, one at a time. They adjusted for age, sex, deprivation, weight, smoking, drinking, exercise and other factors. To guard against flukes, they counted a link only if it cleared a bar about a thousand times stricter than the usual one. They describe the study as exploratory and hypothesis-generating, and their supplement lists every link that passed the usual bar, with its case count, so readers can see well beyond the 83.

What they found. REM stood out. More of it went with lower risk of 83 diseases, more than total sleep time (50) and far more than deep sleep (7), light sleep (7), time awake at night (6) or irregular sleep (3). Because REM minutes rise with total sleep, the team also tested REM as a share of the night, which still went with 54. Every one of the 83 pointed the same way. The typical volunteer logged about 82 minutes of REM a night, and the paper compares people 47.6 minutes apart, roughly someone at the lower quarter (59 minutes) against someone at the upper quarter (107). By that yardstick, dementia's rate was 46% lower, heart failure's 26% lower, and Parkinson's disease's 72% lower. All of these are relative figures.

The paper gives no disease rates by REM level, so we can't turn those percentages into absolute differences. For scale, 458 of about 95,000 volunteers were diagnosed with dementia in hospital during follow-up. That is roughly 5 in 1,000, and about 15 in 1,000 for heart failure (our arithmetic, before any adjustment).

Two corrections to the coverage. The press release gives a hazard ratio of 0.20 for "Parkinson's disease." In the paper, 0.20 belongs to parkinsonism, a separate, much smaller category with just 31 cases. Parkinson's disease itself, with 428 cases, was 0.28. The mix-up is the release's; the paper labels 0.20 as parkinsonism. Daily Sabah's "80% lower risk of Parkinson's disease" repeats the mix-up. Its "people who got sufficient REM sleep" also turns a comparison between people into a threshold, which the paper doesn't give.

What it doesn't show. It can't show that REM protects anyone, and the authors say so: "causal inferences cannot be made." Three features of the data point to other explanations.

First, the direction never varies. Less REM went with more rotator cuff tears, carpal tunnel syndrome and cannabis dependence as well as heart failure and dementia. None of the 83 went the other way. That is the pattern you'd expect if a low REM reading marks people in worse general health, rather than 83 separate effects of REM. The study can't tell those two readings apart.

Second, the washout. When the team set aside diagnoses from the first two years, 52 of the 83 REM links still cleared the strict bar. The authors say this suggests disease already under way "may have influenced sleep patterns." Dementia, heart failure and Parkinson's held.

Third, the wristband infers REM from movement, and the authors call its agreement with lab sleep studies "relatively modest." On average it recorded 17 fewer minutes of REM than the lab did. Illness can change how people move in their sleep. People with REM sleep behaviour disorder act out their dreams, and the disorder often comes years before Parkinson's. The model might score such restless nights as less REM. That's our reasoning; the paper didn't test it. Nor can it rule out medicines: it didn't adjust for them, and many antidepressants suppress REM.

For dementia, other evidence doesn't back it up. An analysis of five US cohorts published in 2025, using a night of lab-grade sleep recording in 4,657 people, found no link between the share of sleep spent in REM and later dementia. It measured REM differently, as a share of one night rather than minutes over a week, and unlike this study it accounted for antidepressant use and APOE, the main genetic risk factor for Alzheimer's.

What it means beyond the study. So, should you chase more REM? This study can't say. Nobody raised anyone's REM and watched what happened. The same goes for the authors' 6–8 hours advice, which rests on the lowest-risk point of curves drawn from the same data. Those hours are on the wristband's own scale: it recorded about 48 minutes more sleep than lab studies do, and it isn't the self-reported sleep behind the usual 7–9 hour guidance, so the two ranges can't be compared directly. Sleeping well is sensible for plenty of other reasons. This study doesn't add a reason to target REM.

Why it's still interesting. The authors call it the most comprehensive map to date of sleep stages measured in ordinary life against a broad range of diseases. REM, not deep sleep, carried most of the signal. A single week on a simple wrist sensor held information about hospital diagnoses nearly a decade later. If that's cause, it's a lead worth testing. If it's an early warning, it's arguably more useful now: a passive check that something is starting. The team published its code and data summaries, so others can test which it is.

Open questions.

Is the wristband seeing REM, or stillness? Repeat the analysis in cohorts whose sleep was staged in the lab, and check the AI model against lab recordings in people with sleep apnea, pain or dream-enacting behaviour. If the device is misreading illness, the advice to get more REM has no basis here, but the wristband could become a screening tool.

Cause or early sign? Look for the links with a gap of five years or more, and follow trials that restore REM, such as treating sleep apnea, for later disease. The answer decides whether REM is something to target or a warning light to watch.

How much do medicines explain? Re-run the analysis with the Biobank's medication records. If drugs that suppress REM account for much of it, the signal is about who takes them.

A note on timing. The paper came out on 17 September 2026 and the coverage followed within days. The wristband recordings themselves are from 2013 to 2015.