REM Sleep and 83 Conditions: What a Wrist Can and Can’t See

Between 2013 and 2015, about a hundred thousand people in Britain strapped a small grey sensor to their wrist and wore it for a week. Nobody asked them to sleep differently. They just slept, and the sensor felt every shift of their arm. More than a decade later, a team of researchers used those seven nights to ask a large question: does the kind of sleep you get line up with what happens to your health afterwards?

The answer, published in PLOS Medicine on 17 September, made headlines because of one number. People with more REM sleep had a lower recorded risk of 83 different conditions. That is a striking list. It is also a good moment to slow down, because the most dramatic result in the paper is the one I trust least, and the reason why says something useful about every sleep score you will ever see on a screen.

Seven nights, 1,049 outcomes

The study, led by Jingsong Luo and Rui Chen, used 95,559 UK Biobank participants, mean age 56, followed for a median of 8.9 years. Each had worn a wrist accelerometer for a week. A deep-learning model called SleepNet, trained on more than 1,100 nights of combined lab sleep recordings and wrist data, turned that movement into estimates of total sleep, REM, deep and light sleep, time awake after falling asleep, and night-to-night irregularity.

The team then did something ambitious. Instead of testing one disease, they tested 1,049 of them, drawn from hospital records. With that many tests, false alarms are guaranteed, so they set a very strict threshold. 156 associations cleared it. REM sleep accounted for 83 of them. Total sleep time accounted for 50. Deep sleep, the stage most of the wellness world worships, managed seven.

That last detail deserves a moment. For a decade, deep sleep has been sold as the restorative stage, the one to chase. In this dataset, REM did far more of the statistical work.

What “more REM” means in this paper

The hazard ratios are reported per interquartile range, which here means roughly 47 extra minutes of REM a night. The typical participant logged about 82 minutes. So the comparison is not between a little more and a little less. It is between people whose nights looked meaningfully different, every night, for a week.

Across that gap, the associations were large. The recorded hazard of dementias was about half. Heart failure was about a quarter lower. Parkinsonism, the umbrella term for Parkinson’s disease and similar movement conditions, showed a hazard ratio of 0.20: an 80% lower risk. The authors also found that 6 to 8 hours of total sleep sat at the low point of risk, and that nights under 6 hours tracked with more conditions, a finding that agrees with a lot of older work.

The biggest number is the one to distrust

An 80% lower risk from a sleep stage is not the kind of effect biology usually hands out. When a result looks that strong, the first question is not “how does REM protect?” but “what else could make the line so steep?”

Parkinsonism has a well-known head start in sleep. Normally, during REM, your muscles go almost completely slack, which is why you do not act out your dreams. In a condition called REM sleep behaviour disorder, that slackness fails and people move, punch or shout during dreams. It is a strong early signal: in a multicentre study of 1,280 patients, Postuma and colleagues found that 28% developed a condition such as Parkinson’s disease within an average of 3.6 years of follow-up, and about three in four of those followed for 12 years did.

Now put that next to the measurement. SleepNet only feels movement. A stage defined partly by stillness is scored from a sensor that cannot see brain waves or eyes. If someone moves during REM, a movement-based model may record that time as light sleep or wake. I want to flag this as my reading, not the authors’ finding. But it means low “REM” in this dataset could partly be an early sign of the same process the study then counts as an outcome. In that case the arrow runs backwards, and the 0.20 is the disease talking first.

The authors saw the general risk. When they excluded everyone diagnosed in the first two years, 95 of the 156 associations survived. That is reassuring for many results. It also means 61 did not, which tells you how much early illness was shaping the picture.

A wrist that only feels movement

The second problem is quieter. The paper reports how well SleepNet agreed with lab polysomnography, the gold standard that records brain waves, eye movements and muscle tone. For sorting each slice of the night into REM, non-REM or wake, its F1 score was 0.49, where 1.0 is perfect. On average it underestimated REM by about 17 minutes and overestimated total sleep by about 48 minutes.

This is not a criticism of the study. With 95,000 people, random errors largely wash out, and a noisy ruler can still find real patterns in a crowd. The problem appears when the same logic is applied to one person. Your single night does not get averaged with 95,000 others. Your REM estimate on Tuesday is a model’s guess, and the error bar on that guess is wider than most of the night-to-night changes you will be tempted to read into it.

That is why we label sleep stages as estimates rather than measurements, including on our own devices. Time asleep and wake-ups are derived from much firmer signals. Stages are a modelled inference on top of them.

What the lab adds

If wrist REM were the only evidence, the safest response would be a shrug. It is not. In 2020, Eileen Leary and colleagues looked at REM measured in the lab in two separate cohorts, a group of older men followed for about 12 years and a mixed-sex Wisconsin cohort followed for about 21. In the first, every 5 percentage-point drop in the share of the night spent in REM went with roughly 13% higher mortality, and the second cohort pointed the same way.

That matters because lab REM does not depend on movement. It cannot be explained by the algorithm problem above. So something real seems to connect REM and long-term health. What remains unknown is which way the link runs, and whether a person who adds REM gains anything. No trial has tested that, and it would be hard to design one.

So the honest summary is narrow. REM appears to be a meaningful signal of how a body is doing over time. It is not yet a lever with a known effect.

Where REM lives in your night

You cannot order REM directly. You can, however, stop cutting it off. Sleep runs in cycles of roughly 90 minutes, and the REM portion of each cycle grows as the night goes on. The early night is deep-sleep heavy. The last third is where most REM sits, as we walk through in sleep stages explained.

This has a plain consequence. Sleep 6 hours instead of 7.5 and you do not lose a fifth of every stage evenly. You lose the end of the night, and the end of the night is REM-heavy. A late bedtime with a fixed alarm is, in effect, a REM tax.

Alcohol is the other common thief. A review of the alcohol and sleep literature by Ebrahim and colleagues found that alcohol before bed delays the first REM period at all doses studied and, at moderate and higher doses, lowers the share of the night spent in REM, while sleep in the second half becomes more broken. The drink that makes you fall asleep faster takes its payment later, in exactly the stage this study put at the top of the table. We go further into that trade in alcohol, sleep and recovery.

What to do with your own REM number

Three things are worth trying this fortnight, and none of them needs a REM score to work.

First, protect the end of the night. If you need more sleep, move bedtime earlier rather than hoping to sleep later. A steady wake time also helps, and regular timing matters about as much as duration.

Second, run a simple alcohol comparison. A week of evenings with no drink within a few hours of bed against your usual week. Watch how you feel on waking and whether the second half of your night looks less broken. Do not judge it by one night.

Third, read stage numbers as weather, not climate. If your Agen Band shows a low-REM night, it is a weak signal from a modelled estimate. A sustained drift over several weeks, alongside how you actually feel, is worth noticing. Persistent changes in sleep, especially acting out dreams or a partner noticing movements during sleep, are a conversation for your doctor, not your app.

The bottom line

This is a big, careful study with an honest limitations section, and its most quotable number is its least reliable one. REM seems to matter, and lab studies back that up. But the striking 80% figure is most likely a mix of early illness and a movement sensor guessing at a stage defined by stillness.

The useful message is older and simpler. Get 6 to 8 hours, keep your timing steady, and do not cut the end of the night or soak it in alcohol. That protects REM without a single score. For how sleep fits into the rest of a measured routine, start with how to build a longevity protocol, or read how REM itself was found in the story of its discovery.