I Track My REM Sleep Every Night. This Study Made Me Stop Trusting the Number.

For about a year I have been treating the little REM bar on my sleep tracker like a report card, so when a study landed last week tying more REM sleep to lower odds of 83 different diseases, some smug part of me felt vindicated. Then I read past the headline and the vindication curdled. The study’s own numbers, the ones sitting a few clicks below the press release, run almost exactly backward from the story being sold, and the authors’ own two-year sensitivity test is what quietly sides with the backward reading.

The study itself is big and serious. On September 17, a team led by Shengzhi Sun at Capital Medical University published in PLOS Medicine a phenome-wide analysis of 95,559 people in the UK Biobank who wore wrist accelerometers for seven straight nights. A deep-learning model sorted those wrist movements into sleep stages, and the researchers tracked more than a thousand disease outcomes over a median of nearly nine years. People with more REM sleep, roughly 47 minutes more, showed lower risk across 83 conditions. More deep sleep tracked with lower risk of seven, including type 2 diabetes and major depression. Six to eight hours a night sat in the sweet spot, and under five hours was linked to higher risk of 37 conditions. The top of the REM list is where the headlines came from: dementia at a hazard ratio of 0.54, Parkinson’s at 0.70, heart failure at 0.74.

Now watch that word “linked” do its heavy lifting. This is an observational study, not a trial. Nobody was assigned to sleep differently. And the moment I saw dementia and Parkinson’s sitting at the very top of the protective list, a small alarm went off, because those two diseases have a specific, well-documented relationship with REM sleep, and it is not the flattering one.

So here is the question I could not stop turning over: why would the brain regions that build REM be the same ones a disease takes out first? REM does not run from some tidy central switch. It runs out of the brainstem, out of small deep structures like the locus coeruleus and the circuits that paralyze your body so you do not act out your dreams. Those exact structures are among the first casualties of Parkinson’s and Lewy body dementia. When they start to fail, the paralysis switch loosens and people begin thrashing, punching, shouting through their dreams. That condition has a name, REM sleep behavior disorder, and it is one of the loudest early warnings in all of neurology: more than 80 percent of people with isolated RBD eventually convert to Parkinson’s, Lewy body dementia, or a related disorder, sometimes a decade or more later. So if the disease is physically eating the machinery that makes REM years before anyone gets a diagnosis, of course the people who will later develop Parkinson’s turn up in the data with less REM. Their brains were already dismantling the thing being measured. The low REM reading isn’t causing the disease; it’s smoke from a fire that started years earlier in the brainstem.

And you don’t have to take my word for the direction of the arrow, because the authors ran the test themselves. After correcting for all the multiple comparisons, they were left with 156 significant sleep-disease associations. Then they did what every honest epidemiologist does with a scary observational finding: they threw out everyone diagnosed within two years of the sleep measurement, on the logic that a disease brewing that close was probably already bending the reading. Only 95 of the 156 survived. Nearly 40 percent evaporated. The authors themselves flag that reverse causality is a serious concern, that prodromal disease may have shaped sleep patterns before diagnosis, and that the findings should be read with caution.

That caution never made it into the “83 diseases” headline. The framing outran the study’s own sensitivity analysis, which is a telling direction for the exaggeration to run.

There’s a second, quieter problem underneath all of it: nobody actually measured REM. Gold-standard sleep staging needs electrodes on your scalp reading brain waves. What this study had was a watch reading wrist motion, run through an algorithm called SleepNet, and the authors are candid that its output shows systematic differences from lab polysomnography and represents movement-based estimates rather than exact physiological sleep stages. That’s a fair trade for studying 95,000 people, you can’t wire up a stadium’s worth of volunteers in a lab, but it means the “REM” in the headline is inferred from how still your arm was, not a direct look at your brain.

I also want to flag who this cohort was, because it decides whether any of it applies to you. The UK Biobank skews healthy, wealthy, overwhelmingly white, and comfortably middle-aged, with an average age around 56, a well-known feature of that database rather than a cross-section of the country. The authors name the healthy-volunteer selection bias themselves and warn it limits how far the findings generalize. This is a portrait of a fairly comfortable slice of Britain, not of you and everyone you know.

So what would I actually do with this? Not chase the REM number on my watch, which is exactly the pitch the wellness economy is already spinning up around this paper. There is no REM in a bottle. The shelf is full of OTC sleep aids like doxylamine, and a nightcap is the other reflex, but alcohol suppresses REM in the first half of the night instead of building it, and this study tested no pill or drink at all. What the paper can honestly tell you is thinner and more useful than the headline: protect your shot at six to eight hours, because that part is cheap, biologically plausible, and about the closest thing here to a lever you can actually pull. But the message worth keeping runs backward from “get more REM to dodge dementia.” If your REM sleep collapses, or if you or your partner notice you suddenly acting out your dreams, that is not a wellness metric to optimize, it’s a neurological finding to take to a doctor. Me, I would stop reading my tracker like a scoreboard, and I would start treating a sudden change in how I sleep as a message from the brainstem worth listening to.

Sources

1. PLOS Medicine – Sun et al., accelerometer-derived sleep stages and risk of incident diseases, UK Biobank (Sept 17, 2026): the study itself, source of the hazard ratios and the 156-to-95 sensitivity analysis.
2. ScienceDaily – “More REM sleep linked to lower risk of 83 diseases” (Sept 2026): the press framing that produced the viral headline.
3. Neuroepidemiology – Morbidity, mortality, and conversion to neurodegenerative diseases in patients with REM sleep behavior disorder: evidence on RBD converting to neurodegenerative disease, the reverse-causation mechanism.
4. Journal of Sleep Research – REM sleep and neurodegeneration (2024): why failing brainstem circuits degrade REM before diagnosis.
5. SLEEP (Oxford) – Altered sleep architecture following presleep alcohol (2024): alcohol suppresses REM in the first half of the night.
6. ClinicalTrials.gov NCT07790744 – Prodromal Parkinson’s disease in subjects with REM sleep behavior disorder: brainstem and locus coeruleus damage as the earliest stage.
7. FDA – ANDA070156, doxylamine succinate OTC sleep aid: confirms doxylamine as an approved over-the-counter sleep aid.