The dream before the alarm

23:00

You close your eyes. Over the next hours you will go down into sleep and come back up several times, in cycles. Across 6,064 cycles recorded in a sleep laboratory in Basel, the median cycle lasted 96 minutes, and the first cycle of the night was consistently shorter than the rest.

The night begins with non-REM sleep. Your eyes lie still under the lids.

In 1953, Aserinsky and Kleitman woke sleepers 23 times while their eyes were still. Nineteen times, the sleeper recalled nothing at all.

Dreams happen here too. In 2017, Siclari and colleagues woke people from non-REM sleep after reading their brain activity. When the activity pointed to a dream, the sleeper described one 33 times out of 36.

In the schematic model, your first REM sleep begins around 00:11, named for the rapid eye movements that give it away. In the drawing of your night it lasts only a few minutes; the later ones lengthen towards a peak, then shorten.

Almost every muscle you control goes slack. A 2023 review of how the brain produces REM sleep lists the exceptions: the diaphragm and the other breathing muscles, the muscles that move the eyes, and the tiny muscles of the ear.

Your breathing changes too. Aserinsky and Kleitman counted 16.9 breaths a minute while the eyes moved, against 13.4 while they were still. If the sound is on, that is the rhythm you hear now.

1951

Late in 1951, Aserinsky made the first all-night recording of a sleeper's eyes and brain waves together. The sleeper was his own eight-year-old son. So he wrote, looking back, in 1996.

He first wanted to call the episodes jerky eye movement periods. Kleitman, he recalled, played the sceptic.

1953

Their paper in 1953 tested what the movements meant. Waking someone during eye-movement periods was much more likely to yield a detailed dream report: 20 times out of 27, usually full of images. Woken while the eyes were still, sleepers recalled nothing 19 times out of 23.

Later work found dreams outside those periods too, as Siclari's did. But the moving eyes marked the moments when a detailed dream report was most likely.

Why don't you act out what you dream? In 1982, Hendricks, Morrison and Mann made lesions on both sides of the pons, part of the brainstem, of 25 cats. The cats still entered REM sleep, but without the slack muscles that normally come with it.

Eight lifted their heads and made movements that looked like orienting, staring, reaching and trying to stand. Five moved only their limbs or heads.

In people, the same failure has a name: REM sleep behaviour disorder. The sleeper acts out dreams. In a study of 1,280 patients across 24 centres, the estimated share who had developed parkinsonism or dementia reached 73.5% within 12 years.

So REM sleep is where dreams are most often found, and it can also show a brain in trouble years before a diagnosis.

Each cycle ends in REM sleep. Over the first cycles the REM periods lengthen; in the second half of the night they peak and then ease off, and what keeps rising to the morning is REM's share of sleep. Aserinsky and Kleitman saw the eye-movement periods come back a second and a third time, and rarely a fourth.

In the schematic model of your night, the longest REM period runs from 04:40 to 05:08, in the second half of the night. Any REM periods after it are drawn a little shorter. Real nights vary a lot: the measured pattern is a curve that peaks mid-to-late, not a rule that the last one is the longest.

With age, REM sleep at the end of the night tends to shrink: in the Basel recordings, older people had shorter REM sleep, particularly at the end of the night.

17 September 2026

A study in PLOS Medicine followed 95,559 adults in the UK Biobank for a median of 8.9 years. Those with more REM sleep were less likely to develop 83 diseases.

For every extra 47.6 minutes of estimated REM a night, roughly the gap between a night with 59 minutes of REM and one with 107, the hazard of dementia over follow-up was 46% lower (hazard ratio 0.54), of heart failure 26% lower (0.74) and of Parkinson's disease 30% lower (0.70).

The press release gave Parkinson's as 0.20, a cut of 80%. In the paper, 0.20 is the figure for parkinsonism, a different diagnosis.

But nobody in the study had their REM sleep recorded from the brain. They wore an accelerometer on the wrist for a week, and a computer model inferred their sleep stages from the movement.

Tested against brain recordings, the model's agreement was moderate (F1 score 0.49, where 1 is perfect). On average it counted 17.1 minutes of REM too few, and for a single night 95% of its errors fell between 56.7 minutes too few and 91 too many. That range is wider than the 47.6 minutes the hazard ratios are quoted per.

The authors expect these errors to be spread evenly between people who later fell ill and people who did not, so the comparison across the group can still hold. What the model cannot do is tell one person how much REM they had.

The authors also write that, as an observational analysis, “causal inferences cannot be made”.

There is also an important counterexample. People taking several kinds of antidepressant, including SSRIs and MAOIs, have what the sleep researcher Siegel called “a massive suppression of REM sleep”. Reviewing the evidence in 2011, he wrote that they “do not go insane and do not have any apparent impairment of memory”, and that long-term use “does not generally have deleterious effects on health”. He pointed to people whose medicines strongly suppress REM without an obvious corresponding memory or health deficit: evidence against a simple claim that less REM itself must be harmful.

Smaller studies did record the brain. In 2017, Pase and colleagues recorded one night of sleep in 321 people over 60 and followed them for up to 19 years; 32 developed dementia. Each percentage point less REM went with about 9% higher risk.

In 2022, Otaiku followed 2,770 older men after one recorded night; 70 developed Parkinson's disease. Those in the lowest quarter for REM had 2.58 times the odds of those in the highest.

In both, the low REM was measured years before the diagnosis.

Put together, REM sleep may be less a shield than a window: a view of the brain's health, years ahead.

Whether low REM is an early sign of disease or one of its causes is an open question: Pase and colleagues judged an early sign “less likely”, because excluding the first three years of follow-up did not remove the link. But excluding the first three years cannot rule out a prodrome that begins much earlier. Otaiku concluded that the sleep abnormalities “precede the development of PD by several years”, without settling which way the arrow points.

Roffwarg and colleagues reported in 1966 that newborns spend about half their sleep in REM.

In adults the share is far smaller and changes slowly. Pooling hundreds of sleep studies, Floyd and colleagues estimated 21.7% at 19, 21.2% at 40, 18.8% at 75 and 20.4% at 85.

For your night, the page uses 21.2%, the value for age 40.

You never see your own REM sleep. If someone sleeps beside you, you can ask them to wake you gently at around 04:54, a time from the schematic model, and ask what you were dreaming.

In 1953, waking sleepers during eye-movement periods yielded a detailed dream report about three times in four.