Sleep architecture

A night's sleep is a repeating cycle of light, deep and dreaming sleep, and the way those stages fit together explains a lot about how sleep works. It also explains why the "deep sleep" figure on your watch is the least trustworthy number it gives you, and why trying to push it up is a project with no proven payoff.

Stop trying to raise your deep sleep and look after the things that have outcome evidence behind them: about seven hours a night, a regular schedule, screening for sleep apnea if you snore, and no alcohol close to bed. If you wear a tracker, use it for bedtime consistency and total sleep and ignore the stage breakdown.

The four stages, and the two rules that run them

Deep sleep depends on how long you have been awake, and REM depends on what time it is. Almost everything else about sleep structure follows from that.

StageShare of the nightWhat it is
N1~5%The drowsy transition into sleep. Easily woken.
N2~45%Light sleep, and the bulk of the night.
N3~25%Slow-wave, or "deep," sleep. Large, slow brain waves; hardest to wake from.
REM~25%Dreaming sleep. Brain highly active, body paralysed.

Those shares are textbook values for a healthy adult and vary between people and with age. A full cycle averages about 90 minutes, and a typical night runs four to six of them.[1]

Deep sleep dominates the first two cycles and is largely finished by the middle of the night; REM periods start short and lengthen toward morning.

That pattern comes from two systems. Deep sleep reflects sleep pressure, which builds while you are awake and drains away overnight.[2] REM is driven by the body clock, peaking shortly after your core temperature bottoms out before dawn. When volunteers lived on a 28-hour day for over a month, deep sleep faded within every sleep period whenever it fell, while REM followed the body clock.[3]

The idea that the first half of the night restores the body and the second the mind is a memory aid, not a finding: in the classic experiment, remembering word pairs improved more over the deep-sleep-rich first half, and a motor skill more over the REM-rich second.[4]

What actually changes as you age

The age change worth caring about is probably broken sleep, not lost deep sleep.

An older pooled analysis of 65 studies found deep sleep and REM both falling with age.[5] A newer one, 169 studies in 5,273 healthy adults using modern scoring rules, found essentially no change in deep sleep per decade, and no change in light sleep or REM either.[6] The two agree on total sleep and disagree hugely on deep sleep, which points to how deep sleep is scored.

What the newer analysis did find was a clear decline in continuity: total sleep down about 10 minutes per decade, more time awake in the night, and more brief arousals. Unlike deep-sleep percentage, that has outcome evidence. Across three large studies of 8,001 people, the amount of the night spent in brief arousals predicted death from any cause in women; in men the picture was less clear.[7]

Whatever deep sleep you do lose, you lose early. In 149 healthy men, deep sleep fell from 18.9% of the night at ages 16–25 to 3.4% at 36–50, then showed no further significant drop into old age; what kept worsening was time awake and REM.[8] That study was men only and compared different people at different ages, and the plateau partly reflects there being little left to lose.

What the stages seem to do

The stages look like readouts of how healthy you are, not dials you can turn.

Deep sleep (N3) is when growth-hormone release peaks, cortisol is at its lowest, and the body's "fight or flight" nervous system is quietest.[9] It is also said to be when the brain's waste-clearance, or glymphatic, system works hardest, a claim now contested: a 2024 mouse study designed to avoid flaws in the earlier experiments found clearance reduced during sleep,[10] while later work from the original group still finds clearance driven by sleep.[11] In people, one small trial that disrupted deep sleep for a night raised spinal-fluid amyloid-β (the protein that clumps in Alzheimer's) about 10% in a nine-person subgroup, which its authors read as reduced release rather than slower removal, and tau did not move.[12] An industry-funded 2026 trial found no difference in overnight blood amyloid or tau between a night's sleep and a night awake.[13]

REM probably carries the stronger signal about long-term health. Across two large studies, each 5% less REM went with 13–17% higher death from any cause,[14] a finding that has drawn a formal published objection.[15] A 2026 study of heart patients with sleep apnea found the same association, and its authors note that less REM may reflect more advanced disease.[16] REM may simply be a marker of general health rather than a cause of survival.

Stage percentages also discard information: across 36,000 overnight recordings, a model reading the raw brainwave signal predicted thinking ability, disease and death better than the hand-scored numbers, REM share included.[17]

It is observational, the disease may be causing the sleep change, and a much larger analysis found no link with deep sleep measured on a single night.

The figure usually quoted is that each yearly percentage-point decline in deep sleep tracked with about 27% higher dementia risk. That came from 346 adults averaging 69 years old, with 52 dementia cases.[18] The authors themselves note that early, undiagnosed dementia may be what is eroding the deep sleep. Carrying the main Alzheimer's risk gene predicted faster deep-sleep decline, which is what that kind of reverse causation looks like.

A 2025 analysis pooling five American studies, 4,657 people and 998 dementia cases, found no association between any measure of sleep architecture and later dementia: not deep sleep, not REM.[19] The two are not a straight contradiction (one night's deep sleep versus its decline over years), but two of the older paper's senior authors are on the newer one, so this is not a rival camp.

Why your "deep sleep %" is a shaky number

The rule is a convention, not biology. A 30-second scoring window counts as deep sleep when at least six seconds of it show slow brain waves above 75 microvolts, measured peak to peak at the front of the head.[20][21] Brainwave height falls with age and differs between men and women, so the threshold bites unevenly. In 2,913 participants of one large American heart study, scored deep sleep fell with age in men but rose in women, a result that vanished only when slow waves were counted by frequency rather than height; the authors concluded the sex difference may be an artefact of the 75-microvolt rule.[22] Scoring rules have also changed over the decades, so "is 15% deep sleep normal?" has no stable answer.[23]

Wearables can't measure it. Against a full lab sleep study, the best consumer device missed a third of true deep-sleep periods, the worst nearly half.[24] Newer hardware hasn't fixed it: in a 2025 test of six current wrist devices in 62 sleep-clinic patients, the Garmin read deep sleep 44 minutes high and the Apple Watch 25 minutes low on average, so the brand you buy moves the number more than anything you do at night.[25] Rings do no better.[26] The device-by-device figures are on Sleep trackers.

It wobbles even when measured properly. Put the same older adult in a sleep lab two nights running, with a human scoring both, and deep-sleep percentage is only moderately reproducible. The underlying continuous brainwave signal is steadier.[27]

Devices such as Fitbit are trained to copy those human-scored labels,[28] so optimising your score means optimising a noisy estimate of a scoring convention.

Should you try to increase deep sleep?

No. Your brain already defends it, the things that raise it have not been shown to help, and the one everyday thing that seems to raise it is harmful.

Your brain defends it. In a study of 35 young adults, a week of either 6 or 10 hours in bed barely moved deep-sleep activity; REM absorbed the loss. After about 40 hours awake, deep-sleep activity rebounded to 140–152% of normal.[29] That rebound is weaker in middle age,[30] and it says nothing about damage from sleep apnea or alcohol. Most importantly, a defended deep-sleep number is not a recovered brain: under sustained short sleep, deep-sleep activity stays near normal while alertness and performance keep getting worse.[31]

Exercise and warm baths help sleep, not deep sleep. Pooling 66 studies, regular exercise improved total sleep, efficiency and how fast people fell asleep, with no clear deep-sleep benefit; a single workout had only a small effect.[32] A warm bath or shower 1–2 hours before bed helps you fall asleep faster and sleep better, but not by adding deep sleep.[33]

Sound-stimulation gadgets change the brainwaves and nothing else. Soft noise pulses timed to your slow waves, the idea behind acoustic stimulation headbands, do boost them. Across 12 studies and 206 people, the average effect on memory was close to nothing, and the effect shrank steadily with each year of publication, from a large benefit in 2013 to a moderate harm by 2021.[34] The only pooled analysis to report a large benefit has been retracted.[35]

Drugs that raise deep sleep rarely help. Across 27 randomised trials of such drugs, the extra deep sleep rarely translated into better thinking.[36] The clearest case is gaboxadol, a sleeping drug that reliably raised deep sleep in insomnia patients.[37] In two three-month trials it beat placebo on self-reported sleep and daytime function in the first (927 patients), by only about twenty minutes of sleep, and did neither in the second.[38] It was dropped in 2007, the company citing "limited or variable efficacy" and psychiatric effects seen above therapeutic doses in a separate study.

Alcohol probably raises it, and is bad for you. A 2013 review concluded that across doses, ages and sexes, alcohol increases deep sleep in the first half of the night.[39] Eight sleep researchers formally challenged that review's deep-sleep statistics,[40] and the authors disputed the challenge.[41] A 2025 meta-analysis found only that a high dose, about five drinks, gets you into deep sleep sooner, while REM is disrupted from about two drinks and worse the more you drink.[42] Either way, the tracker number may go up on a night whose REM is worse. See alcohol.

What to do instead

  1. Get enough total sleep, around seven hours. Death rates rise in both directions from about seven hours, roughly twice as fast on the long side as the short.[43] These are self-reported hours in observational studies, and illness causing long sleep is the usual explanation for the long side. Unbroken sleep matters too: in the small experiment that isolated deep sleep, poorer home sleep efficiency over the preceding six nights went with higher tau (14 people), while the deep-sleep disruption did not move it.[44]
  2. Keep your timing regular. In one large study of about 61,000 UK adults who wore activity trackers, how regular their sleep timing was predicted death from any cause better than how long they slept.[45] See Circadian rhythms.
  3. Get screened for sleep apnea if you snore, someone has seen you stop breathing, or you have high blood pressure. Untreated apnea genuinely wrecks sleep, and unlike your deep-sleep score it can be measured and treated. See Sleep-disordered breathing.
  4. For chronic insomnia, do cognitive behavioural therapy for insomnia (CBT-I), the first-line treatment. See Treating chronic insomnia.
  5. Don't drink close to bedtime, for the sake of REM and the second half of the night.

Who should be careful with trackers

Anxiously chasing sleep scores has a name, orthosomnia, coined in a report of three patients whose pursuit of perfect tracker numbers, the authors suggest, may have worsened their insomnia.[46] Whether trackers cause this is barely tested; the one small controlled comparison found no difference between insomnia patients given a tracker (plus a self-help booklet) and those given a paper diary.[47]

A tracker is fine for bedtime consistency and total sleep, though in older adults even that is less certain: one small study found all four devices tested reading total sleep low,[48] others have found the opposite.[49] Supplement claims built on tracker "deep sleep scores" are close to uninterpretable. See Sleep supplements.

Further reading

  • Patel AK, Reddy V, Shumway KR, Araujo JF. Physiology, Sleep Stages. StatPearls 2024.[50]
  • Skorucak J, Arbon EL, Dijk DJ, Achermann P. Response to chronic sleep restriction, extension, and subsequent total sleep deprivation in humans: adaptation or preserved sleep homeostasis? Sleep 2018.[51]
  • Boulos MI, et al. Normal polysomnography parameters in healthy adults: a systematic review and meta-analysis. Lancet Respiratory Medicine 2019.[52]
  • Leary EB, et al. Association of rapid eye movement sleep with mortality in middle-aged and older adults. JAMA Neurology 2020.[53]
  • Himali JJ, et al. Association between slow-wave sleep loss and incident dementia. JAMA Neurology 2023.[54]
  • Yiallourou S, et al. Sleep architecture and dementia risk in adults: an analysis of 5 cohorts from the Sleep and Dementia Consortium. Sleep 2025 — 4,657 participants, 998 cases.[55]
  • Davidson S, Sharman R, Kyle SD, Tarassenko L. Is it time to revisit the scoring of slow wave (N3) sleep? Sleep 2025 — the 75-microvolt criterion and the artefactual sex difference.[56]
  • Chinoy ED, et al. Performance of seven consumer sleep-tracking devices compared with polysomnography. Sleep 2021.[57]
  • Schyvens AM, et al. A performance validation of six commercial wrist-worn wearable sleep-tracking devices for sleep stage scoring compared to polysomnography. Sleep Adv 2025 — n=62, recruited from a hospital sleep clinic.[58]
  • Roth T, Lines C, Vandormael K, Ceesay P, Anderson D, Snavely D. Effect of gaboxadol on patient-reported measures of sleep and waking function in patients with Primary Insomnia: results from two randomized, controlled, 3-month studies. J Clin Sleep Med 2010 — inconsistent patient benefit across the two trials.[59]
  • Baandrup L, et al. Pharmacological interventions to enhance slow wave sleep and cognitive performance: a systematic review. Sleep Med Rev 2026 — 27 randomised trials.[60]
  • Ebrahim IO, et al. Alcohol and sleep I: effects on normal sleep. Alcohol Clin Exp Res 2013.[61]
  • Baron KG, et al. Orthosomnia: are some patients taking the quantified self too far? J Clin Sleep Med 2017.[62]

— § —