Sleep architecture

Sleep alternates between non-rapid-eye-movement (NREM) stages and rapid-eye-movement (REM) sleep in an orderly pattern that a technician can read like a score. That structure is what "sleep architecture" means. This page covers what the stages are, what a normal night looks like, how the pattern shifts with age, what each stage appears to do — and why the popular project of maximising deep sleep rests on much weaker ground than it seems.

What the evidence says

Strong:

• Sleep cycles through NREM stages N1, N2, and N3 (slow-wave, or "deep") and REM, in cycles averaging about 90 minutes, four to six times a night. Deep sleep is front-loaded into the early cycles; REM lengthens toward morning.^[1]
• Deep sleep is homeostatically defended: a week of sleep restriction leaves slow-wave activity almost untouched, and it rebounds to 140–152% of baseline after deprivation.^[2]
• Consumer devices cannot stage sleep reliably, missing roughly a third to half of true deep-sleep epochs.^[3]

Moderate:

• What clearly degrades with age is sleep continuity — less total sleep, lower efficiency, more time awake — rather than stage composition.^[4]
• Deep sleep coincides with peak growth-hormone release.^[5]
• The N3 scoring rule is a measurement convention, not a biological threshold, and its fixed 75-microvolt amplitude criterion produces sex differences the same recordings do not show when rescored.^[6]
• Raising deep sleep has not, on the evidence so far, been shown to improve anything. Acoustic stimulation enhances slow oscillations without improving memory,^[7] and across 27 randomised trials of drugs that raise slow-wave sleep, the extra deep sleep rarely translated into cognitive benefit.^[8]
• REM, not deep sleep, carries the outcome signal. Compared head-to-head across two large cohorts, each 5% reduction in REM was associated with 13–17% higher all-cause mortality.^[9]

Weak / preliminary:

• The link between declining deep sleep and dementia is observational, the authors concede early dementia may be causing the sleep change,^[10] and a much larger pooled analysis found no association between any sleep-architecture measure and later dementia.^[11]
• Exercise raises deep sleep only slightly and acutely; the largest meta-analysis to measure sleep stages reports no slow-wave benefit from regular training.^[12]
• Whether deep sleep clears amyloid-β from the brain is less settled than it is usually made to sound: the crossover trial that disrupted slow-wave sleep and sampled cerebrospinal fluid reads its own result as reduced amyloid release rather than faster removal,^[13] and the industry-funded 2026 plasma trial's direct sleep-versus-deprivation comparison was null.^[14]

Caution:

• Alcohol reliably increases deep sleep in the first half of the night while degrading the night in every other way.^[15]
• Anxiously chasing tracker metrics has a clinical name — orthosomnia — coined in a report of three patients whose sleep worsened as they tried to perfect their scores.^[16] Whether trackers cause the problem or just attract people who already have it is untested: the one small randomised comparison found no difference between insomnia patients given a tracker and those given a paper diary.^[17]

One note on the numbers below. Where a figure below is followed by a range marked "95% CI", that range is where the true effect most plausibly lies; when it spans zero, the result is compatible with no effect at all.

The four stages

Those shares are conventional textbook values for a healthy adult, and vary between people and across the lifespan. A full cycle averages about 90 minutes, and a typical night runs four to six of them.^[18]

The stages are not spread evenly. Deep sleep dominates the first two cycles and has largely finished by the middle of the night, while REM periods start short and lengthen toward morning. This is why the first half of the night is loosely called the physically restorative half and the second half the mentally restorative one. Treat that as a mnemonic rather than a finding: the classic experiment that split the night this way found declarative memory — remembering word pairs, the most "mental" task it tested — improved more over the deep-sleep-rich first half, while a motor skill improved more over the REM-rich second.^[19]

That front-loading is not arbitrary; it falls out of the two systems that govern sleep. Slow-wave activity is the electrical signature of Process S, the homeostatic sleep pressure that builds while you are awake and discharges roughly exponentially across successive cycles.^[20] REM is instead under circadian control, cresting shortly after your core body temperature bottoms out before dawn. In a forced-desynchrony experiment where volunteers lived on a 28-hour day for over a month, slow-wave activity decayed within every sleep episode regardless of when it fell, while REM tracked the body clock.^[21]

In short: deep sleep answers to how long you have been awake, and REM answers to what time it is. Almost everything else about architecture follows from those two rules.

How the pattern changes with age

The received wisdom is that slow-wave sleep collapses as you get older. The evidence is messier.

The classic reference — a meta-analysis of 65 studies and 3,577 people — found slow-wave sleep and REM both falling with age while light sleep and wakefulness rise.^[22] But a newer meta-analysis of 169 studies in 5,273 healthy adults, using modern scoring rules, put the per-decade change in deep sleep at −0.1%, not statistically significant, with null changes in N2 and REM as well.^[23] The two agree closely on total sleep time while diverging roughly twentyfold on deep sleep — which points at how N3 is scored rather than at either study's quality.

What the newer analysis found unambiguously degrading was continuity: total sleep time down about 10 minutes per decade, sleep efficiency down 2.1%, time awake after falling asleep up nearly 10 minutes, and more arousals. That is the age change to take seriously — and unlike the stage percentages, it has outcome evidence behind it. Across three large cohorts and 8,001 people, how much of the night was spent in brief arousals predicted all-cause mortality in women; in men the picture was less clear, with half the estimates compatible with no effect.^[24]

Where the older literature is clearer is that whatever deep sleep you lose, you lose early. Among 149 healthy men aged 16 to 83, slow-wave sleep fell from 18.9% of the night in the 16–25 group to 3.4% in the 36–50 group — and then showed no further significant decline into old age. What deteriorated after midlife was time awake (up ~28 minutes per decade) and REM (down ~10 minutes per decade).^[25] By the time you are 50 and worrying about deep sleep, most of the deep sleep you were ever going to lose is already gone. (That study was men only and cross-sectional, and the midlife plateau is partly a floor effect — there is little left to lose, which is not the same as being protected.)

What the stages appear to do

N2 is where sleep spindles — brief bursts of oscillation — appear, and they are associated with consolidating what you learned during the day. It is also, simply, most of your night.

N3 is when growth-hormone secretion peaks, cortisol reaches its low point, and sympathetic tone is lowest.^[26] It is also when the brain's waste-clearance system is usually said to be most active — though that is now contested rather than settled. A 2024 study designed to avoid the flaw in the original experiments found the opposite in mice: clearance was reduced during sleep and under anaesthesia, and deeper slow-wave activity went with less clearance, not more.^[27] A 2025 rebuttal identified a mechanism that would explain the anaesthetic half without conceding the sleep half.^[28] The best experimental evidence that N3 specifically matters comes from one small crossover trial: 17 adults had their slow-wave activity selectively disrupted with acoustic tones while cerebrospinal fluid was sampled. Amyloid-β rose about 10% in responders, and the effect was specific to slow-wave disruption rather than to total sleep time.^[29] Ju's own authors read this as slow-wave sleep reducing amyloid release rather than speeding its removal — total protein and the glial marker YKL-40 did not move. That is a real result — and it is a nine-person responder subgroup over a single night, measuring a biomarker rather than an outcome. Notably tau did not move at all; what tau tracked was six-night sleep efficiency.

A 2026 randomised crossover trial approached the same question from the other end — measuring what arrives in the blood rather than what is left behind in the brain. In 39 healthy adults aged 49–66, the direct comparison found nothing: the overnight change in plasma amyloid-β and tau was statistically indistinguishable after a night of normal sleep and after a night without sleep, on every one of the ten comparisons the trial reported.^[30] The paper's headline is a model result rather than a measurement: feeding the trial's electroencephalogram (EEG) and cardiovascular readouts into a statistical model does predict higher morning biomarker levels, and the authors read the null direct comparison as clearance and production cancelling each other out during sleep. That includes the title's "amyloid and tau" — the tau markers were as null as the amyloid ones. Two further reasons to hold it lightly. The measurements came from an investigational in-ear device rather than established instrumentation. And the trial was funded by Applied Cognition, which makes that device and employs the first author and three co-authors, with two further authors at the company selling the plasma assays — a strong result reported by interested parties. It is the most direct attempt yet to measure the pathway in humans; it is not independent confirmation.

REM carries the stronger outcome signal of the two. When all stages were compared head-to-head across two large cohorts, a machine-learning model ranked REM, not deep sleep, as the stage most important for predicting survival, with each 5% reduction associated with 13–17% higher all-cause mortality.^[31] The accompanying editorial cautioned that REM quantity is "likely a biomarker of general health rather than having a direct causal link to mortality" — which is the right way to read all of these associations. The finding has also drawn a formal published objection, with a reply from the authors — a dispute worth knowing about before leaning on it,^[32] and a 2026 cohort of patients with coronary disease and sleep apnea has since reproduced the association.^[33] Stages look like readouts of how healthy you are, not dials you can turn. A 2026 analysis of 36,000 overnight recordings put that to the test directly: a model reading the raw sleep electroencephalogram predicted cognition, disease and death better than the hand-scored numbers did — including REM fraction and spindle density.^[34] The information is in the signal. It is the act of compressing it into stage percentages that throws it away.

The dementia association is the one most often quoted: each percentage-point-per-year decline in slow-wave sleep tracks with about 27% higher dementia risk.^[35] It deserves more scepticism than it gets. The cohort was 346 adults averaging 69 years old, with 52 dementia cases. The paper states that "it is possible that preclinical dementia disrupts the homeostatic mechanisms that regulate [slow-wave sleep]" — and, tellingly, carrying the APOE ε4 risk gene predicted faster slow-wave decline. That is what reverse causation looks like: the disease eroding the sleep, rather than the reverse. A 2025 consortium analysis makes the doubt concrete. Pooling five American cohorts with overnight sleep studies — 4,657 people and 998 dementia cases, against 346 people and 52 cases above — it found no association between any measure of sleep architecture and later dementia: not deep sleep, not REM, not slow-wave power.^[36] The two studies are not straightforwardly in conflict, and the newer one says so: it measured how much deep sleep people had on one night, while the older one measured how fast it was declining over years. But two of the earlier paper's senior authors are on the newer one, which is not a result you can attribute to a rival camp.

Why "deep sleep %" is a shaky number

Four separate problems stack on top of each other.

The scoring rule is a convention, not a biological threshold. An epoch counts as N3 when at least 20% of it — six seconds out of thirty — consists of slow-wave activity, which the American Academy of Sleep Medicine (AASM), whose manual defines how sleep is scored, sets at 0.5–2.0 Hz with a minimum amplitude of 75 microvolts peak-to-peak, measured at frontal electrodes.^[37][38] But EEG amplitude falls with age and differs between men and women, so the threshold bites unevenly. Among 2,913 participants in the Sleep Heart Health Study, visually scored deep sleep decreased with age in men but increased in women — a biologically implausible result that survived every amplitude-based rescoring of the same recordings and vanished only when scoring switched to counting slow waves by frequency instead. The authors concluded the sex differences "may be artifactual rather than physiological, and a result of the 75 µV amplitude criterion."^[39] In the same cohort, women averaged 106% more slow-wave sleep than men.^[40] Comparing your percentage against an unstratified "normal range" is close to meaningless — and because scoring rules changed over the decades, "is 15% deep sleep normal?" has no answer that holds across the literature.^[41]

Wearables cannot measure it. Against polysomnography, epoch-by-epoch sensitivity for deep sleep across the six of seven consumer devices that stage sleep ranged from 0.53 to 0.68 — the best missed a third of true deep-sleep epochs, the worst nearly half.^[42] Newer hardware has not fixed it. A 2025 validation of six current wrist devices in 62 adults referred to a hospital sleep clinic — mean age 46 — found errors running in opposite directions by brand.^[43]

(A sixth device, the Withings Scanwatch, was further out still at +73 minutes, but it merges deep and dreaming sleep into one category, so its number is not comparing like with like.)

Neither of those studies tested a ring, which is the device most likely to be on a reader's finger. Rings have since been validated separately, and they stage sleep no better than the watches do.^[44]

The biases run in opposite directions by more than an hour end to end — these are group averages across different subsets of the 62 participants rather than one person wearing two watches, but the spread is the point: which brand you buy changes the answer more than anything you do at night.

And the number is unstable even when it is measured properly. Send the same older adult to a sleep lab on two consecutive nights, have a human technician score both, and the deep-sleep percentage you get back is only moderately reproducible, with the paper rating sleep-macrostructure measures as a class poorly reproducible.^[45] What is reproducible in the same recordings is the underlying continuous signal: slow-wave activity, spindles, K-complexes.

And the devices are trained on the biased ruler. They learn to reproduce visually scored labels — the ones carrying the 75-microvolt problem above — as their manufacturers' own validation papers describe.^[46] So even a well-validated device faithfully reproduces the bias. You would be optimising a noisy estimate of a scoring convention. That is the pattern behind all four problems. The electrical signal is real; the percentage built on top of it by thresholding is the artefact.

Should you try to increase deep sleep?

This is the question most people arrive with, and the honest answer is no — not because deep sleep doesn't matter, but because it is already defended, hard to move safely, and not demonstrably a lever.

Your brain defends it for you. In a randomised crossover study, 35 adults spent a week on either 6 or 10 hours in bed, then ~40 hours awake, then a recovery night.^[47] Across a week of severe restriction or generous extension, slow-wave activity barely moved — REM absorbed the loss — with no habituation over the week; after the deprivation, recovery slow-wave activity rebounded to 140–152% of baseline. The authors conclude the data "imply that there is no significant adaptation to insufficient or excessive sleep." Shortchange your sleep and the brain protects deep sleep at the expense of everything else, then repays itself with interest. Four limits: the protocol used a mild 6–10-hour range in healthy young adults, the rebound is known to be weaker in middle age,^[48] self-correction after a late night says nothing about deficits caused by apnea or alcohol — and, most importantly, a defended deep-sleep number is not the same as a recovered brain. Under sustained short sleep, slow-wave activity settles at close to its normal level while waking performance and alertness keep deteriorating, and stay impaired after slow-wave activity has returned to baseline.^[49] The homeostat protects the measurement, not you.

The things that supposedly raise it mostly don't. The canonical meta-analysis of 66 studies found a single exercise bout has only a small effect on slow-wave sleep, and for regular exercise it reports benefits to total sleep time, efficiency, onset latency and subjective quality — but no slow-wave benefit at all.^[50] A warm bath 1–2 hours before bed genuinely helps, but the evidence is for falling asleep faster, sleeping more efficiently, and rating sleep better — not for N3.^[51]

Devices that "enhance" it change the EEG and nothing else. Closed-loop acoustic stimulation — pink-noise pulses phase-locked to your slow oscillations — does reliably enhance slow oscillations. It does not reliably improve memory. Pooled across the entire literature — 12 studies, 206 people — the average effect on overnight memory was close to nothing, and the range compatible with the data runs from a modest harm to a modest benefit (standardised effect 0.14, 95% CI −0.14 to 0.42).

What makes this more than a null result is the pattern underneath it. The one thing that predicted how large an effect a study found was the year it was published, which accounted for 91.8% of the variation between studies — no other study characteristic explained the decline. The predicted effect fell from a large benefit in 2013 (0.99, 95% CI 0.49–1.49) to a significant effect in the opposite direction by 2021 (−0.39, 95% CI −0.73 to −0.05). Two findings in the same paper explain how that could happen: roughly a third of participants could hear the stimulation, so they were not blinded, and the memory test itself has essentially no test–retest reliability (0.01, 95% CI −0.18 to 0.21) — measure the same person twice and you get unrelated answers.^[52] That is the signature of an effect that was never there, reported by authors one of whom holds a stake in a neurostimulation company — a conflict that runs against the null they found. One replication says it in its title: the stimulation enhances sleep oscillations but not memory.^[53] The one meta-analysis that ever reported a large positive effect in this literature has since been retracted.^[54] And a 2026 five-arm experiment in 102 people, larger than anything in the pooled set, reproduced the same dissociation: the stimulation changed the electroencephalogram and not the memory.^[55]

A drug once did exactly what was asked of it, and the benefit never followed reliably. Gaboxadol was developed on precisely this logic and robustly, dose-dependently increased slow-wave sleep in insomnia patients.^[56] Development was dropped in 2007 after two randomised three-month trials produced only inconsistent patient benefit — in the first, 927 patients, the drug beat placebo on self-reported sleep and on daytime function, but by about twenty minutes of sleep; in the second it did neither.^[57] The company's own summary was "limited or variable efficacy", and the stated reasons for dropping it included psychiatric effects seen at supra-therapeutic doses in a separate abuse-liability study. The hypnogram moved reliably and dose-dependently; what patients reported barely followed — about twenty minutes of self-reported sleep in one trial, and nothing in the other.

And the one thing that reliably works is bad for you. Across doses, ages and sexes, alcohol increases slow-wave sleep in the first half of the night — an effect more robust than its REM suppression.^[58] A nightcap will improve your deep-sleep score while delaying REM onset, cutting total REM, and fragmenting the rest of the night. See alcohol.

If a number goes up when you do something harmful, and doesn't move when you do things that help, it is not a health target.

What to do instead

The levers with real outcome evidence behind them are the unglamorous ones:

1. Get enough total sleep — around 7 hours. This is the variable with mortality data behind it: risk rises in both directions from about seven hours, and about twice as fast on the long side as the short one.^[59] Read that with the same caution this page applies elsewhere — these are self-reported durations in observational cohorts, and illness causing long sleep is the usual explanation offered for the upper arm. Sleep continuity belongs here too: it was six-night sleep efficiency, not any stage measure, that tracked tau in the one experiment that isolated the stages.
2. Keep your timing regular. Regularity outperforms duration as a mortality predictor. See Circadian rhythms.
3. Get screened for sleep apnea if you snore, have witnessed apneas, or have hypertension. Untreated apnea genuinely shreds sleep architecture — and unlike your deep-sleep percentage, it is both measurable and treatable. See Sleep-disordered breathing.
4. If you have 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 REM and the second half of the night, not for your deep-sleep score.

Use a wearable for bedtime consistency and total duration, which it tracks reasonably well, and ignore the stage breakdown. Accuracy in older adults is less well established — the one study in adults aged 56–80 found every device underestimating total sleep time. Be aware, too, that supplement claims resting on tracker "deep sleep scores" are close to uninterpretable given the accuracy data above — see Sleep supplements.

Further reading

• Patel AK, Reddy V, Shumway KR, Araujo JF. Physiology, Sleep Stages. StatPearls 2024.^[60]
• 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.^[61]
• Boulos MI, et al. Normal polysomnography parameters in healthy adults: a systematic review and meta-analysis. Lancet Respiratory Medicine 2019.^[62]
• Ohayon MM, et al. Meta-analysis of quantitative sleep parameters from childhood to old age in healthy individuals: developing normative sleep values across the human lifespan. Sleep 2004.^[63]
• Van Cauter E, Leproult R, Plat L. Age-related changes in slow wave sleep and REM sleep and relationship with growth hormone and cortisol levels in healthy men. JAMA 2000.^[64]
• Himali JJ, et al. Association between slow-wave sleep loss and incident dementia. JAMA Neurology 2023.^[65]
• Leary EB, et al. Association of rapid eye movement sleep with mortality in middle-aged and older adults. JAMA Neurology 2020.^[66]
• Ju YS, et al. Slow wave sleep disruption increases cerebrospinal fluid amyloid-β levels. Brain 2017.^[67]
• Dagum P, et al. The glymphatic system clears amyloid beta and tau from brain to plasma in humans. Nature Communications 2026 — n=39; funded by Applied Cognition, whose device produced the measurements.^[68]
• Chinoy ED, et al. Performance of seven consumer sleep-tracking devices compared with polysomnography. Sleep 2021.^[69]
• Baron KG, et al. Orthosomnia: are some patients taking the quantified self too far? J Clin Sleep Med 2017.^[70]
• 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.^[71]
• 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; the device table above.^[72]
• Lundahl J, Staner L, Staner C, Loft H, Deacon S. Short-term treatment with gaboxadol improves sleep maintenance and enhances slow wave sleep in adult patients with primary insomnia. Psychopharmacology (Berl) 2007 — the drug did exactly what was asked of it.^[73]
• 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.^[74]
• Dijk DJ, Czeisler CA. Contribution of the circadian pacemaker and the sleep homeostat to sleep propensity, sleep structure, electroencephalographic slow waves, and sleep spindle activity in humans. J Neurosci 1995 — the forced-desynchrony experiment.^[75]
• Kredlow MA, et al. The effects of physical activity on sleep: a meta-analytic review. J Behav Med 2015 — 66 studies; no slow-wave benefit from regular training.^[76]
• Haghayegh S, et al. Before-bedtime passive body heating by warm shower or bath to improve sleep: A systematic review and meta-analysis. Sleep Med Rev 2019.^[77]
• Ebrahim IO, et al. Alcohol and sleep I: effects on normal sleep. Alcohol Clin Exp Res 2013.^[78]
• Harlow TJ, Jané MB, Read HL, Chrobak JJ. Memory retention following acoustic stimulation in slow-wave sleep: a meta-analytic review of replicability and measurement quality. Front Sleep 2023 — 12 studies, 206 subjects; publication year explains 91.8% of the heterogeneity.^[79]
• Baandrup L, et al. Pharmacological interventions to enhance slow wave sleep and cognitive performance: a systematic review. Sleep Med Rev 2026 — 27 randomised trials.^[80]
• 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.^[81]
• Ganglberger W, et al. Brain Health from Sleep EEG: A Multicohort, Deep Learning Biomarker for Cognition, Disease, and Mortality. NEJM AI 2026 — 36,000 recordings.^[82]
• Miao A, et al. Brain clearance is reduced during sleep and anesthesia. Nat Neurosci 2024.^[83]
• Hauglund NL, et al. Norepinephrine-mediated slow vasomotion drives glymphatic clearance during sleep. Cell 2025.^[84]
• Mullins AE, et al. The stability of slow-wave sleep and EEG oscillations across two consecutive nights of laboratory polysomnography in cognitively normal older adults. J Sleep Res 2025.^[85]
• Dijk DJ. Human data at odds and in confirmation of the two-process model of sleep regulation — a perspective. NPJ Biol Timing Sleep 2026.^[86]
• Yin J, et al. Relationship of Sleep Duration With All-Cause Mortality and Cardiovascular Events: A Systematic Review and Dose-Response Meta-Analysis of Prospective Cohort Studies. J Am Heart Assoc 2017.^[87]
• Henin S, et al. Closed-loop acoustic stimulation enhances sleep oscillations but not memory performance. eNeuro 2019.^[88]

---