Sleep architecture, not sleep hours: what a shortened night actually costs

Applied Physiology #3 - Two hours removed from the end of a night do not take a slice of each sleep stage. They take one stage almost entirely, and it is the stage nobody notices losing.

Third in the applied physiology series. The opening piece treated recovery capacity as an account you learn to read, the second identified the demand on that account paying the highest return, and this one asks where the account gets refilled. A fourth will ask which measurements are worth taking at all.

Sleep advice arrives almost entirely as a quantity. Get your eight hours, hit your sleep score. Quantity is the easiest thing to say and the easiest thing to sell a device against, and it turns out to describe rather poorly what a night did for you.

A night runs as an ordered sequence. Deep sleep loads at the beginning, rapid eye movement sleep accumulates toward morning, and different control systems govern the two. Shorten the sequence and you do not lose a proportional share of each part, because the body protects the beginning and pays for it out of the end. That reorganises most of the practical advice, and it explains why two people sleeping identical totals can recover very differently.

This is an educational and strategic perspective, not personal medical advice.
The views are the author's own and not statements by Atlas Cove Lda.

How a night is put together

Two processes set sleep timing. One accumulates: the longer wakefulness continues, the greater the drive to sleep, and sleeping discharges that drive. The other oscillates on an internal cycle of about twenty-four hours and pays no attention to how long you have been up. Borbély and colleagues reappraised the model built on this pairing, Process S and Process C, thirty-four years after the original, and it remains the framework that best reproduces both the timing of human sleep and its depth (Borbély et al., 2016). The accumulating side has a well-studied chemical correlate in adenosine, a by-product of cellular energy use that gathers through waking hours. Caffeine binds those receptors without activating them, which matters further down this page.

Because two systems are involved, clock time becomes a variable in its own right. An identical block of sleep taken at a different hour is not an identical block of sleep, since the circadian process sits at a different point in its cycle and permits or suppresses different things while you are under. A red-eye and a normal night at home are not interchangeable, whatever the stopwatch says.

Why deep sleep comes first and REM comes last

Separating the two experimentally takes an unusual design. In forced desynchrony, participants live on a day length their internal clock cannot lock onto, which scatters sleep across every phase of the circadian cycle and lets the homeostatic and circadian contributions be read apart.

Dijk and Czeisler applied this to eight men living on twenty-eight hour days for around a month without clocks or daylight. The results separated cleanly. REM showed a pronounced circadian rhythm whose peak arrived shortly after core body temperature bottomed out. Slow wave activity behaved otherwise, declining steadily inside each sleep episode and carrying only a modest circadian component, one that failed to follow the rhythm of sleep propensity (Dijk and Czeisler, 1995).

A further result in that paper does the practical work. REM was also sensitive to elapsed time asleep, and its highest values anywhere in the protocol occurred where a scheduled episode finished at the participant's habitual waking hour. The window in which REM is most available coincides precisely with the window an early alarm deletes. No amount of going to bed earlier changes where in the night REM prefers to sit.

What a short night actually removes

If sleep were uniform, removing a fifth of it would cost a fifth of every component. It does not work that way.

Brunner, Dijk and Borbély gave nine healthy young participants three baseline nights of eight hours, then four consecutive nights restricted to four hours, holding the bedtime constant and moving the waking time earlier. That protocol removes the second half of the night, which is exactly what a demanding stretch of work does to most people. Slow wave sleep survived the restriction largely unaffected. Stage one, stage two and REM absorbed the loss (Brunner et al., 1993).

So the system triages under pressure, and it triages in a consistent direction: defend the front, charge the shortfall to the back. Anyone whose sleep is compressed at the waking end is therefore running a REM deficit specifically, whatever their weekly average suggests.

The stage that has mostly gone by midlife

Deep sleep gets protected because a considerable amount of maintenance is scheduled into it: the largest growth hormone pulse of the twenty-four hour cycle, peak tissue repair activity, and the daily maximum in parasympathetic tone, which connects this directly to the autonomic account this series opened with.

Less widely known is the timetable of its decline. Van Cauter, Leproult and Plat assembled recordings from 149 healthy men spanning ages sixteen to eighty-three. In the youngest group, aged sixteen to twenty-five, deep slow wave sleep took up 18.9 percent of total night-time sleep. For the thirty-six to fifty year olds the same figure was 3.4 percent. Lighter stages had absorbed what went missing, and REM had yet to fall significantly at that stage. Growth hormone secretion dropped in parallel across that span and held a significant association with slow wave sleep that was independent of age (Van Cauter et al., 2000).

Read together with the previous section, this reorders the priorities most people are working from. For a reader in their forties or fifties, the majority of the deep sleep decline has already happened, while REM remains substantially intact and is the part being surrendered every weekday morning. Protecting deep sleep is still worth doing. It is simply no longer the part with the most left to lose.

What REM appears to be doing

REM leaves no obvious trace when it goes missing, which is much of why its loss is tolerated for years at a stretch.

Wagner, Gais and Born approached its function through the shape of the night itself. Participants learned emotional and neutral material, then slept across either an early window dominated by slow wave sleep or a late window dominated by REM. Retention of the emotional material was better across the late window. The authors treated that as evidence compatible with REM supporting the formation of emotional memory (Wagner et al., 2001).

Two things are worth keeping straight. The finding concerns emotional processing, and it compares windows rather than demonstrating that REM deprivation causes measurable harm. The popular account meanwhile credits REM with a good deal more, motor skill consolidation in particular, where the evidence points more consistently toward lighter non-REM stages and spindle activity. Kept at its documented size, the conclusion is that the back half of a night performs work the front half does not, and losing it removes something specific rather than trimming everything a little.

The two halves of the night at a glance

The front of the night, dominated by slow wave sleep

• Governed by the homeostatic process, so it depends mainly on how long you have been awake.

• Associated with the day's largest growth hormone pulse, peak tissue repair activity and maximum parasympathetic tone.

• Under sleep restriction it is largely preserved, because the body defends it first.

• Across the lifespan it collapses early, 18.9 percent in the late teens and early twenties against 3.4 percent by the late thirties and forties.

• Weakest point in the evidence is that the lifespan figures come from men only, compared across age bands rather than followed over time.

The back of the night, dominated by REM

• Governed by the circadian process, peaking shortly after the core temperature minimum, so its position is set by clock time rather than by tiredness.

• Associated with emotional memory processing on the available evidence, and considerably less firmly with motor learning than the popular account suggests.

• Under sleep restriction it takes the loss, along with the lighter stages.

• Across the lifespan it holds up into midlife and declines later.

• Weakest point in the evidence is that its function is inferred from split-night memory experiments in small samples rather than from long-term deprivation studies, which would be difficult to run ethically.

Getting to sleep is a heat-loss problem

Standard guidance says to keep the bedroom cool. The mechanism underneath is more specific and more actionable than the guidance built on it.

Kräuchi and colleagues tested a range of physiological candidates for predicting how quickly people fell asleep under controlled laboratory conditions. Top of the list came the gradient between distal and proximal skin temperature, which captures how much warmer the hands and feet run than the trunk. It outperformed the core reading itself, the speed at which that core reading was moving, heart rate, the onset of melatonin, and what participants reported about their own sleepiness. Their reading was that selective widening of the blood vessels in distal skin, and the heat loss this produces, is what promotes rapid sleep onset (Kräuchi et al., 2000).

The practical translation is counterintuitive and frequently got backwards. Sleep onset is helped by losing heat through the extremities, so warm hands and feet are working with the mechanism while a cold room gives the shed heat somewhere to go. Chilled feet and a chilled bedroom are pulling in opposite directions. Socks with a window open is a coherent combination rather than a contradictory one.

Alcohol and caffeine reshape a night without shortening it

Both change what a night contains while leaving its length untouched, which is why neither appears in a duration-based view of sleep.

For alcohol, a 2025 systematic review and meta-analysis pooled twenty-seven studies. Shortened sleep onset, the effect a nightcap is generally taken for, appeared only at high doses of around five standard drinks, and effects on total sleep time, sleep efficiency and wake after sleep onset could not be established with confidence. What emerged clearly was a dose-response relationship affecting REM: onset delayed and duration reduced from roughly two standard drinks, deteriorating as intake rose (Gardiner et al., 2025). That is not an argument for abstinence. It is an argument about what is being purchased. The sedative effect most people believe they are buying is not well evidenced at ordinary doses, while the cost falls on the stage already most exposed to an early alarm.

Caffeine acts on the other control system. By occupying adenosine receptors without triggering them, it suppresses the perception of sleep pressure while that pressure builds underneath, and with an elimination half-life of roughly five to six hours in most adults, a mid-afternoon dose remains materially present at bedtime. In a randomised placebo-controlled crossover, Drake and colleagues gave 400 mg of caffeine, roughly what two strong coffees deliver, either immediately before the usual bedtime or three or six hours ahead of it. All three timings cost over an hour of sleep. With six hours of separation, deep sleep came in near forty-nine minutes, against something closer to seventy-one under placebo (Drake et al., 2013).

The instructive part concerns perception. At six hours out, the objective measure registered the disruption while the participants' own diaries did not. The effect was undetectable from the inside rather than absent, which reasonably explains why advice about afternoon coffee is so widely and so sincerely ignored.

Can you catch up at the weekend?

Partly, and less completely than the arithmetic of a weekly total implies.

The restriction study above also tracked what happened afterwards. Across the first two recovery nights, total sleep time and REM sleep both increased above baseline and sleep latency shortened, which is a genuine rebound and indicates the system actively seeks to repay what it lost (Brunner et al., 1993). So recovery sleep is real, and a repaid night is better than an unrepaid one.

The complication is that weekend recovery is usually purchased by moving the waking time, which is the variable the final section of this article identifies as carrying the strongest outcome evidence. A Saturday lie-in repays some sleep debt and simultaneously delivers a large timing shift to the circadian system, reversed again on Sunday night. Treating the week as a total to be balanced is precisely the habit that damages the input that predicts most. The honest position is that catching up helps with the debt and works against the regularity, and that nobody has run the trial that weighs one against the other directly.

What a sleep tracker can and cannot see

Most readers will test this against a wrist-worn device, so it is worth being precise about which of its numbers deserve weight.

Chinoy and colleagues ran seven consumer trackers head to head with laboratory polysomnography over three consecutive nights, using thirty-four healthy young adults, and disrupted one of those nights on purpose. Detection of sleep was strong throughout, epoch-by-epoch sensitivity of 0.93 or higher on every device, and most matched or bettered research-grade actigraphy at identifying wake. Specificity was far weaker, between 0.18 and 0.54. Stage classification was inconsistent. On average most devices missed somewhere between a third and a half of deep sleep, and the same again for REM, with the mislabelled epochs usually reassigned to light sleep, and accuracy fell further on the nights when sleep was worse (Chinoy et al., 2021).

Two consequences follow. A single night's stage breakdown is an estimate rather than a measurement, and it is least reliable on exactly the disturbed nights that prompt someone to open the app. Averaging across a week is not a clean escape either, since a weekly number landing near reality may have got there because errors in opposite directions happened to offset, rather than because any single night was read correctly.

What these devices do record accurately is the time you went to bed and the time you got up. That is a timing measurement rather than a staging estimate, and as the next section shows, timing is where the strongest outcome evidence sits. The most useful number on the screen is generally the one nobody looks at.

Regularity outpredicted duration

Windred and colleagues derived a sleep regularity index from over ten million hours of accelerometer recordings across 60,977 UK Biobank participants with a mean age of sixty-three, then followed mortality for an average of just over six years. Set beside the least regular quintile, the remaining four carried all-cause mortality risk lower by somewhere between twenty and forty-eight percent. When regularity and duration were entered into equivalent models, regularity was the stronger predictor, its minimum hazard ratio reaching 0.52 against 0.69 for duration in minimally adjusted models, and 0.70 against 0.76 in fully adjusted ones (Windred et al., 2024).

The strength of that result should be stated carefully. This is observational data, and the authors state plainly that regularity may be a cause of premature mortality risk or a marker of it. Nobody has demonstrated that imposing a regular schedule extends life, and somebody whose sleep timing has become erratic because they are becoming unwell would generate this same association. Read at its proper weight it still leaves something unusual: the best-evidenced sleep input available costs nothing, requires no equipment, and is almost always the first thing surrendered to a calendar.

The trade-off deserves stating rather than hiding. Keeping the hour you rise within sixty minutes of itself, every day of the week, gets paid for in recurring and specific ways: the dinner that overruns, the concert on a Tuesday, the flight booked for its fare, the shapeless Sunday morning. Anyone presenting this as costless has not attempted it in a country where dinner starts at ten.

How to check your own, in two weeks

Two checks are worth more here than any single measurement, and neither requires buying anything.

1. The weekday to weekend gap. Write down the hour you genuinely leave the bed, daily, for a fortnight. Average the five working days, average the two at the weekend, and read the distance between them. Anything past about sixty minutes means your circadian system is handed one instruction each Friday and its reverse each Sunday. Closing it is the highest-value change most people have available, and it costs nothing.

2. The six hour caffeine line. Note the time of your last caffeinated drink and add six hours. If the result falls after your intended lights-out, the evidence above says the night is being altered whether or not you can feel it, and this is changeable tonight rather than eventually.

Neither check requires a device, and both answer a decision. That test, whether a measurement will actually change what you do, is the one we apply to every number in the protocol we build.

Where the evidence is weak

A reference page should be explicit about the limits of what it has just argued, so here is where this case is thinnest.

The lifespan data is male and cross-sectional. All 149 participants in Van Cauter's dataset were men, and the age comparisons are between groups rather than within individuals followed over time. Nothing in it licenses a claim about women, and it cannot rule out cohort effects.

The architecture experiments are small. Eight men in the forced desynchrony protocol, nine in the restriction study. These designs are demanding to run, which explains the sample sizes, and it also means the numbers should be treated as descriptions of a mechanism rather than as population estimates. The restriction study's own conclusion says slow wave sleep was "largely" unaffected, and that qualifier is doing real work.

The caffeine trial is small and industry-funded. Twelve participants, sleep measured with a single-channel home EEG headband rather than laboratory polysomnography, and the study was funded by an investigator-initiated grant from the device manufacturer. The direction of the finding is consistent with the pharmacology and the effect size is large, which is why it is included, but a larger independent replication would strengthen it considerably.

The wearable data comes from healthy young adults. Mean age around twenty-eight, no sleep disorders, tested in a laboratory. A reader in their fifties reading their own stage chart is extrapolating twice over, and there is no strong reason to expect device accuracy to improve in an older population with more fragmented sleep.

The mortality evidence is associative. Discussed above and worth repeating here, because it is the finding most likely to be over-quoted once it leaves this page.

What would change the conclusion. A randomised trial that fixed waking time in one group and total sleep duration in another, followed for long enough to read hard outcomes, would settle whether regularity is causal. It has not been run, it would be expensive and hard to keep people compliant, and until it exists the honest summary is that regularity is the best-supported bet rather than a proven lever.

Where Atlas Cove fits

An Atlas Cove week treats sleep timing as a load variable sitting alongside training and autonomic data, rather than as background conditions noted once and forgotten. A variable that determines what a night returns belongs in the same column as the work that spends it. The same logic governs the argument for strength work in a crowded week: the question is always which input returns the most per hour, given the state the person is actually in.

People finish the week able to name the hour of their own night that most repays protecting, and to say why that answer differs between two people running identical schedules. What holds after the six days decides whether any of it was worth doing.

Eight hours describes an opportunity and says nothing about what got taken up inside it. Whether a given night delivered the repair work, the emotional processing, or the steadiness of timing that beats both as a predictor is a separate question entirely, and answering it well rarely calls for more time in bed. It calls for the same time, held.

This is an educational and strategic perspective, not personal medical advice.
The views are the author's own and not statements by Atlas Cove Lda.

Questions people actually ask

Is deep sleep or REM sleep more important?

They do different jobs, so the useful question is which one you are currently short of. Deep sleep carries the growth hormone pulse and the repair work, and it loads into the first part of the night. REM carries emotional processing and accumulates toward morning. If your nights are being cut at the waking end, you are short of REM specifically, whatever your total says. If you are in your forties or beyond, most of the age-related deep sleep decline has already occurred, which makes REM the part with more still to protect.

How many hours of sleep do I actually need?

The evidence covered here does not answer that in hours, which is part of the point. What it does show is that when sleep is compressed, the loss falls on particular stages rather than spreading evenly, and that in a cohort of nearly 61,000 adults the consistency of sleep timing predicted mortality more strongly than duration did. A stable schedule at a slightly shorter duration is a better-supported target than a variable schedule averaging more.

Does catching up on sleep at the weekend work?

Partly. Recovery nights after restriction show increased total sleep and increased REM, so the rebound is real. The difficulty is that most weekend catch-up happens by shifting the waking time by a couple of hours, which is a direct hit to the regularity that carries the stronger outcome evidence. Repaying the debt and protecting the schedule pull against each other, and no trial has weighed them directly.

Are sleep trackers accurate?

For distinguishing sleep from wake, yes, and comparably to research-grade actigraphy. For sleep stages, no: tested against laboratory polysomnography, most consumer devices miss between thirty and fifty percent of deep and REM sleep, and they perform worst on disrupted nights. Treat the stage breakdown as an estimate and the bedtime and waking timestamps as reliable, then steer by the timestamps.

Does alcohol help you sleep?

It helps you fall asleep only at high doses, around five standard drinks in the pooled evidence, and it disrupts REM sleep from roughly two drinks with the effect worsening as intake rises. The sedation people notice is real; the restorative sleep they assume follows it is not well supported.

Sources

1. Borbély, A. A., Daan, S., Wirz-Justice, A., & Deboer, T. (2016). The two-process model of sleep regulation: a reappraisal. Journal of Sleep Research, 25(2), 131-143. DOI: 10.1111/jsr.12371

2. Dijk, D. J., & Czeisler, C. A. (1995). Contribution of the circadian pacemaker and the sleep homeostat to sleep propensity, sleep structure, electroencephalographic slow waves, and sleep spindle activity in humans. The Journal of Neuroscience, 15(5 Pt 1), 3526-3538. DOI: 10.1523/JNEUROSCI.15-05-03526.1995

3. Brunner, D. P., Dijk, D. J., & Borbély, A. A. (1993). Repeated partial sleep deprivation progressively changes the EEG during sleep and wakefulness. Sleep, 16(2), 100-113. DOI: 10.1093/sleep/16.2.100

4. Van Cauter, E., Leproult, R., & Plat, L. (2000). Age-related changes in slow wave sleep and REM sleep and relationship with growth hormone and cortisol levels in healthy men. JAMA, 284(7), 861-868. DOI: 10.1001/jama.284.7.861

5. Wagner, U., Gais, S., & Born, J. (2001). Emotional memory formation is enhanced across sleep intervals with high amounts of rapid eye movement sleep. Learning & Memory, 8(2), 112-119. DOI: 10.1101/lm.36801

6. Kräuchi, K., Cajochen, C., Werth, E., & Wirz-Justice, A. (2000). Functional link between distal vasodilation and sleep-onset latency? American Journal of Physiology-Regulatory, Integrative and Comparative Physiology, 278(3), R741-R748. DOI: 10.1152/ajpregu.2000.278.3.R741

7. Gardiner, C., Weakley, J., Burke, L. M., Roach, G. D., Sargent, C., Maniar, N., Huynh, M., Miller, D. J., Townshend, A., & Halson, S. L. (2025). The effect of alcohol on subsequent sleep in healthy adults: a systematic review and meta-analysis. Sleep Medicine Reviews, 80, 102030. DOI: 10.1016/j.smrv.2024.102030

8. Drake, C., Roehrs, T., Shambroom, J., & Roth, T. (2013). Caffeine effects on sleep taken 0, 3, or 6 hours before going to bed. Journal of Clinical Sleep Medicine, 9(11), 1195-1200. DOI: 10.5664/jcsm.3170

9. Chinoy, E. D., Cuellar, J. A., Huwa, K. E., Jameson, J. T., Watson, C. H., Bessman, S. C., Hirsch, D. A., Cooper, A. D., Drummond, S. P. A., & Markwald, R. R. (2021). Performance of seven consumer sleep-tracking devices compared with polysomnography. Sleep, 44(5), zsaa291. DOI: 10.1093/sleep/zsaa291

10. Windred, D. P., Burns, A. C., Lane, J. M., Saxena, R., Rutter, M. K., Cain, S. W., & Phillips, A. J. K. (2024). Sleep regularity is a stronger predictor of mortality risk than sleep duration: a prospective cohort study. Sleep, 47(1), zsad253. DOI: 10.1093/sleep/zsad253