Sleep fragmentation vs sleep duration: which one correlates with what

Three cohorts, thousands of adults, wearable-measured sleep. The broken-up part of the night keeps showing up in frailty, mortality and cognition; the total hours often do not. All of it is correlational, and the article says exactly where that limit sits.

mixedStudies disagree, or the mechanism is contested.Revised 04 Oct 2026

Five small clay beads resting in a row on a pale wooden surface with uneven gaps

Same number of hours, very different night. Duration is one number; fragmentation is a shape.

The question people actually search for at three in the morning is not “how many hours should I sleep”. It is “why is my night in pieces”. Those are different questions, and only one of them has been measured properly in large cohorts. The honest headline from the objective sleep literature is this: the broken-up part of the night is what keeps showing up in the outcomes, and the total hours often are not. Every result below is correlational. Nothing in this article can tell you what will happen to you if you sleep better, only that the two sleep features do not travel together, and only one of them keeps appearing in the risk data.

What the two measures actually are

Total sleep time is the sum: the hours between falling asleep and the final awakening, minus the time awake in between. It is the number on the watch, and roughly the number people quote from memory.

A fragmentation index is a shape rather than a quantity. On actigraphy it is usually built from the percentage of the recording spent moving plus the share of immobile periods lasting a minute or less — in other words, how many times the body shifted and how many micro-awakenings punctuated the night. The two can diverge completely. Someone who is awake twenty minutes every hour and then falls back asleep has lost hours; someone who lies still and awake for ninety minutes without moving scores almost no fragmentation. The first night is a bad night. The second one is a genuinely different physiological state, and the instruments treat them differently.

The older-adult cohort: fragmentation and death, in the same participants

The National Social Life, Health and Aging Project followed community-dwelling older adults born between 1920 and 1947. A random third of one wave was invited into a sleep sub-study; 727 consented and 615 were analysed, wearing a wrist actigraph for 72 hours, an average of 2.93 nights (Sleep, PMID 33406254). Fragmentation, total sleep time, percent sleep and wake after sleep onset were all computed.

The reported results are the interesting part. Per ten-unit increase in the fragmentation index, the odds of being frailer at follow-up were 1.70 times (95% CI 1.02 to 2.84) and the odds of dying over five years were 2.12 times (95% CI 1.09 to 4.09). Greater wake after sleep onset carried an odds ratio of 1.24 (95% CI 1.02 to 1.50) for mortality, and lower percent sleep 0.41 (95% CI 0.17 to 0.97). Total sleep time was among the actigraphic parameters collected and is not among the associations the paper reports. Note the interval widths: a lower bound of 1.02 is barely distinguishable from nothing. These are real signals with soft edges, not established risks.

The midlife cohort: duration showed nothing at all

The Coronary Artery Risk Development in Young Adults study measured sleep by actigraphy in 526 people in their mid-thirties to late forties, then tested their cognition eleven years later (Neurology, PMID 38170947). Mean sleep duration was 6.1 ± 1.1 hours and the mean fragmentation index 19.2 ± 8.1%. After adjusting for smoking, body mass index, depression, physical activity, hypertension and diabetes, the highest versus lowest tertile of fragmentation carried odds of poor cognitive performance of 2.97 on the digit-symbol test (95% CI 1.34 to 6.56), 2.42 on fluency (95% CI 1.17 to 5.02) and 2.29 on the Montreal Cognitive Assessment (95% CI 1.06 to 4.94). The association did not differ by sex or race.

The line in the paper that should be quoted more often is the next one: objective sleep duration and subjective sleep quality were not associated with cognition in midlife. Same people, same nights, same cohort. The broken pattern carried the signal; the hours did not.

The Swedish men: two profiles, one elevated risk

Twenty-four-hour actigraphy in 2,667 men aged 65 and over, followed for twelve years, was sorted by machine learning into three sleep-circadian profiles: actively healthy sleepers (64.0%), fragmented poor sleepers (14.1%) and long frequent nappers (21.9%). Against the healthy group, fragmented poor sleepers had a hazard ratio of 1.35 for dementia (95% CI 1.02 to 1.78) and 1.32 for cardiovascular events (95% CI 1.08 to 1.60), after multivariable adjustment (Communications Medicine, PMID 40696147). The long nappers showed no association with dementia, 1.09 (95% CI 0.86 to 1.38), and a cardiovascular ratio of 1.16 (95% CI 0.98 to 1.37) that does not clear the conventional threshold either.

Note what the profile split does: it separates people by the structure of their sleep rather than its length, and the group that stands out is the fragmented one.

Why the correlational label is not a formality

It is tempting to read three cohorts pointing the same way as a settled cause. It is not, and the gap is specific rather than vague.

There is no experiment in which researchers assigned people to have broken sleep. The closest evidence runs the other way. A meta-analysis of 41 randomized controlled trials of sleep restriction — deliberately cutting sleep, not fragmenting it — found increased hunger of 13.4 points on a 100-point scale, an extra 252.8 kcal per day of intake, a weight gain of 0.34 kg, and reduced insulin sensitivity at a standardized mean difference of -0.70 (Sleep Medicine Reviews, PMID 30870662). That is a graded body of evidence, and it applies to a different intervention.

For fragmentation specifically, the mechanism question is explicitly open. A review of laboratory work on sleep and metabolism states that there is limited evidence indicating that sleep fragmentation without reduction in sleep duration also reduces insulin sensitivity (Metabolism, PMID 29510179). In other words, the one clean experimental question that would separate the two mechanisms has not been answered.

There is also reverse causation, in both directions. Fragmented sleep can be an early marker of illness, of pain, of nocturia, of mood problems, of medication timing. And people who sleep badly report differently from how they sleep, which is why every study above used an accelerometer instead of a questionnaire.

What this does and does not settle

It settles a framing question. If you are choosing which sleep feature to pay attention to, the literature points at the structure of the night, not just its length. A shorter night that is consolidated and a longer one chopped into pieces are not the same object, and treating the hours as the whole story misses the measure that keeps associating with outcomes.
It does not settle causation. Every number here is an odds or hazard ratio from an observational cohort, with confidence intervals that in several cases touch 1.0.
It does not settle the mechanism. The proposed pathways — glucose handling, inflammation, blood pressure — are plausible and largely untested directly under fragmentation-without-shortening conditions.
It says nothing about individual prognosis. A ratio from a cohort is a population statement, and it does not transfer to a particular person’s night.
It says nothing about compounds. The cohorts above measure a night; they do not test a molecule. That literature exists, and it is short — what the research on DSIP actually measured is the version of it that keeps its own limits in view.
The actigraphy era is young. Most of this work is a decade old or less, uses device definitions that differ between cohorts, and has not yet been replicated across independent samples at this scale.