Five cohorts, 44 sleep measurements, no association with dementia risk

"Sleep is good for your brain" is among the least controversial statements in health communication. A study published in Neurology assembled the best available data to test the specific version of that claim — which measurable features of sleep predict dementia — and returned a null result [s1].

Null results in a well-powered study with good measurement are worth more attention than they usually get.

What the consortium had to work with

The Sleep and Dementia Consortium pools five US population-based cohorts in which participants underwent methodologically consistent, home-based polysomnography — a full overnight sleep study in their own beds — alongside self-reported habitual sleep, neuropsychological assessment and dementia surveillance [s1]. All participants were free of dementia and stroke at baseline [s1].

The pooled cognitive analysis included 5,958 participants with a mean age of 70 (SD 11), 32% female [s1]. The incident dementia analysis included 5,471 participants with a mean age of 74 (SD 12), 31% female, among whom 1,134 developed dementia over a median follow-up ranging from 5 to 19 years across cohorts [s1].

Polysomnography produces far more variables than any analysis can sensibly test. The authors addressed that with a data-driven step: a cluster-around-latent-variables analysis reduced 44 sleep metrics to nine composites, which together explained 49% of the total variance in the original measures [s1].

The nine were sleep quantity and efficiency; sleep fragmentation; light non-REM predominance; N3 predominance; spindle number and duration; REM sleep bouts; respiratory disturbances; slow oscillation–spindle coupling; and spindle amplitude [s1].

That list covers essentially every mechanism proposed to link sleep to neurodegeneration, including the two that dominate current research interest: deep-sleep architecture, and the coupling between slow oscillations and sleep spindles thought to underpin overnight memory consolidation.

The results

For cognitive performance, two composites reached significance. Greater sleep quantity and efficiency — longer, more consolidated sleep — was associated with better global cognition, with a pooled β per one-unit change in the composite of 0.03 (95% CI 0.004 to 0.06; p = 0.033) [s1]. Stronger slow oscillation–spindle coupling was associated with better global cognition, pooled β 0.04 (95% CI 0.003 to 0.07; p = 0.039) [s1].

For incident dementia, none of the nine composites was significantly associated with risk [s1].

The authors state it directly: sleep composites were not significantly associated with dementia risk, while longer, more consolidated sleep and stronger slow oscillation–spindle coupling were associated with cognitive performance [s1].

How to read a null of this kind

Three readings are available and they are not equivalent.

The first is that these sleep features genuinely do not predict dementia. Possible, and this is the best-designed test of it so far.

The second is measurement timing. Polysomnography captures one or two nights, often decades before a diagnosis. A single-night snapshot of a highly variable behaviour may simply not carry enough signal to survive years of follow-up, however well it is measured.

The third is that the cognitive findings and the dementia null are consistent with a small, continuous relationship between sleep and cognitive function that never becomes large enough to shift the probability of crossing a diagnostic threshold. The effect sizes on cognition support that reading: pooled βs of 0.03 and 0.04 are small [s1].

What the study does undercut is the strong version of the claim — that fixing measurable sleep architecture is a lever on dementia risk. That version now has a well-powered test against it.

A cohort that found the opposite, on a different variable

A study published days earlier in the Journal of Sleep Research looked not at sleep architecture but at change in self-reported sleep duration over time [s2].

It followed 3,057 participants from the Swedish population study 'Good Aging in Skåne', mean age 68.4, all of whom had normal sleep duration of 6 to 8 hours at baseline, for up to 20 years (mean 8.39 years) [s2]. Change was defined as shortened or prolonged if a participant lost or gained more than one hour since their previous visit, and sleep variability as the standard deviation of sleep duration across visits [s2]. Models adjusted for health-related factors including time spent in bed without sleeping [s2].

Longitudinal patterns of prolonged sleep — an increase of more than an hour — were associated with significant cognitive decline in fully adjusted models, and with a 131% increased risk of dementia compared with stable sleep (HR 2.31, 95% CI 1.77 to 3.03; p < 0.001) [s2]. Greater sleep variability was also associated with steeper cognitive decline and with increased dementia risk (HR 1.49; p < 0.001) [s2].

Shortened sleep was not significantly associated with cognitive decline [s2].

The two findings do not conflict

They measure different things. The consortium measured what sleep looked like on a given night [s1]. The Swedish cohort measured how a person's sleep changed across years [s2].

Lengthening sleep by more than an hour in later life is a plausible early consequence of the neurodegenerative process rather than a cause of it — which is precisely why a hazard ratio of 2.31 should not be read as a warning against sleeping longer. Adjusting for time in bed without sleeping [s2] addresses one alternative explanation, but not this one.

Read together, the pair suggests that the useful sleep signal for dementia may be trajectory rather than architecture — and that trajectory, if it is a signal at all, is more likely a symptom than a target.

This article is informational and is not medical advice.