The direction of the arrow here is unusually well pinned. A polygenic score compresses the combined influence of a person's genes on their probability of a disease into a single figure [16], and those variants were settled long before anyone recorded a night of sleep [23]. Fragmented sleep in the older group cannot have produced the genetic risk sitting beside it [2]. That rules out the reverse-causation worry that shadows most sleep-and-dementia work. But the design cannot say whether repairing sleep would move risk, which is the more useful claim, and nothing here answers it.
Puneet Talwar of the GIGA ULiege laboratory reads the mechanism the other way, arguing that the awakenings are "not insignificant" and that certain profiles "could promote the accumulation of proteins involved in Alzheimer's disease" [8]. That is a causal proposition, and this study cannot test it: what the release describes is genetic risk estimates compared against sleep characteristics, with no follow-up of who is later diagnosed [20].
The age contrast is the part worth arguing about. It survives the obvious objection. The better-powered comparison is the one that came up empty [22], so the absent link in young adults is not easily written off as too few participants. Worth noting that the older band covers 19 years against 13 for the younger one [24], which means "middle age" here spans people at rather different points.
The brainstem thread sits oddly against that null. The locus coeruleus is about the size of a grain of rice and helps regulate wakefulness, attention and sleep [9]; Gilles Vandewalle calls it "difficult to observe" but says it "appears to play a role in the early mechanisms linked to the disease" [10], and it is described as among the earliest regions where abnormal protein shows up [13]. The same 2025 imaging work tied healthy locus coeruleus function to REM sleep quality [12]. So sleep and brainstem condition appear to travel together in young people, while sleep and inherited risk do not [2]. Whether the genetic signal needs decades to become visible in a sleep recording, or the two analyses are indexing different things, the release does not settle.
What is missing here is any sense of size. The release gives no effect estimate, no per-group counts, and no account of how sleep was measured or over how many nights [19]. Genetics are only a partial contributor to the disease in the first place [17], which caps how much a genotype correlation can carry on its own.
My read: the honest label is a candidate stratification variable for research cohorts [18]. It has no business being offered as a test to anyone yet, and the team's own caution about scores that stayed low and cannot speak to any individual says roughly the same thing [6].
