Polysomnography is the gold standard for measuring sleep, and it is also, for most families, a single awkward night: wires glued to the scalp, a strange bed, a technician down the hall, and a result that describes one night out of a thousand. Dementia does not arrive on one night. It arrives across seasons. That gap between what the best instrument can measure and what the disease actually does is the problem a paper published online on August 10, 2026 in Experimental Neurobiology sets out to work around. Sang-ho Lee, Cheil Moon and colleagues did not look for a new risk factor. They looked for a new way to read data that millions of people are already generating on their own wrists.
The method is the news. Most studies of wearable sleep data collapse each night into a handful of averages: total sleep time, minutes of deep sleep, number of awakenings. Lee and colleagues threw the averages out and cut the night into 10-minute slices, then tracked what happened inside each slice over time, so that the shape of a night, and not just its totals, became the unit of analysis. They applied it to daily sleep data from thirteen participants (5 healthy controls, 4 with amnestic mild cognitive impairment, and 4 with mild Alzheimer's disease) monitored for three months with a Fitbit Charge 2.
What came out were what the authors call group-dependent temporal signatures. The aMCI and Alzheimer's groups showed shorter deep sleep in the early part of the night, reduced REM sleep roughly three hours after sleep onset, and consistently elevated light sleep and wake after sleep onset across the night. Read as a picture rather than a number, that is a night that never quite sinks and never quite settles. None of it is a diagnosis, and the paper does not present it as one; the authors describe their output as candidate digital phenotypes, meaning features worth testing, not findings worth acting on.
The honesty is written into the abstract itself. The authors state that "All reported findings are exploratory in nature and require replication in larger, independent cohorts." Thirteen people is a feasibility sample, not a cohort. Four people per clinical group means a single unusual sleeper can bend a curve. And the study design, which compares groups already clinically classified, cannot say whether these patterns appear before a diagnosis does, which is the only question that would make them useful for early detection.
There is a second boundary, and it sits in the hardware. The signals the method leans on are stage-level: deep sleep, light sleep, REM, wakefulness. Those are exactly the estimates consumer trackers find hardest. In a 2021 validation study in the Journal of Medical Internet Research, 59 first responders wore a Fitbit Charge 2 alongside polysomnography; epoch-by-epoch accuracy was 0.86 for REM sleep and 0.82 for wake after sleep onset, but 0.78 for N3 and 0.55 for light N1+N2 sleep. A framework that reads the timing of deep and light sleep is therefore reading a device's best guess at the stages it guesses worst, which is an argument for replication against sleep-lab recordings rather than an argument against the idea.
For a family, the practical meaning is modest and worth stating plainly. Nobody should read a tracker's sleep-stage graph and see a diagnosis, or an all-clear, in it. What this work offers is a direction of travel for researchers: if the useful signal in sleep is about when things happen during the night rather than how much of them there is in total, then the cheap device already on the bedside table may be collecting the right raw material, and the job becomes proving it in hundreds of people rather than thirteen.