BIONEWS
A model trained on overnight brain recordings from 7,000 adults found that brains appearing older than their real age tracked a sharply higher dementia risk — a population level signal, not a diagnosis.
An adult lies in a clinic bed, wired up for an overnight sleep study. By morning, a different kind of readout might come back from the EEG leads, the small metal discs that record the brain's electrical activity through the scalp. Not a chart of sleep stages this time, but a single number: a "brain age" estimated by a machine-learning model that reads 13 microscopic features of the brain's electrical activity during sleep. The model was built by researchers at UC San Francisco and Beth Israel Deaconess Medical Center, and its performance was first put to the test in an individual participant data meta-analysis published in JAMA Network Open, a peer-reviewed medical journal.
The team applied the model to roughly 7,000 adults pooled across five separate cohorts. None had dementia at enrollment; baseline ages ran from 40 to 94. About 1,000 went on to develop dementia during follow-up windows that stretched from 3.5 to 17 years. Across that pooled group, every 10-year gap between the model's estimated brain age and a person's chronological age came with a 39% higher chance of developing dementia (hazard ratio 1.39, 95% CI 1.21–1.59), the researchers report. People whose brains read younger than their calendar age tracked lower risk.
Earlier pooled studies, working from the same kind of overnight data, mostly came up empty. Standard sleep summaries such as total time in bed, time spent in deep sleep, sleep efficiency, and how often someone wakes failed to predict who would later develop dementia. The new model picks up a different layer: not how long someone slept, but the fine texture of the brain's electrical activity while they did.
"Traditional sleep metrics, such as sleep stages and efficiency, miss the multidimensional physiological information embedded in sleep," said senior author Yue Leng, MBBS, PhD, a UCSF psychiatrist who studies sleep and aging. The features the model leans on — the shape of individual sleep spindles and the distribution of slow-wave activity — are too small and too numerous for a human scorer to weigh by eye.
The 39% is a relative-risk figure, not a personal prediction. Drawn from a model rather than a single biomarker, it can shrink or grow depending on who is in the room, how their EEG was recorded, and how "dementia" was diagnosed across five different cohorts. The JAMA Network Open analysis was designed to average across that variation, not to deliver a verdict on any one overnight recording.
The study also does not yet show that acting on a high brain-age gap changes what happens next. Flagging someone as higher-risk, then watching whether earlier counseling, blood-pressure control, sleep-apnea treatment, or cognitive follow-up alters their trajectory, is the missing clinical step. Leng's argument points in that direction: the value, she and her coauthors argue, is a cheap, non-invasive triage layer that runs on sleep data many clinics already collect.
A 2018 ScienceDirect paper first estimated brain age from sleep EEG, and a 2023 Nature Scientific Reports study linked cognitive health to features extracted from sleep brainwaves. The JAMA Network Open analysis is the first to pool individual participant data across cohorts at this scale, which is what gives the 39% number its weight.
For now, the readout stays in the research file. A sleep-clinic pilot that flags a high brain-age gap and follows the patient for a decade would be the kind of prospective evidence the field still needs.