When Your Brain Is Older Than You Are, Sleep Knows First

The number that may matter most for dementia risk is not the one on your birthday card. It is the difference between your chronological age and your brain age, an estimate derived not from a blood test or a brain scan but from the electrical signals your sleeping brain produces in the small hours. According to a new meta-analysis of more than 7,000 people tracked for up to 17 years, that gap may be among the most reliable early warnings of cognitive decline we have yet found in community populations.

Every decade of excess brain age, as measured from overnight EEG, was associated with a 39% increase in dementia risk. The finding, published today in JAMA Network Open and led by researchers at UC San Francisco and Beth Israel Deaconess Medical Center in Boston, held up after accounting for education, BMI, smoking, sleep medication use, physical activity, most major comorbidities, and even the APOE e4 allele, the strongest known genetic risk factor for Alzheimer’s disease.

The central idea is deceptively simple. Train a machine-learning algorithm on EEG recordings from cognitively healthy people across the adult lifespan, and it learns what “normal” brain electrical activity looks like at each age. Present it with a new recording, and it produces an estimate of the brain’s biological age. The gap between that estimate and the person’s actual age is what the researchers call the brain age index. A score of minus ten suggests a brain behaving roughly a decade younger than its owner. A score of plus ten suggests the opposite, and, as it turns out, considerably worse odds.

What makes this work where standard sleep analysis consistently fails is that the model draws on the microstructure of sleep rather than its gross architecture. Clinicians have long measured how much time people spend in each sleep stage, how efficiently they sleep, whether they have apnoea. A prior pooled analysis of largely the same cohorts found essentially no link between those conventional metrics and dementia risk. “Broad sleep metrics don’t fully capture the complex multidimensional nature of sleep physiology,” said Yue Leng, associate professor of psychiatry at UCSF and the study’s senior author.

The 13 features the algorithm actually uses tell a different story. Sleep spindles, bursts of high-frequency activity during light sleep that are central to memory consolidation, contribute. So does the coupling between spindles and slow oscillations, the mechanism by which the sleeping brain is thought to replay and cement daytime experiences. Delta wave power during deep sleep enters the model. And perhaps most striking, the statistical “peakedness” of brain wave amplitude distributions during light and intermediate sleep, a measure called kurtosis that is thought to reflect K-complex activity, the sudden large spikes that mark a brain coping well with the night, turned out to be the single strongest protective signal in the whole analysis.

“Brain age is calculated from sleep brain waves,” said Leng. “We know that brain activity during sleep provides a measurable window into how well the brain is aging.” That the window is visible during ordinary overnight sleep, recorded unobtrusively at home on equipment no more alarming than a sticky electrode or two, is what gives the finding its practical weight. Dementia biomarkers based on cerebrospinal fluid require a lumbar puncture. Amyloid PET imaging costs thousands of dollars. A sleep study, potentially conducted with a consumer-grade wearable in a few years, is a considerably more appetising proposition.

The sample from which these conclusions are drawn is large and diverse by the standards of this kind of research: 7,105 participants drawn from five long-running US cohort studies, ranging in age from 40 to 94 when their sleep was recorded, followed for as little as 3.5 years and as many as nearly 17. Just over 1,000 ultimately developed dementia. The cohorts include multi-ethnic urban populations (the Multi-Ethnic Study of Atherosclerosis), community-dwelling older men (the Osteoporotic Fractures in Men Study), and the long-running Framingham Heart Study offspring cohort, among others. Consistency of the brain age effect across all of them, and across men and women and younger and older age groups alike, suggests the signal is real rather than the artefact of any single study’s design.

The neurobiological plausibility is not hard to sketch. Spindle generation depends on thalamo-hippocampal circuits, precisely the network that Alzheimer’s disease dismantles early. Higher concentrations of tau protein and amyloid in cerebrospinal fluid have been associated in other work with deteriorating spindle activity and slower oscillations, suggesting the brain age index may be picking up on changes that precede symptoms by years, possibly much longer. The fact that adjusting for APOE genotype barely moved the association suggests the EEG signal is not simply a proxy for genetic susceptibility; it seems to be capturing something additional, perhaps reflecting whatever degree of resilience or vulnerability the network has accumulated through the combination of genetics, lifestyle, and plain luck.

That raises the obvious question: can any of it be modified? First author Haoqi Sun, who developed the machine-learning model, was measured in his answer. “Better body management, such as lowering body mass index and increasing exercise to reduce the likelihood of apnea, may have an impact,” he said. Prior studies do show that treating sleep-disordered breathing can shift EEG patterns in measurable ways. But Sun was quick to add: “There’s no magic pill to improve brain health.” The brain age index is a marker, not a target. Identifying someone whose brain is running years ahead of schedule opens the question of why; it doesn’t immediately tell you what to do about it.

The present study cannot resolve that, nor does it try to. The cohorts use different dementia ascertainment methods, which introduces noise; the MESA cohort relied partly on hospitalisation codes, which are known to undercount. The model itself was trained on clinical populations that skew White, which may limit its precision in more diverse groups, though the multi-ethnic MESA data offer some reassurance. Wearable EEG devices have not yet been validated for brain age estimation specifically, so the path from research finding to scalable clinical tool still requires work.

What the analysis does establish, across a large, varied, long-running data set, is that the signature of accelerated brain ageing is present in sleep EEG years before dementia manifests, and that existing clinical tools have been largely blind to it. The usual measures of sleep quality are, in effect, asking the wrong question. The brain’s actual age, written in its electrical rhythms during the night, has been available to read for some time. We are only now beginning to understand the language.

DOI / Source: https://doi.org/10.1001/jamanetworkopen.2026.1521

Related