Exposome study ties 76 everyday exposures to dementia risk in half a million UK adults

Most of what is known about the causes of dementia arrived one suspect at a time: a study of smoking here, of air pollution or schooling there. A team at Harvard has tried the opposite approach, testing 193 features of people’s lives against dementia at once in half a million British adults.

Of those 193, 76 were linked to dementia over 15 years of follow-up, spanning income, housing, work, habits, diet, medicines and the air outside people’s homes. Rolled together into a single score, the exposures separated a top quarter of participants with about three times the dementia rate of the bottom quarter. The list also shows the limits of the approach: alongside familiar suspects such as smoking and low income, it flags never eating eggs and frequent sunbed use, both linked to more dementia, and time at a computer, linked to less. The study, led by Chenshuang Li with Peng Gao and Chirag Patel as senior authors, at the Harvard T.H. Chan School of Public Health and Harvard Medical School, was published on September 29 in Alzheimer’s & Dementia.

What is the exposome?

The idea behind the study is the exposome, a term for the totality of human environmental exposures from prenatal development through adulthood, as two University College London researchers put it in a 2025 review. Genes are fixed at conception; the exposome covers everything else that happens to a person, from the neighborhood they grow up in to the job they hold and the air on their street.

Dementia research has long pointed at parts of it. In the Harvard authors’ summary, the Lancet Commission on dementia identified 14 modifiable risk factors accounting for around 40% of cases, all of them things that could in principle be changed. Its list runs from lower education and hearing loss to smoking, high blood pressure, social isolation, air pollution and visual impairment. Those factors cluster. A person who rents may also earn less, live by a busier road and smoke, so a study of any one of those things inherits the others. The Harvard team argues that traditional single-exposure studies cannot capture these interdependencies, and borrowed a method from genetics to try.

How 193 exposures were screened in half a million Britons

The data came from UK Biobank, which recruited about half a million people aged 40 to 69 between 2006 and 2010 and has followed their health records since. At recruitment each person answered long questionnaires and had their home address linked to estimates of pollution, noise and green space. The researchers kept every non-genetic measure with enough complete answers, 193 in all, and sorted them into five groups: social, physical, chemical, lifestyle and the surrounding environment.

They then checked each exposure, one by one, against who went on to develop dementia, in the style of a genome-wide scan, where thousands of gene variants are tested and only those clearing a very strict statistical bar count as hits. Every test allowed for age, sex, ethnicity, overall health, the assessment center and APOE status, the best-known gene variant behind Alzheimer’s risk.

Over a median of 15.3 years, 8,881 participants developed dementia of any kind, about 1.8% of the cohort, including 4,000 with Alzheimer’s disease.

Which exposures went with more dementia, and which with less

The screen found 76 exposures significantly associated with dementia risk, meaning their links cleared that strict bar: 45 pointing toward higher risk and 31 toward lower.

Money and housing loomed large. Unemployment carried about 72% higher risk, financial difficulties 39% more and lower educational attainment 28% more. Compared with owning a home, renting went with 64% higher risk and living in a mobile structure, such as a trailer, with 60% higher. The authors suggest these may reflect financial strain, chronic stress and less access to health-promoting resources.

Habits showed up too. Current smoking, drinking outside recommended limits, frequent daytime dozing and sleeping too much or too little all went with higher risk. On the other side, regular exercise, heavy do-it-yourself jobs such as gardening and carpentry, walking for pleasure, living with a partner, higher household income and owning more cars all went with lower risk.

Air pollution was on the list as well. Fine particles and nitrogen dioxide, a traffic pollutant, went with higher risk across most subgroups of participants. That fits independent evidence: a 2023 meta-analysis judged that fine particles might be a risk factor for dementia, with thinner evidence for nitrogen dioxide, though most of the studies it pooled were at high risk of bias.

Several of these match the Lancet Commission’s list. Others are harder to place. People who said they never ate eggs had about 76% more risk than people who did, and more time on a computer was associated with reduced risk.

Why never eating eggs made the dementia list

A screen like this finds whatever travels with dementia, and much of what travels with it is not a cause.

The clearest example is the single strongest predictor in the authors’ model: the number of medications taken. Every 2.7 or so extra medicines, the study’s standard unit for this measure, went with about 19% more risk. The authors write that medication count likely reflects underlying comorbidity burden, meaning other illnesses, not any effect of the drugs themselves. Someone on many prescriptions often has heart disease, diabetes or other conditions that are themselves tied to dementia.

Reverse causation is the second trap. Dementia develops slowly, and its early stages can change how people eat, sleep, move and spend their time years before diagnosis. A pooled analysis of 19 cohorts in the BMJ showed how much this matters. Physical inactivity measured in the decade before diagnosis went with higher dementia risk, but when activity was measured ten or more years earlier there was no difference in dementia risk between active and inactive people. The UK Biobank exposures were recorded once, at recruitment, and some participants were within a few years of diagnosis. In this study, too, the links for mood and sleep factors were highest at early follow-up and faded later, which fits that pattern.

Confounding, where an exposure stands in for something unmeasured, is the third. Avoiding eggs, sunbed use and computer time could each mark differences in health, wealth or education that the models did not fully capture. The authors acknowledge this, writing that the modifiable factors they found “need to be evaluated to rule out reverse causation and confounding and to gauge clinical impact.”

How well the exposome score sorted dementia risk

To test whether the exposures could predict dementia, the team trained a machine-learning model on two thirds of the participants and checked it on the remaining third. The usual accuracy measure runs from 0.5, a coin toss, to 1, a perfect forecast. A baseline model containing only age, sex, ethnicity, overall health and APOE status scored 0.678. Adding the exposures raised it to 0.718, a modest gain.

A score built from the same exposures answers a different question. The accuracy figure says how well the model ranks one person against another; the score shows how far apart groups at the two ends end up, and there the gap was wider. The authors say the score “effectively stratified the population into distinct risk groups.” Compared with the lowest quarter, dementia rates rose step by step: about 43% above it in the second quarter, 83% above it in the third and about three times as high in the top quarter. The model worked best in people under 65, where it scored 0.810, and gave its weakest performance in the high genetic risk group, carriers of the risky APOE variant, where it scored 0.586, barely above chance.

For comparison, a different team built a simpler dementia score from the same cohort in 2023 using 11 items, among them age, education, family history, deprivation, diabetes, stroke, depression and blood pressure. It reported an accuracy of 0.80 in the UK Biobank test sample and 0.77 in a separate group of British civil servants. The two studies were set up differently, so the figures are not a head-to-head contest, but a short list of health measures reached a higher figure than the 193 exposures did.

The share of people still free of dementia in the high- and low-score groups began to differ within 2.5 years of recruitment, sooner than the gap between APOE groups appeared. The authors note that this early divergence may be partly attributable to confounding, and early dementia changing people’s answers would produce the same pattern.

Where the exposome study falls short

UK Biobank volunteers are not typical. Only 5.5% of those invited participated in the baseline assessment, and a 2017 comparison found evidence of a “healthy volunteer” selection bias: participants smoked less, drank less daily and were better off than the population. The study’s authors list the same problem, along with others: exposures were assessed at a single baseline time point, many were self-reported, dementia was identified from health records that can miss cases, and the cohort was mostly White British.

The score has not been tested outside UK Biobank. The authors write that external validation in independent cohorts is needed before it could be used anywhere else.

What the exposome results mean

The study maps how dementia risk is spread across the parts of people’s lives that are not written in their genes, and it confirms that money, housing, smoking, inactivity and polluted air sit among them. It does not show that changing any single item on its list, such as starting to eat eggs, would change anyone’s risk.

The authors see a use for the score in finding people at high risk for prevention studies, writing that it “has the potential to contribute to clinically relevant prevention strategies” for dementia. That remains to be tested.

Do not stop or change a prescribed medicine without talking to your doctor. Whether a particular person is at raised risk of dementia is assessed by their doctor.

The Harvard score sorts dementia risk modestly better than age and genes alone, and the causes behind most of its 76 links are still to be untangled.