The decision. You cannot buy a dedicated brain by assembling a brain-health programme. The one lifestyle change with a randomized trial showing it prevents the diagnosis of dementia — not just a better score on a cognitive test — is lowering high blood pressure, and that is a lever most people at risk are already being told to pull for their heart. Everything else on the popular list divides into two piles: cardiometabolic rocks (blood pressure, diabetes, weight, physical activity) that plausibly carry a real but modest dementia dividend because they are already being managed for cardiovascular disease, and a long tail of diet, supplement and single-food candidates whose signal shows up on soft cognitive-test scores and then vanishes at the hard diagnosis. The honest ceiling: control your cardiovascular risk, stay active, protect your hearing and sleep, and treat the rest as unproven.
How confident, and in what. Confidence is low-to-moderate, and it is uneven. The blood-pressure result rests on pooled randomized data and is about as solid as this field gets; the multidomain-bundle verdict is well-tested and clear; almost everything else is observational, low-certainty, and open to the reverse causation that a decades-long dementia prodrome invites. The loop is open — no analysis here has been checked against whether a real person who followed this advice actually avoided dementia. The wiki grades whether the reasoning is coherent and faithful to its sources, never whether it is right about the world.
The one discipline that governs the whole page. A cognitive-test score is a surrogate; incident dementia is the disease. A moved test score is not a prevented dementia unless the link between them is itself evidenced, and here it usually is not -> Surrogate Outcomes. Read every claim below for which endpoint the trial actually hit.
1. The crux: a brain-health bundle versus one cardiometabolic lever
Two bets have been run against dementia, and they must be kept apart. One is a multidomain lifestyle bundle — diet, exercise, cognitive training and vascular monitoring packaged and delivered as a unit. The other is a single lever — pull one cardiometabolic rock and watch dementia risk. The bundle has been tested three times and pooled once; it does not clear the bar. The single lever has been tested once at scale, on the disease itself, and it does.
The bundle moves the test score more reliably than it moves the disease. FINGER, the trial everyone cites, ran a 2-year four-component programme in at-risk elderly (CAIDE score >=6, near-normal cognition) and improved a cognitive-composite score by 0.022 SD per year versus an active control (95% CI 0.002-0.042), Cohen’s d 0.13 (Ngandu et al., 2015). That is real, statistically significant, and small — the lower bound sits near zero, and the trial never measured dementia incidence. FINGER moved a surrogate in a selected stratum. It did not show dementia was prevented.
MAPT is the sharper test, because its design matches FINGER most closely: a 3-year four-arm French trial of the same kind of bundle on the same surrogate. At the whole-population level it was null — the combined-intervention effect (0.093 points over 3 years, CI 0.001-0.184) fell to adjusted p=0.142 once corrected for multiple comparisons, and even at face value sits below the composite’s minimal clinically important difference (Andrieu et al., 2017). MAPT’s own CAIDE>=6 and amyloid-positive subgroups did reproduce a benefit — pointing, like FINGER, toward a high-risk responder story on the surrogate.
That responder story dies at the hard endpoint. preDIVA ran a 6.7-year unselected cluster-RCT of nurse-led vascular care, measured clinical dementia diagnosis directly, and found nothing: HR 0.92 (95% CI 0.71-1.19) (van Charante et al., 2016). The trial was powered to detect a 33% incidence reduction — far larger than a vascular-care effect would plausibly deliver — so this null is best read as underpowered for a small effect, not proof nothing happened. Still, it is the first time the bundle itself, rather than a proxy for it, met the disease it is meant to prevent, and it came up empty.
Do more domains help? No. A network meta-analysis of 109 RCTs (23,010 cognitively-unimpaired older adults) ranked the levers on the global-cognition composite. The top combination was physical exercise plus cognitive training (SMD 0.26, 95% CI 0.10-0.42); the fuller four-component bundle landed lower, at SMD 0.14 (0.02-0.27), no better than exercise alone (Mendes et al., 2025). Adding domains did not add benefit — and this whole ranking is on the test-score surrogate, in the same class of trials FINGER and MAPT ran, so it does not convert into the hard-endpoint prevention the pooled null denies.
2. Does targeting the highest-risk person rescue the bundle?
The natural way to save a bundle that fails on average is to give it to the people most likely to respond — those at highest baseline dementia risk. The well-powered test says this does not work.
Coley pooled individual-participant data from MAPT and preDIVA (n=5205, up to 12 years’ follow-up, 486 incident dementia cases) and searched for a responder. The pooled intervention effect on all-cause dementia is flat — HR 0.98 (95% CI 0.80-1.21) — and stays flat inside every one of 11 pre-specified subgroups. The two cells that had looked promising on the surrogate stayed null on the disease: for incident Alzheimer’s, CAIDE>=6 gave HR 1.03 (0.76-1.40) and preDIVA’s untreated-hypertension responder cell HR 0.72 (0.44-1.18, non-significant). A data-driven search free to combine any variable at any cutpoint found no responder subgroup either (Coley et al., 2025).
This is a clean verdict against route-(b) personalization by dementia risk — not merely untested but evidenced against on the hard endpoint -> Baseline Risk and the Relative-Absolute Split. It is one of the corpus’s better instances of the maintainer’s standing suspicion that over-personalization is the likelier failure: a subgroup that shone on a surrogate did not survive the disease-endpoint test.
Note what this does not say. It does not prove the bundle is worthless for everyone — preDIVA and the pool are underpowered for a small effect. It says the specific, non-decomposable package tested three times has not been shown to prevent dementia, and aiming it at the highest-risk person does not change that. The FINGER 7-year extended follow-up on incident dementia, which could still move this verdict, is not held ->.
3. The levers, ranked by what each demonstrably moves
Behind the bundle question sits the list of single factors. The Lancet Commission estimates 14 modifiable factors could be associated with around 45% of dementia cases as a population-attributable fraction (Livingston et al., 2024). Read that as a population ceiling under strong causal and additivity assumptions, not a personal 45% you can claim — the fractions do not sum for an individual, and they borrow relative risks assumed causal. What matters for a decision is which levers actually move the disease, and how big each is in absolute terms.
Blood-pressure lowering — the one randomized win on the disease. Peters pooled individual-participant data from five double-blind, placebo-controlled antihypertensive trials (28,008 people, 861 incident dementia cases): a sustained ~10/4 mmHg reduction cut incident dementia, OR 0.87 (95% CI 0.75-0.99) (Peters et al., 2022). In absolute terms this is small — dementia occurred in 2.9% of the treated group versus 3.3% of placebo over a median 4.3 years, an absolute risk difference near 0.4 percentage points, roughly one dementia case averted per 250 people treated for four years. The trials stopped early once their cardiovascular endpoint was met and dementia accrues slowly, so this is plausibly a floor. The relative benefit did not vary by age, baseline pressure or stroke history, so the decision runs on baseline risk (route-a), not on finding a special responder (Peters et al., 2022). And the observational U-shape — where very low pressure in old age looks harmful — did not survive randomization: the dose-response stayed linear down to at least 100/70 mmHg -> The U-Shaped Association Artifact.
Two things bound this win. It is single-lever, not a brain regimen — just blood-pressure control, the same intervention already recommended for the heart, so for most at-risk adults it is a second patient-important outcome bought on a rock already being pulled, not a new thing to add -> Layer 1 - Ranking Interventions for a Stratum. The full cardiovascular case for BP-lowering, and its own harms and absolute sizing, live at Blood Pressure Lowering and Cardiovascular Events. And this is the last evidence of its kind: it is no longer ethical to randomize people to placebo blood pressure, so no larger trial will supersede it.
Diabetes — a counted rock, with a drug-class question that stays prescriber-zone. Diabetes raises Alzheimer’s risk (RR ~1.39-1.57) and vascular dementia more strongly, and is already one of the cardiometabolic rocks (Kuate Defo et al., 2023). An umbrella review of 27 observational studies (N~3.05 million) finds which glucose-lowering drug associates with different dementia risk — metformin (RR 0.83), GLP-1 receptor agonists (0.35), SGLT2 inhibitors (0.39) and pioglitazone (0.74) lower; sulphonylureas (1.39) and meglitinides (1.87) higher (Kuate Defo et al., 2023). But confounding by indication runs straight through this — metformin is first-line for milder disease, the secretagogues come later in more severe diabetes — heterogeneity is near I2=99%, no randomized trial backs any figure, and certainty is low to very low. Which agent a person takes is a prescriber decision outside this wiki; the appraisal here is only that a cognitive dividend for the newer agents is plausible and unproven.
Physical activity — survives the reverse-causation check. A 58-cohort meta-analysis (n=257,983) finds the standard association, RR 0.80 (0.77-0.84), and tests it against the obvious objection that early, undiagnosed dementia makes people less active: restricted to the 16 studies with 20+ years of follow-up, the estimate holds at RR 0.79 (0.71-0.87) (Iso-Markku et al., 2022). That is the strong-adjudication direction the check is meant to supply. It is not proof of a causal slope — the three highest-quality young-baseline, long-follow-up studies lose significance (RR 0.79, 0.62-1.01), and a cognitive-reserve confound the design cannot remove stays live. One clean negative: protection does not depend on ApoE4 genotype, so activity is not a lever to withhold or intensify by genetic risk — a route-(b) null that simplifies the advice (Iso-Markku et al., 2022). Net: this firms an existing rock, it does not add a new factor.
Hearing loss — a large exposure, a weak treatment lever. Uncorrected hearing loss carries one of the largest attributable fractions and a tight exposure association: HR 1.35 (95% CI 1.26-1.45) across 50 cohorts (Yu et al., 2024). But a big risk factor is not an effective treatment. ACHIEVE, the one randomized trial of hearing aids built for this question (N=977, ages 70-84), found no difference in 3-year cognitive decline versus a health-education control: 0.002 SD (95% CI -0.077 to 0.081, p=0.96) (Lin et al., 2023). A pre-specified higher-risk subgroup (the ARIC cohort) showed a 48% smaller decline, but that is one subgroup of one trial at a lenient alpha, and the harder endpoint — incident cognitive impairment or dementia — was null in every stratum including that one (HR 0.90, 95% CI 0.61-1.33) (Lin et al., 2023). Honest reading: treat hearing loss as a well-evidenced risk marker and a reasonable bet concentrated at high baseline risk, not an established general-population dementia intervention -> Hearing Loss and Dementia.
4. The diet, sleep and single-food candidates
Below the rocks sits a crowded field of diet and lifestyle candidates. Each shows a statistically real association, and each weakens under the same three questions: does it survive restriction to the hardest subtype (Alzheimer’s), does it survive adjusting for the cardiometabolic factors already counted, and does a clean dose-response appear? The recurring pattern across single food components is the finding itself: the signal lives on soft cognitive-test endpoints and vanishes at the hard diagnosis -> Single Food Components and Cognitive Outcomes.
- Ultra-processed food associates with higher all-cause dementia risk (RR 1.44, 95% CI 1.09-1.90), but the moderate-intake category is null, no subtype individually reaches significance, and the association is lost after adjusting for type 2 diabetes and total energy — so it runs largely through rocks already counted (Henney et al., 2023).
- Fruit and vegetables associate with lower risk of cognitive disorders (OR 0.82, 95% CI 0.75-0.90), but Alzheimer’s specifically is null (0.88, 0.76-1.01), and the effect is weakest in the prospective-cohort designs least prone to reverse causation (Zhou et al., 2022).
- Flavonoids, nested inside those same fruits and vegetables, repeat the pattern one level down — cognitive decline moves (OR 0.88) but dementia (0.97) and Alzheimer’s (0.90) are both null (Peng et al., 2026).
- Soy isoflavone supplements show it from the trial side: 16 RCTs pooled to a small composite test-score gain (SMD 0.19, 95% CI 0.07-0.32), carried by memory, but every trial ran <=2 years and none counted a dementia case (Cui et al., 2019) -> Soy Products.
- Dietary DHA (mainly oily fish) associates with lower decline (RR 0.82, 95% CI 0.72-0.93), but omega-3 supplements were tested in MAPT and did nothing — a marker-versus-lever gap (Wei et al., 2023).
- Dairy is null at every endpoint: across 15 cohorts (312,580 people) the highest-versus-lowest contrast is RR 0.94 (95% CI 0.82-1.07), and the one dose-response dip near 150 g/day is a between-population artifact (Villoz et al., 2024) -> Dairy.
Sleep tracks with dementia across the board — insomnia RR ~1.13, long sleep over 8 hours RR ~1.66 for Alzheimer’s — but the long-sleep arm in particular looks like a preclinical marker of dementia already under way rather than a cause, and no trial has tested whether treating a sleep disorder lowers dementia incidence (Zhang et al., 2025). It is a screening-relevant candidate, not an established lever; the metabolic case for sleep lives at Sleep.
The MIND / Mediterranean diet is the pattern most often sold for the brain, and it carries the field’s cleanest surrogate-versus-disease tension. The observational base is broadly positive — Huang’s pooled cohorts put global cognitive function at +0.042 z per 1-SD MIND score (95% CI 0.020-0.065) (Huang et al., 2023). The one randomized test (Barnes, NEJM 2023, n=604, 3 years) found no significant between-group difference — a mean difference of +0.035 SD (95% CI -0.022 to 0.092, p=0.23), with null brain-MRI outcomes (Barnes et al., 2023). The trial point estimate is small, positive, and statistically consistent with the observational slope: the RCT bounds the effect (a large one is ruled out) without overturning the small association -> MIND Diet - Observational Benefit vs Randomized Null. So MIND is defensible on its cardiometabolic merits, not as a proven brain intervention. No RCT of any diet pattern on incident dementia is held — Barnes measured cognition, not the disease.
Where the candidates leave you. None of these is an independent fifteenth, sixteenth or seventeenth target stacked on top of the cardiometabolic rocks. Ultra-processed food, fruit and vegetables, flavonoids and DHA are at most routes to pull blood pressure, glycaemia, weight and vascular health — the same rocks, reached through the plate. Chasing individual foods or isolated-component supplements (flavonoid extracts, isoflavone pills, fish oil) for the brain over-counts overlapping, confounded, soft-endpoint signals -> Layer 1 - Ranking Interventions for a Stratum.
Peripheral candidates, held at their evidence. Cognitive stimulation at work associates with lower dementia (HR 0.77, 95% CI 0.65-0.92, surviving a 10-year lag), but a childhood-IQ confound the design cannot remove keeps it below the cardiometabolic rocks (Kivimäki et al., 2021) -> Cognitive Stimulation at Work and Dementia. Air pollution shows a modest exposure-response (PM2.5 per 2 µg/m³ HR ~1.04, crossing the null in the most conservative pooling) (Wilker et al., 2023). Periodontal disease and cognitive reserve stay observational-only. Hold each at the certainty the fabric supports; do not promote an observational exposure to an established lever.
The bottom line
For a reasonably healthy adult who wants to lower dementia risk, the evidence points somewhere less exciting than a brain-training subscription and more useful. The multidomain brain-health bundle, tested three times against a test score and once against the diagnosis, then pooled and searched for a responder, has not been shown to prevent the disease — and targeting the highest-risk person does not rescue it. Lowering high blood pressure does prevent it, in randomized data, on the hard endpoint, at a small but real absolute size (about one case averted per 250 people over four years) — and for most at-risk adults that lever is already being pulled for the heart. Staying physically active firms a rock that survives the reverse-causation check. Protecting hearing and sleep are reasonable bets concentrated at high risk. The diet and single-food candidates are, on today’s evidence, routes to the cardiometabolic rocks rather than independent brain levers, and the isolated supplements are the weakest bet of all.
Two verdicts stand together without contradicting each other, because they answer different questions: a non-decomposable package has not been shown to work, and one well-understood exposure has. The broader shared-lever thesis this rests on — that most age-related diseases yield to the same handful of cardiometabolic rocks — lives at Age-Related Diseases.
Open loop. Nothing here has been graded against a realized outcome — whether a person who followed this actually avoided dementia. This is an appraisal of the evidence, coherent and faithful to its sources; it is not a validated prediction, and the confidence attached to each lever reflects the quality of the evidence, not a track record.
Evidence box
Question ’For an adult who wants to lower their risk of dementia: which modifiable exposures actually change dementia incidence rather than only a cognitive-test score, does a dedicated multidomain brain-health bundle prevent the disease or does the evidence run through single cardiometabolic levers already pulled for the heart, how large is each effect in absolute terms, and does targeting the highest-risk person rescue a bundle that fails on average?‘ Evidence included 22 sources — 15 gold, 7 high Overall certainty Low-moderate (see Rating Certainty of Evidence) Source-selection note All sources are gold or high tier. Last updated 2026-09-09 · Independently reviewed: No · Full edit history