Almost everything that lowers your dementia risk is what already protects your heart: blood pressure, blood sugar, weight, cholesterol, not smoking, staying active. Only one of those levers has a randomized trial showing it cuts dementia diagnoses rather than just test scores — lowering blood pressure in older people with hypertension, and even there the benefit is small. The dedicated “brain-health bundle” (diet plus exercise plus brain-training plus vascular checks) moves a cognitive test score at best, comes up null on the disease itself, and no trial has found a high-risk subgroup it helps. Single foods, supplements and named diets like MIND show promising associations that shrink or vanish once anyone runs the experiment. The headline is calibrated uncertainty, not a list of confident brain-specific fixes -> Age-Related Diseases, Surrogate Outcomes. Do the cardiometabolic basics early; skip the supplements.
Most of what lowers dementia risk is what protects your heart
The Lancet Commission’s headline is «nearly half of all cases of dementia being associated with 14 potentially modifiable risk factors» — a modelled population attributable fraction (Livingston et al., 2024). That figure usually arrives stripped of three caveats. It is a population attributable fraction, not a personal guarantee — it assumes every factor is fully removed and is fully causal. Most of the 14 are the same cardiometabolic rocks already pulled for the heart: hypertension, diabetes, obesity, physical inactivity, smoking, high LDL. And they overlap, so their separate percentages cannot be added — a large share of the “dementia” benefit is the same benefit counted once for the heart and again for the brain.
In practice, a person managing blood pressure, glucose, weight and activity for cardiovascular reasons has already pulled most of the dementia levers. Sizing any brain-specific addition means asking what it adds at the margin, on top of rocks a healthy adult is pulling anyway -> Layer 1 - Ranking Interventions for a Stratum, Age-Related Diseases. The levers work earlier and longer: midlife hypertension and hearing loss carry more weight than the same factor late in life (Livingston et al., 2024). What you measure to track them is what you already measure for the heart — blood pressure, HbA1c, waist, LDL — not a brain scan.
The dedicated brain-health bundle moves a test score, not the disease
Four trials tested a packaged multidomain program. Read them by endpoint and a clean split appears: the bundle nudges a cognitive test battery (a surrogate) but does not lower dementia incidence (the disease) -> Multidomain Lifestyle Intervention and Cognitive Decline, Surrogate Outcomes.
FINGER moved a surrogate, and only a little
FINGER randomized at-risk older Finns to a two-year diet-exercise-cognition-vascular bundle. The between-group difference in the neuropsychological test battery was +0.022 Z-score points per year (95% CI 0.002-0.042, p=0.030) (Ngandu et al., 2015) — statistically real, small, and on a test score, not a diagnosis. FINGER did not run long enough to count dementia cases; its disease-endpoint verdict waits on the extended follow-up.
MAPT and preDIVA: null on the surrogate, null on the disease
MAPT tested a similar bundle (with and without omega-3) over three years and found no significant benefit on its cognitive composite overall (Andrieu et al., 2017). preDIVA ran the harder test — a six-year nurse-led vascular-care program against incident dementia — and came up null: HR 0.92 (95% CI 0.71-1.19, p=0.54) (van Charante et al., 2016). So the one bundle trial powered for the disease did not move it.
No high-risk subgroup rescues the bundle
The obvious defense — it works, just in the right people — was tested directly. Coley pooled MAPT and preDIVA and searched systematically for a responder subgroup: «We did not identify any subgroups of older adults in whom multidomain interventions significantly reduced incident dementia» (Coley et al., 2025). This is a route-(b) claim that failed its own test: a responder subgroup needs positive interaction evidence, and the well-powered pooled analysis found none -> Baseline Risk and the Relative-Absolute Split. It is a worked case against over-personalizing on hope rather than evidence.
More components is not more benefit
A network meta-analysis of single- and multidomain programs found that stacking components did not beat the best single ones, and the whole comparison sits on cognitive-surrogate endpoints, not dementia incidence (Mendes et al., 2025). The bundle’s appeal is theoretical; the evidence does not reward the stack.
The single levers, ranked by what each actually moves
Blood pressure — the one randomized win on the disease
Pooling five trials of blood-pressure-lowering in older adults, Peters found an odds ratio of 0.87 (95% CI 0.75-0.99) for incident dementia (Peters et al., 2022). This is the only modifiable lever with randomized evidence on the disease itself, and it is route (a) — the relative effect is roughly uniform, so absolute benefit rises with baseline risk. In absolute terms the effect is small: on the order of 0.4 percentage points over ~4 years (NNT ~250). Real, worth having, already indicated for the heart — not a brain-specific add-on.
Peters tested for non-linearity and found none substantive (Peters et al., 2022), so the observational hint that very low pressure in old age is harmful is evidence against a J/U-shaped harm, not for one -> The U-Shaped Association Artifact. The fuller cardiovascular case, harms and absolute sizing live at Blood Pressure Lowering and Cardiovascular Events.
Diabetes drugs — confounded by indication
Confounding by indication dominates whether a particular glucose-lowering drug class protects the brain: sicker patients get different drugs, so head-to-head class rankings from observational data are unreliable (Kuate Defo et al., 2023). Treat glucose control as one of the shared cardiometabolic rocks; do not chase a specific agent for its claimed brain benefit on current evidence.
Physical activity — direct effect on dementia is unproven
Being active is a big rock for the heart, and inactivity clearly causes the vascular diseases that raise dementia risk. Whether activity protects the brain directly and independently is genuinely contested. Iso-Markku’s meta-analysis finds the association survives long follow-up (RR 0.80, 0.77-0.84) (Iso-Markku et al., 2022). But Kivimaki’s individual-participant analysis finds it vanishes beyond ten years (HR 1.01, 0.89-1.14), consistent with reverse causation — the dementia prodrome makes people less active years before diagnosis (Kivimäki et al., 2019). The clean evidence in both nearly converges on the null -> Does Physical Activity Protect Against Dementia Beyond Ten Years. Be active regardless — the vascular warrant is not in dispute; the brain-specific one is insufficient.
Hearing loss — a large marker, a weak treatment lever
Hearing loss is one of the strongest risk markers: HR 1.35 (95% CI 1.26-1.45) for incident dementia (Yu et al., 2024). But an exposure association is not a treatment lever. The ACHIEVE randomized trial of hearing aids found no overall benefit on three-year cognitive change; a pre-specified high-risk subgroup improved (hypothesis-generating, not confirmatory) (Lin et al., 2023) -> Hearing Loss and Dementia. Correcting hearing has its own quality-of-life case; do not oversell it as proven dementia prevention.
Education, mentally demanding work, staying social
Higher cognitive engagement across life is associated with lower risk — mentally demanding work carried HR 0.77 in Kivimaki’s cohort (Kivimäki et al., 2021) -> Cognitive Stimulation at Work and Dementia. These support the cognitive-reserve idea but are observational and hard to separate from lifelong confounders. Reasonable to favor, weak as a claimed intervention.
Diet and single foods: the signal fades at the diagnosis
The MIND diet — an observational benefit a trial did not confirm
Observational cohorts tie the MIND diet to better cognitive function (pooled +0.042 per SD, 0.020-0.065) (Huang et al., 2023). The one randomized test, Barnes, returned a between-group difference of +0.035 SD (95% CI -0.022 to 0.092) — a positive point estimate that does not reach significance (Barnes et al., 2023). Both arms of the tension survive: the trial’s interval rules out a large effect but not a small one, and its point estimate sits inside the observational range, so the RCT bounds the claim rather than overturning it -> MIND Diet - Observational Benefit vs Randomized Null. And no trial has tested MIND or Mediterranean on incident dementia at all. Eat this way for its cardiometabolic merits; do not bank on it as a proven brain diet.
The single-food ceiling
Line up the single-component arms — fruit and vegetables, flavonoids, soy, dairy — and the same shape repeats: a modest signal on a soft or surrogate endpoint that disappears at the hard Alzheimer’s or dementia diagnosis -> Single Food Components and Cognitive Outcomes. Fruit and vegetables: cognitive-impairment OR 0.76 but AD null, 0.88 (0.76-1.01) (Zhou et al., 2022). Flavonoids: decline OR 0.88 but dementia null, 0.97 (0.79-1.19) (Peng et al., 2026). Soy’s only randomized signal is a short-term memory test score, with no dementia-event evidence (Cui et al., 2019) -> Soy Products. Dairy is null throughout (Villoz et al., 2024) -> Dairy.
No single food or supplement is an established hard-endpoint lever. Do not chase foods or pills for cognition; the value is in the whole pattern routing through the cardiometabolic rocks.
Sleep and the weaker candidate markers
These carry a plausible directional signal but no established intervention on incident dementia — hold them at the certainty the evidence supports, not above it.
- Sleep — disordered sleep is associated with cognitive decline and dementia, but long sleep duration behaves like a marker of the prodrome rather than a cause; no trial shows that changing sleep prevents dementia (Zhang et al., 2025) -> Sleep.
- Ultra-processed food — associated with higher dementia risk in cohorts, confounded and observational (Henney et al., 2023).
- Air pollution — a small association whose interval crosses the null in the current meta-analysis (Wilker et al., 2023).
- Omega-3 / fish — cohort signal, no confirmed dementia-prevention effect from supplementation (Wei et al., 2023).
What we still don’t know
- Does the bundle work on the disease over a longer horizon? FINGER’s extended follow-up on incident dementia is the pending test.
- Is there any RCT of a diet pattern on incident dementia? None exists; the MIND trial measured cognition, not the disease.
- Does physical activity protect the brain independent of its vascular effect? Insufficient evidence -> Does Physical Activity Protect Against Dementia Beyond Ten Years.
- Do hearing aids help the highest-risk? ACHIEVE’s positive subgroup is hypothesis-generating only.
- The loop is open (R1): no source here grades a lever against a realized dementia outcome under a lifetime of intervention. A clean appraisal is not a validated promise.
What to do
Pull the cardiometabolic rocks early and keep them pulled — blood pressure, glucose, weight, LDL, not smoking, staying active. Blood-pressure control is the one lever with randomized evidence on the disease, and the rest carry their heart benefit for free. Correct hearing loss for its own sake and because the risk marker is large, without expecting proven dementia prevention. Eat a Mediterranean/MIND-style pattern for its cardiometabolic value, not as a validated brain intervention, and skip single-nutrient supplements for cognition — none has hard-endpoint support. Do not buy a packaged “brain-health program” expecting it to beat the basics; the bundle has never beaten the disease and no subgroup rescues it. Realistically, expect a modest shift in the odds, largest for those starting at high cardiovascular risk — worth doing, honestly small.
Evidence box
Question ’For an adult who wants to lower their risk of dementia: which modifiable exposures change dementia incidence versus only a cognitive-test score, what does a dedicated multidomain brain-health bundle do relative to acting on single cardiometabolic levers, how large is each effect in absolute terms, and does a person”s baseline risk change what the bundle achieves?‘ Evidence included 23 sources — 16 gold, 7 high Overall certainty Low-moderate (see Rating Certainty of Evidence) Source-selection note All sources are gold or high tier. Last updated 2026-10-03 · Independently reviewed: No · Full edit history