The verdict
You do not need a separate anti-dementia diet, an anti-cancer diet, and an anti-heart-disease regimen. A short list of shared modifiable levers — staying physically active, keeping blood pressure, LDL, blood sugar and body fat down, not smoking, going easy on alcohol — lowers the risk of several leading age-related diseases at once. That overlap is why these levers rank first: pull one, and several outcomes move together.
Two honest bounds keep that verdict from over-promising. First, most of the per-disease effects behind it are modest and observational — estimated across populations, not measured in any one person — and the levers do not transfer uniformly from disease to disease: Parkinson’s barely follows the pattern, and a lever that helps one eye disease does nothing clear for its neighbour.
Second, the strongest new interventional evidence the wiki holds runs through a single lever, not a purpose-built regimen: lowering blood pressure cuts incident dementia in randomized trials, a small effect in absolute terms but a real one — bought on a blood-pressure pill already prescribed for the heart, not from a dedicated brain-health program. The dedicated version of that program — bundling exercise, diet coaching, cognitive training and vascular monitoring specifically to prevent dementia — has been tried directly and has not cleared that bar: it moves cognitive test scores in some trials, but not, so far, the disease itself. Most of the diet-detail levers people spend the most attention on — ultra-processed food, extra fruit and vegetables, flavonoids, fish-oil capsules — work, where they work at all, through the same cardiometabolic pathways as the shared core. They reinforce those levers; they do not add new ones alongside them.
Start with why the same short list keeps reappearing.
One set of levers, several diseases
Line up the modifiable-risk-factor lists the wiki holds for four leading age-related outcomes — dementia, cardiovascular disease, cancer, Parkinson’s disease — and a small set of exposures recurs across most of them, in the same direction: excess body fat, physical inactivity, diabetes and high blood sugar, high blood pressure, high LDL, smoking, and heavy alcohol use. No single disease body states this; each was written for its own outcome. The recurrence itself is the finding.
| Exposure | Dementia | Parkinson’s | Cancer | Cardiovascular disease |
|---|---|---|---|---|
| Physical inactivity | RR 0.80 (0.77-0.84) protective | RR 0.79 (0.68-0.91) protective | a protective cause | established lever |
| Excess body fat | midlife RR 1.31 (1.02-1.68) | not significant | convincing/probable cause, 12 of 17 cancer sites | established lever |
| Diabetes / high blood sugar | HR 1.24 per 5-yr-earlier onset | RR 1.31 (1.10-1.57) | via body fat | established lever |
| High blood pressure | untreated HR 1.42 (1.15-1.76) | — | — | established lever |
| High LDL | >3 mmol/L HR 1.33 | — | — | established lever |
| Smoking | midlife RR 1.30 (1.18-1.45), harm | RR 0.64 (0.60-0.69), protective — an artifact, not a lever (see below) | a cause | established harm |
| Heavy alcohol use | >21 units/week HR 1.22 (1.01-1.48) | — | a cause | J-shaped; the protective-looking low-dose arm is artifact-suspect |
(Livingston et al., 2024) (Chen et al., 2021) (World Cancer Research Fund & American Institute for Cancer Research, 2018)
Read this table as co-membership evidence, not a shared magnitude. Each column is a different outcome, measured in a different population, over a different follow-up — a row does not mean “this exposure costs everyone the same amount”; it means the same exposure shows up, in the same direction, in disease bodies that were built independently of one another. Whether a given relative risk translates into a large or a small absolute benefit for any one person still depends on that person’s baseline risk for each disease — a question the later sections take up disease by disease.
On alcohol specifically: the harm direction above is the one to act on — past roughly 21 units a week, dementia risk rises, and alcohol is a recognized cause of several cancers with no evidence of a safe floor. The apparent protection at low intake is the textbook case of a U-shaped association that does not survive scrutiny of who ends up in the “non-drinking” comparison group — former drinkers who quit because they were already getting sick get folded into the reference category, which manufactures a protective-looking dip at low doses -> The U-Shaped Association Artifact. Nothing here should be read as licensing a “safe” drinking amount; the number worth carrying is the harm threshold, not a floor.
This is exactly the situation Layer 1’s ranking logic is built for. Ranking by expected effect size and certainty does not mean ranking one outcome at a time: a lever that moves three diseases outranks one that moves only one, at an equal effect on each -> Layer 1 - Ranking Interventions for a Stratum. Physical activity and the vascular-metabolic cluster (blood pressure, LDL, glycaemia, adiposity, smoking) sit at the top of both the Big Rocks (Elderly) and Big Rocks (Median) rankings for exactly this reason — not because any single per-disease effect is unusually large, but because the same short list keeps paying out across the whole outcome menu at once. The shared core is not a coincidence sitting beside the ranking; once multi-outcome breadth counts toward a lever’s rank, the shared core is the ranking.
Why does the same list keep recurring? The most economical explanation is that these exposures injure a common cardiometabolic-vascular substrate — the blood vessels and the systemic metabolic machinery that dementia, cardiovascular disease and (via body fat and inflammation) some cancers all depend on. That is a candidate mechanism, not an outcome finding: it is consistent with the overlap above but not independently tested here, and it does not crown any single exposure as “the cause” of age-related disease in general — an effect arrives only when a whole set of conditions is present, and ranking by breadth is a statement about where to act, not about what explains the outcome -> Layer 1 - Ranking Interventions for a Stratum.
One exposure in the table already breaks the pattern outright — smoking’s apparent protection against Parkinson’s, flagged above — and later sections return to why Parkinson’s and a few other diseases resist the shared-lever logic more broadly.
Dementia: does anything actually prevent it?
Whether pulling those levers actually prevents a disease — not just a risk marker — is best tested on the one where it has been tried head-on: dementia. Two different bets have been run against dementia incidence and cognition, and they need to be kept apart. One bet is a multidomain lifestyle bundle — diet, exercise, cognitive training and vascular monitoring packaged together and delivered as a unit. The other is a single-lever intervention — pulling one cardiometabolic rock and watching what happens to dementia risk. The bundle has been tested three times and pooled once, and it does not clear the bar. The single lever has been tested once, at scale, and it does.
The bundle, tested head-on, moves the test score more reliably than the disease
FINGER, the 2015 proof-of-concept trial, is the positive result everyone cites: a 2-year, 4-component programme in at-risk elderly (CAIDE score >=6, near-normal cognition) 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 and statistically significant, and it is also small — the CI’s lower bound sits close to zero, and dementia incidence itself was never measured; a 7-year extended follow-up was planned but is not yet held. FINGER shows the cognitive-composite surrogate moved in a selected high-risk stratum. It does not show dementia was prevented.
MAPT, a 3-year, 4-arm French trial testing the same kind of bundle (physical activity, cognitive training, nutrition, with or without omega-3) on the same kind of surrogate, is the sharper test because it is the closest design match — and at the whole-population level it was null. The combined-intervention group’s raw 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 it sits far below the composite’s estimated minimal clinically important difference (Andrieu et al., 2017). MAPT’s own CAIDE>=6 subgroup did reproduce a benefit (p=0.023), and so did its amyloid-positive subgroup — pointing, like FINGER, toward a high-baseline-risk responder story on the surrogate.
That responder story does not survive contact with a hard endpoint. preDIVA, a 6.7-year, unselected cluster-RCT of nurse-led vascular care (no cognitive training, no supervised exercise — a thinner bundle than FINGER’s), measured clinical dementia diagnosis directly and found nothing: HR 0.92 (95% CI 0.71-1.19) (van Charante et al., 2016). The trial did move blood pressure (-2.06 mmHg in the intervention arm) and was powered to detect a 33% incidence reduction — an order of magnitude larger than a BP-mediated effect would plausibly deliver — so this null is honestly read as underpowered-for-a-small-effect rather than a demonstration that nothing happened. Still, it is the first time the bundle, rather than a proxy for it, was checked against the disease it is meant to prevent, and it found nothing.
Coley’s pooled individual-participant analysis of MAPT and preDIVA (n=5205, up to 12 years of follow-up, 486 incident dementia cases) is the decisive test of whether targeting the high-risk stratum rescues the bundle on the hard endpoint. It does not. 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 subgroups that had looked promising on the surrogate stayed null on the hard endpoint too: in the incident-Alzheimer’s- disease analysis, CAIDE>=6 gave HR 1.03 (0.76-1.40) and preDIVA’s own 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).
A large network meta-analysis sharpens the surrogate side of the same story and adds a more-is-not-better twist. Mendes 2025 pooled 109 RCTs (23,010 cognitively-unimpaired older adults) and ranked the lifestyle levers on the global-cognition composite: the top combination was physical exercise plus cognitive training (SMD 0.26, 95% CI 0.10-0.42), while the fuller four-component bundle — diet, exercise, cognitive training and health education — landed lower, at SMD 0.14 (0.02-0.27), no better than exercise alone (0.14) (Mendes et al., 2025). So on the surrogate, adding more domains did not help — the exercise-plus-training pair beat the richer stack. But every one of these is a cognitive-test-score effect, in the same class of trials FINGER and MAPT ran; none is dementia incidence, and this within-surrogate ranking does not convert into the hard-endpoint benefit the pooled null denies.
So the honest summary of the bundle-as-tested is this: it moves a cognitive test score in some circumstances, it has not been shown to prevent the diagnosis, and deliberately targeting people at the highest baseline dementia risk — the most natural way to try to rescue it — does not change that. Route-(b) personalization by dementia risk is evidenced against here, not merely untested -> Baseline Risk and the Relative-Absolute Split.
One lever that works: lowering blood pressure
Against that backdrop, the strongest interventional dementia evidence the fabric holds is not a bundle at all. Peters 2022 pooled individual-participant data from five double-blind, placebo-controlled antihypertensive trials (HYVET, SYST-EUR, PROGRESS, ADVANCE, SHEP; 28,008 people, 861 incident dementia cases) and found that lowering blood pressure by a sustained ~10/4 mmHg 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 of about 0.4 percentage points, or roughly one dementia diagnosis averted for every 250 people treated for four years. The trials stopped early once their cardiovascular endpoint was met, and dementia accrues more slowly than cardiovascular events, so this is plausibly a floor rather than the full effect.
The relative benefit did not vary by age, baseline blood pressure, or stroke history — it looks uniform across those strata, so the decision runs on baseline risk (route-a), not on identifying a special responder group (Peters et al., 2022) -> Baseline Risk and the Relative-Absolute Split. About half the effect (53%, 95% CI 27-76%) traces statistically to the blood-pressure change itself, with the rest running through other trial or drug effects (Peters et al., 2022).
And the U-shaped pattern seen in observational studies of blood pressure and dementia in old age — where very low pressure looks harmful — does not survive randomization: the dose-response in these trials is linear down to at least 100/70 mmHg, with no rise in dementia risk at the oldest ages -> The U-Shaped Association Artifact. This is the last evidence of its kind that will ever be generated — it is no longer considered ethical to randomize people to a placebo arm on blood pressure — so it will not be superseded by a larger trial.
What this buys is real but bounded on every axis the plan flags. It is single-lever, not a dedicated brain regimen: no cognitive training, no diet counseling, no exercise programme — just blood-pressure control, the same intervention already recommended for cardiovascular disease. It is a second patient-important outcome bought on a rock already being pulled, not a new free-standing intervention to add -> Layer 1 - Ranking Interventions for a Stratum, Blood Pressure Lowering and Cardiovascular Events. And it is small in absolute per-person terms, exactly the scale a shared cardiometabolic lever contributing one slice of a multi-cause disease should be expected to produce.
None of this reopens the bundle question: Peters tested one lever, not a package, and its win does not rehabilitate FINGER’s or MAPT’s cognitive-composite signal into a dementia-prevention claim. Equally, the bundle’s failure does not mean nothing prevents dementia — it means the specific non-decomposable package tested three times did not, while a single, well-understood lever tested on the hard endpoint did.
The diabetes lever splits by drug class
Diabetes itself is already counted as a shared cardiometabolic rock — across the evidence base it raises Alzheimer’s risk (RR 1.39-1.57) and vascular dementia risk more strongly (RR 1.91-2.49) (Kuate Defo et al., 2023). A newer question sits inside that finding: for the large population already on a glucose-lowering drug, does which drug matter for cognition? An umbrella review of 27 observational studies (N=3,046,661) finds a split — metformin (RR 0.83, 0.71-0.96), GLP-1 receptor agonists (0.35, 0.16-0.78), SGLT2 inhibitors (0.39, 0.20-0.76) and pioglitazone (0.74, 0.55-0.98) associate with lower dementia risk, while sulphonylureas (1.39, 1.04-1.87) and meglitinides (1.87, 1.43-2.45) associate with higher risk (Kuate Defo et al., 2023).
The catch is confounding by indication: metformin is first-line therapy for earlier, milder disease, while sulphonylureas and meglitinides are typically used later, in more severe diabetes with more hypoglycaemia — so part of the apparent drug effect is a disease-severity contrast, not a brain effect, and metformin’s own benefit weakens with longer follow-up, the signature of healthy-adherer bias (Kuate Defo et al., 2023). With no randomized trial behind any of these numbers, heterogeneity near I2=99% for most drug classes, and two of the ten signals losing significance in the authors’ own sensitivity analysis, certainty here is low to very low. Which drug a person with diabetes is prescribed is a prescriber-zone decision outside this wiki’s scope; what belongs here is only the appraisal — a possible cognitive dividend for the newer agents, a possible cognitive penalty tracking the older secretagogues’ hypoglycaemia risk, both soft.
Physical activity holds up under scrutiny
Physical inactivity is already a counted big rock; what a first-hand look at its dementia evidence adds is a specific check that the association is not simply reverse causation — people in the early, undiagnosed stages of dementia becoming less active, rather than inactivity causing dementia. A 58-cohort SR+MA (n=257,983) finds the standard association, RR 0.80 (0.77-0.84), and then tests it against exactly that objection: restricted to the 16 studies with 20 or more 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 reverse-causation check is meant to supply.
It is not, however, a demonstrated causal slope: the three highest-quality studies with both a young baseline and 20+ years of follow-up lose significance (RR 0.79, 0.62-1.01), and a residual confound the design cannot remove — people who are more active may simply have started with more cognitive reserve — stays live (Iso-Markku et al., 2022). One clean negative: the protection does not depend on ApoE ε4 genotype (carriers RR 0.81, non-carriers RR 0.72, no significant interaction in 9 of 11 studies) — activity is not a lever to withhold or intensify by genetic risk, a route-(b) null that reinforces rather than complicates the recommendation (Iso-Markku et al., 2022). Net effect: this firms an existing rock with a second patient-important outcome; it does not add a new factor to the list.
Hearing loss: a large exposure, a weak treatment lever
Hearing loss carries one of the largest population-attributable fractions among the fourteen modifiable factors, and the exposure association is tight: a 50-cohort meta-analysis puts uncorrected hearing loss at HR 1.35 (95% CI 1.26-1.45) for incident dementia (Yu et al., 2024). But a large risk factor is not the same claim as an effective treatment, and the correction — hearing aids — has been tested in the one randomized trial built for the purpose, and it failed on its primary outcome. ACHIEVE (N=977, ages 70-84) found no difference in 3-year cognitive decline between people randomized to hearing aids and a health-education control: 0.002 SD (95% CI -0.077 to 0.081, p=0.96) (Lin et al., 2023).
A pre-specified subgroup — the higher-risk ARIC cohort, as opposed to healthier self-referred volunteers — did show a 48% smaller 3-year decline (0.191 points, 95% CI 0.022-0.360, p=0.027), but this is one subgroup of one trial at a lenient pre-specified alpha, and the harder endpoint (incident cognitive impairment or dementia) was null in every stratum, including that one (total HR 0.90, 95% CI 0.61-1.33) (Lin et al., 2023). So the honest reading is the exposure-vs-intervention split the fabric holds elsewhere: hearing loss is a well-evidenced risk marker, but fitting a hearing aid to prevent dementia is, at best, a reasonable bet concentrated at high baseline risk — not an established general-population intervention -> Hearing Loss and Dementia, Baseline Risk and the Relative-Absolute Split.
The diet and sleep candidates: real signals, routed through the same rocks
A cluster of newer observational findings — ultra-processed food, fruit and vegetables, flavonoids, dietary omega-3, and sleep disorders — each show a statistically real association with dementia risk, and each is weaker than it first looks once checked against the same three questions: does it survive restriction to the hardest subtype (Alzheimer’s), does it survive adjustment for the cardiometabolic factors already counted, and does a dose-response actually show up.
High ultra-processed food intake associates with higher all-cause dementia risk (RR 1.44, 95% CI 1.09-1.90), but the moderate-intake category is null, no dementia subtype individually reaches significance, and the association is lost once adjusted for type 2 diabetes and total energy intake — meaning it runs substantially through levers already on this page’s list (Henney et al., 2023).
High fruit-and-vegetable intake associates 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 studies least prone to reverse causation and strongest in the cross-sectional and case-control studies most prone to it (Zhou et al., 2022). Flavonoids, a component nested inside those same fruits and vegetables rather than an independent exposure, show the identical 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 the same shape from the trial side: Cui pooled 16 RCTs and found the composite cognitive-test score improved (SMD 0.19, 95% CI 0.07-0.32), carried by memory (SMD 0.15, 0.03-0.26), but every trial measured a test score over two years or less and none counted a case of dementia (Cui et al., 2019) -> Soy Products.
Dietary DHA (mainly from oily fish) associates with lower cognitive-decline risk (RR 0.82, 95% CI 0.72-0.93), but omega-3 supplements have been tested directly in a randomized trial (MAPT) and did nothing, alone or added to the multidomain bundle (Wei et al., 2023) — a marker-vs-lever gap, not a reason to buy fish-oil capsules. Sleep disorders track with dementia risk across the board — insomnia (RR 1.13, 95% CI 1.04-1.23), sleep-disordered breathing (1.39 for Alzheimer’s), long sleep over 8 hours (1.66 for Alzheimer’s) — but the long-sleep signal 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) -> Sleep and Cognitive Decline.
None of these candidate levers is one of the established fourteen modifiable factors; each belongs in the reinforce the rocks, don’t add to them category — routes to pull blood pressure, glycaemia, weight and vascular health, not fifteenth, sixteenth and seventeenth independent targets.
Where this leaves the crux
Both halves stand. The multidomain bundle, tested three times against a cognitive test score and once against the diagnosis itself, then pooled and searched for a responder subgroup, has not been shown to prevent dementia — and targeting the people at highest risk does not rescue it. Lowering blood pressure, a single lever already pulled for cardiovascular disease, does prevent it, in a randomized trial, on the hard endpoint, at a small but real absolute magnitude. Neither finding overturns the other, because they are not answers to the same question: one is about a non-decomposable package, the other about one well-understood exposure.
Cancer, heart disease, diabetes — the increments beyond the rocks
Cancer and the cardiometabolic diseases tell the same shared-rock story with fewer surprises — and a shorter list of extras.
Heart disease and diabetes add almost nothing beyond the shared core. Blood pressure, LDL, glycaemia, adiposity, smoking and inactivity carry nearly the whole cardiovascular and type-2-diabetes burden, and that ranking is worked out in Big Rocks (Elderly) and Big Rocks (Median) rather than re-derived here. One framing point belongs on this page instead: sizing the lifestyle rock against the drug landscape. A mature, low-harm drug for a single marker — a statin for LDL, an antihypertensive for blood pressure — captures most of the benefit that marker has to give, which shrinks the marginal lifestyle rock for that outcome specifically. It does not shrink the rest: a drug manages a marker, while lifestyle that removes the underlying driver keeps the structural leverage and the other channels a single-target drug does not touch.
Diabetes prevention is where the head-to-head has actually been run. Lifestyle vs Metformin for Diabetes Prevention randomized 3234 adults with prediabetes to placebo, metformin (850 mg twice daily), or an intensive lifestyle program (7% weight-loss target, 150 minutes/week of activity). Over an average 2.8 years, diabetes incidence fell from 11.0 to 7.8 cases per 100 person-years on metformin (a 31% relative reduction, 95% CI 17-43%) and to 4.8 on lifestyle (58%, 48-66%) (Knowler, 2002). Lifestyle beat the drug outright — 39% lower incidence head-to-head — and needed roughly half as many people treated to prevent one case (NNT 6.9 vs 13.9 over three years). Metformin’s benefit is stratum-specific: near-null in the lean with near-normal fasting glucose (3% reduction, CI -36 to 30) and strongest in the more obese and more hyperglycemic (53%, 36-65); lifestyle worked broadly across every subgroup tested.
Hold the outcome scope honestly, though. Twenty-one years later, the same three arms showed no reduction in hard cardiovascular events from either intervention — metformin HR 1.03 (0.78-1.37), lifestyle HR 1.14 (0.87-1.50), both non-significant, in a cohort where widespread out-of-study statin and antihypertensive use diluted whatever effect either arm still had (Goldberg et al., 2022). Preventing or delaying a diabetes diagnosis is demonstrated; that this converts into fewer heart attacks or a longer life is not shown by this trial, in this population.
Cancer adds genuine disease-specific levers that CVD and T2D do not carry. Physical activity, adiposity and alcohol are already shared levers (see above); cancer’s own increments are about what to eat, and what to eat less of. Body fatness is the largest of them by breadth — WCRF grades it a convincing or probable cause across 12 of the 17 cancer sites in its Continuous Update Project, with IARC independently adding thyroid cancer, multiple myeloma and meningioma (World Cancer Research Fund & American Institute for Cancer Research, 2018).
Against that breadth, red and processed meat are a single-site refinement: processed meat carries a convincing colorectal-cancer association (RR 1.16 per 50 g/day, 95% CI 1.08-1.26), red meat a probable one that is not itself statistically significant when pooled (RR 1.12 per 100 g/day, 1.00-1.25) (World Cancer Research Fund International, 2018). WCRF’s own gram target — up to about 350-500 g cooked red meat a week, very little processed meat — is chosen to balance meat’s micronutrient contribution against that colorectal risk (World Cancer Research Fund & American Institute for Cancer Research, 2018).
NutriRECS re-pooled the same cohort evidence into absolute terms and found that a realistic 3-servings/week cut buys about 7 fewer cancer deaths per 1000 people over a lifetime — low certainty, and null on eight other cancer outcomes (Johnston et al., 2019). Fibre runs the other way, probably protective for colorectal cancer above roughly 30 g/day -> Dietary Fibre and Health. On WCRF’s own grading scale, meat and fibre are both smaller, more site-specific levers than the adiposity rock beneath them -> Body Fatness and Cancer Risk.
Most of the population’s cancer burden sits entirely outside this diet/activity/adiposity list — tobacco, infection (HPV, hepatitis, H. pylori) and occupational exposure carry a comparable or larger share and fall outside this deliverable’s lifestyle scope.
Keeping the body working — frailty, muscle, bone, falls
For the oldest adults the decisive outcome is often not whether a disease arrives but whether the body keeps working — a different lever set. Length and function stay separate axes here, not one score: an intervention can push back a disease’s arrival, preserve mobility, or do neither, and this section keeps the three apart rather than folding them into a single healthspan number.
Frailty is a marker with teeth, and also something you can change. Frailty — reduced physiological reserve across multiple organ systems — predicts a wide set of patient-important outcomes. In Vermeiren’s pooled meta-analysis of 31 studies and 158,764 community-dwelling adults over 65, frail versus robust carried roughly double the hazard of death (HR 1.83, 95% CI 1.68-1.98), of disability in basic activities of daily living (HR 1.62, 1.50-1.76), and a raised hazard of falls (HR 1.24, 1.12-1.37) (Vermeiren et al., 2016). That alone would make frailty only a stratifier.
Racey’s meta-analysis of 26 trials in prefrail and frail older adults closes the gap: physical-activity programs moved mobility (standardized mean difference 0.60), activities of daily living (0.50), and frailty status itself (risk ratio 0.58, moderate certainty) (Racey et al., 2021). The dose has to come down for this stratum — a program built for the general older adult can be too intense once someone is already frail — but the direction holds: frailty identifies who has the most to gain, and training moves the thing identified.
Grip strength and muscle mass are cheap numbers that track both function and mortality — measurements, not targets to chase. In UK Biobank (n=502,293), each 5 kg of lower grip strength carried a 16-20% higher hazard of all-cause mortality, adjusted for age, comorbidity, activity and more (Celis-Morales et al., 2018); a later 48-cohort, 3.14-million-adult pooled analysis that folds in UK Biobank put the weakest third of grip at 1.58 times the all-cause mortality of the strongest third (HR 1.58, 1.40-1.78) (López-Bueno et al., 2022).
Low appendicular muscle mass predicts mortality on its own terms too — people who died carried about 0.18 standard deviations less muscle mass than survivors, an association not fully explained by their strength (de Santana et al., 2021). Both measures are substantially heritable and disease-sensitive, so they are markers as much as levers — no trial shows that raising grip strength or muscle mass, on its own, lowers mortality. Measure them to place someone in a risk stratum; the lever behind them is resistance training and protein.
Resistance training drives muscle and strength; protein is a modest, bounded adjunct. Across 49 RCTs, adding protein to resistance training added a further 2.5 kg to one-repetition-maximum strength and 0.3 kg of lean mass on top of what training alone delivered (Morton et al., 2017). The oft-quoted target sits around 1.62 g of protein per kg of total body weight per day, as actually weighed — not lean mass — and it is a soft, statistically non-significant knee (95% CI 1.03-2.20, p=0.079), better read as a wide region than a precise point: a floor worth reaching rather than a ceiling to fear, since higher intakes show no harm to healthy kidneys.
In the obese the number is not merely lower — it is undefined until the denominator is named, because total and lean weight diverge roughly twofold and the trial sample behind it carried no obese stratum at all. That is a gap, not a finding, and it stays a gap for older and energy-restricted adults too -> Protein and Resistance Training for Muscle and Strength. A separate older-adult maintenance target sits lower, around 1.0-1.2 g/kg/day (1.2-1.5 in illness) — a different objective, on a different evidence base -> Protein Intake for Older Adults.
Falls are prevented by balance, not strength, and the effect is unusually solid. A Cochrane review of 108 trials (23,407 participants, mean age 76) found exercise cut the rate of falls by 23% (rate ratio 0.77, 95% CI 0.71-0.83) at HIGH-certainty GRADE — a rare thing in this domain: an RCT-based effect on a directly observed, patient-important outcome (Sherrington et al., 2019). But the active ingredient is specifically balance and functional training (rate ratio 0.76, HIGH certainty); resistance training alone showed no fall benefit on its own (rate ratio 1.14, 0.67-1.97, very low certainty) — a different mechanism from the muscle and mortality levers above, not a contradiction of them.
Fracture reduction follows downstream, more weakly evidenced (RR 0.73, 0.56-0.95, LOW certainty), because exercise reaches bone through two distinct channels and this is only the larger one: preventing the fall, not measurably strengthening the bone itself, whose own loading effect is smaller and surrogate-only -> Exercise and Bone Mineral Density. And once someone is already frail this signal weakens — the frail-only pooling above found no significant fall reduction (RR 0.80, 0.51-1.26, very low certainty) even while mobility and frailty status still improved, so balance training earns the most confidence started before frailty sets in, not after.
Where the incidence lever inverts: weight loss helps the disease list and can harm the frail. Losing weight is a genuine lever against cancer and cardiometabolic incidence (above) and against knee osteoarthritis symptoms and joint load. But unintentional weight loss is itself one of the criteria that defines frailty in the Fried phenotype, and deliberate weight loss can harm a frail older adult by accelerating the loss of muscle and reserve it is meant to protect. The same exposure that lowers risk in an overweight, disease-incidence stratum becomes a route-(c) contraindication in the frail elderly — the sign inverts by stratum, and a recommendation built for the middle-aged should not be carried over to the oldest-old without checking which stratum it is being applied to -> Shared Modifiable Levers Across Age-Related Diseases.
The shared-lever logic is powerful but not universal — two places break it, and the breaks are instructive.
Where the shared logic breaks — Parkinson’s and the eyes
Two corners of this deliverable follow the shared-lever pattern only partly, or not at all, and both are worth stating precisely rather than folding into the general case.
Parkinson’s is the sharper break. Chen’s 2021 umbrella review pooled 46 systematic reviews across six categories and more than 80 factors, and it grades its own evidence base bluntly: under AMSTAR-2, none of the 46 rated high or moderate quality, and «Seven SRs (15.2%) were judged to be of low methodological quality, while the remaining 39 (84.8%) were judged to be of critically low quality» (Chen et al., 2021). Of 50 statistically significant associations found, only 19 survived to prospective (cohort) evidence; the remaining 31 rest on case-control or mixed designs, more prone to recall and selection bias. That ceiling governs every number below.
Physical activity is the one shared lever that clearly holds: RR 0.79 (95% CI 0.68-0.91, 8 cohort/ nested-case-control studies, I2=0) (Chen et al., 2021), strongest in men (RR 0.68) and at moderate-to-vigorous intensity (RR 0.71), null in women (RR 0.91, 95% CI 0.72-1.14). This figure traces, inside Chen’s umbrella, to Fang et al. 2018’s own meta-analysis of physical activity and PD — a source the fabric does not separately hold. That is a named gap, not an oversight: without the Fang MA itself, the Parkinson’s activity signal holds only at the direction and magnitude Chen reports, and no Parkinson’s-specific dose-response, population-attributable fraction, or absolute-risk figure can be built on top of it.
Diabetes is the other shared lever that holds, modestly: RR 1.31 (95% CI 1.10-1.57, 4 cohort studies) «diabetes significantly increased the risk of PD» (Chen et al., 2021). But three of the cardiometabolic rocks that anchor the shared-core case elsewhere in this deliverable were assessed here and came back non-significant: adiposity — «no significant associations were found between PD development and BMI or serum cholesterol» (Chen et al., 2021) (BMI per 5 kg/m2, RR 1.00, 95% CI 0.89-1.12); blood pressure — «no significant association was found between hypertension and PD (pooled RR: 0.98, 95% CI: 0.82 to 1.17, … 2 cohort studies)» (Chen et al., 2021); and lipids — LDL-C found no significant association either (RR 0.58, 95% CI 0.31-1.07, wide interval, 3 studies).
Pulling the shared rocks therefore does far less for Parkinson’s than for dementia, cancer or heart disease, and does it through a narrower channel — physical activity and, modestly, glycaemic control, not the full cardiometabolic set.
The direction flips outright on smoking and coffee, and the flip is the instructive part. Smoking’s association with lower PD risk (RR 0.64, 95% CI 0.60-0.69) and coffee’s (RR 0.67) and caffeine’s (RR 0.55) are PD’s strongest apparent protective signals — running opposite to every other disease in this deliverable. The review attributes them to reverse causation over Parkinson’s years-to-decades prodrome: early, undiagnosed PD lowers activity, smoking and coffee intake before diagnosis, which a cohort design cannot fully rule out.
Chen’s own reading of the smoking arm is explicit that it changes nothing about smoking policy: «its protective effect for PD development does not impact public health strategies for reducing smoking» (Chen et al., 2021). This is the one place in the whole deliverable where the shared-lever logic would license a genuinely harmful action — smoke to protect the brain — and it is exactly where the signal is an artifact, the case The U-Shaped Association Artifact exists to catch.
The eyes show a milder version of the same lesson: a lever proven for one organ does not automatically transfer to its neighbour. Aune’s 2026 SR-MA of prospective cohorts reaches WCRF probable causality for physical activity and cataract — RR 0.90 (95% CI 0.86-0.94, 10 cohorts, ~1.9 million participants) — but for age-related macular degeneration the same paper finds the evidence «limited, and no conclusion could be drawn»: RR 0.92 (95% CI 0.84-1.01), an interval crossing 1 (Aune et al., 2026).
The authors state the contrast directly: «This meta-analysis provides further support for an inverse association between physical activity and risk of cataract, but an association with age-related macular degeneration was less evident» (Aune et al., 2026). The AMD result is insufficient evidence, not a demonstrated null — the wide interval and thin dose-response data (3 cohorts) cannot exclude a real small effect — but physical activity is not (yet) a proven AMD-specific lever the way it is for cataract.
What does reach AMD is the familiar vascular set, and it adds nothing new to the ranking. Babaker’s 2025 SR-MA of 18 observational studies (~44,440 participants) finds smoking (OR 1.86, 95% CI 1.33-2.6), hypertension (OR 1.24, 1.09-1.4), cardiovascular disease (OR 1.44, 1.11-1.87) and diabetes (OR 1.44, 1.3-1.6) all significantly associated with late AMD, while BMI, cerebrovascular disease and blood lipids were not (Babaker et al., 2025). These are prognostic (route-a) associations, not intervention effects, and they are the same big rocks already ranked on mortality and cardiometabolic grounds — AMD is one more organ carrying the cost of the dominant exposures, smoking above all, not a fresh lever.
A Mediterranean-diet pattern shows a separate, somewhat better-evidenced signal for AMD progression specifically (prospective cohort pool: HR 0.77, 95% CI 0.67-0.88) (Marques-Couto et al., 2025) — but MedDiet is already a big rock on cardiometabolic grounds, so for someone who would adopt the pattern anyway this is confirmatory, not additive.
The lesson both exceptions teach is the same one: “shared” in this deliverable means shared for the diseases and levers it has actually been shown to hold for, never a default assumption. Parkinson’s is a disease where most of the shared core does not transfer; the eyes show that even within one organ system, one lever (physical activity) can transfer to one disease (cataract) and not its neighbour (macular degeneration).
Length, trajectory, and how the ranking shifts by age
The wiki holds no single scalar for “health,” so a lever’s payoff has to be named by which axis it moves: length of life, or the shape of the years before death — function, independence, freedom from pain and cognitive decline. These are not the same currency and should not be collapsed into one number.
Dementia is where this distinction bites hardest. Delaying dementia onset only compresses the years lived with the disease if onset moves back further than life expectancy extends — the compressed-morbidity outcome this deliverable’s outcome menu names. This is a plausible entailment of “pull the levers earlier and keep them low longer,” not a measured finding : none of the trials held here — Peters’ BP-lowering IPD meta-analysis, the FINGER/MAPT/preDIVA/Coley bundle — report a shift in age-at-onset or years-of-dementia-free-life. They report whether dementia was diagnosed by the end of follow-up, an event count, not a trajectory.
Livingston’s Commission states the timing rule plainly — «decrease risk factor levels early (the earlier, the better) and keep them low throughout life (the longer, the better)», and «it is never too early or too late to reduce» dementia risk (Livingston et al., 2024) — but that is a policy recommendation, not a measurement of the trajectory it implies.
The gap runs through the function axis generally: the one SR the fabric holds that tries to measure a composite functional trajectory directly (intrinsic capacity plus quality of life) flags that its own outcome measures are the ones this field measures worst — self-reported, unblindable, Hawthorne-prone — so a positive reading is discounted, not taken at face value Intrinsic Capacity and Multidimensional Healthspan. Naming this gap is itself the decision-relevant move: a clean incidence result should not be read as though it had settled a trajectory question it never asked.
The ranking of levers is not static across age, and the reason is arithmetic more than biology. A 10-year risk window understates the case for acting in a 30- or 40-year-old with an unfavourable but modest risk-factor profile, because a decade captures only the first slice of decades of compounding exposure. ESC states this for cardiovascular risk directly: «The 10-year CVD risk in relatively young, apparently healthy people is on average low, even in the presence of high risk factor levels, but the lifetime CVD risk is in these circumstances very high» (European Society of Cardiology, 2021). The logic transports to dementia and the other diseases in this deliverable — midlife exposure to the shared core (LDL, blood pressure, obesity, smoking, diabetes) drives late-life incidence across diseases, so a young adult’s low 10-year single-disease number understates a lever that pays into several late-life outcomes at once.
What does not transport is the apparatus: ESC’s LIFE-CVD model and its CVD-free-years currency are calibrated for cardiovascular disease only, and the fabric holds no equivalent lifetime model for dementia or Parkinson’s — so “act early” is warranted qualitatively here, but no cross-disease lifetime-years figure can be quoted . ESC also bounds its own frame: below 40, «CVD risk predictions, as well as predictions of lifetime benefit of risk factor treatment, are likely to be imprecise», and drug treatment is not usually considered at that age — lifestyle carries the weight instead (European Society of Cardiology, 2021).
At the other end of life the ranking shifts again, mostly toward function. In a lean, physically active, non-smoking elderly person the cardiometabolic big rocks are already largely pulled, so the marginal levers move toward resistance training and adequate protein, balance-and-strength work against falls, and frailty status itself — covered in full above and in Big Rocks (Elderly). Baseline risk does real work here without needing any subgroup claim: the same relative effect buys more absolute benefit in a higher-baseline-risk stratum. That is part of why Peters’ BP-lowering dementia result (OR 0.87, detailed above) was small in absolute terms in the trial population as enrolled, and would plausibly buy more in a population starting from a higher baseline dementia risk Baseline Risk and the Relative-Absolute Split.
The same logic runs the other way in preDIVA, which enrolled an unselected older population already receiving good background vascular care and found no reduction in incident dementia over 6.7 years (HR 0.92, 95% CI 0.71-1.19, p=0.54). The trial’s own reading names the headroom problem directly: «This absence of effect might have been caused by modest baseline cardiovascular risks and high standards of usual care» (van Charante et al., 2016) — a route-(a) point in miniature: a lever layered on top of an already-modest baseline risk and already-good usual care has little absolute room left to work with, whatever the relative logic says.
One structural limit closes the ranking question. The 14-factor dementia population-attributable-
fraction map, WCRF’s preventable-fraction estimates, and every other PAF-style figure in this
deliverable are population quantities, not per-person ones, and the individual factors’ PAFs overlap
and therefore do not sum. Removing a modifiable exposure does not remove that share of one person’s own
risk, and no single aggregate cross-disease burden figure can be built by adding PAFs across the
diseases covered here — a structural limit (G, needs aggregation the fabric cannot perform), named
rather than computed.
What does not move the needle much
That leaves the exposures people spend the most attention on — and where attention runs inverse to effect. Several diet-detail and lifestyle candidates get argued about at length; held against the cardiometabolic rocks above, most turn out to move a little, act through those same rocks rather than adding a new one, or not to move at all. A confirmed small effect is as reportable a result as a confirmed large one — an honest map of this territory is not “everything checks out,” and it is not “nothing here matters” either.
Three diet-detail candidates each carry a small, low-certainty cognitive signal that tracks the cardiometabolic rocks rather than sitting beside them as a 15th factor. None reaches a randomized trial, and by the authors’ own account none plausibly ever will, since withholding a whole diet pattern from people for years to test it is not something an ethics board would approve (Henney et al., 2023).
Ultra-processed food, fruit and vegetables, flavonoids
Ultra-processed food. A pooled analysis of ten observational studies found high intake associated with higher all-cause dementia risk (RR 1.44, 95% CI 1.09-1.90), but moderate intake was not significant, no dose-response was demonstrated, every individual dementia subtype (Alzheimer’s, vascular, MCI) fell short of significance on its own, and the signal disappeared once studies adjusting for diabetes or total energy intake were isolated (Henney et al., 2023) — a pattern consistent with the association running through diet quality and metabolic disease already counted, not a processing-specific brain effect. A narrower review of broader cognitive outcomes reached the same place by a largely overlapping evidence base, so it refines rather than independently confirms the finding (Smith et al., 2026).
Fruit and vegetables. Intake shows a similar-sized inverse association with cognitive disorders (OR 0.82, 95% CI 0.75-0.90) that is null for Alzheimer’s disease specifically (0.88, 0.76-1.01) and weakens rather than strengthens in the prospective-cohort studies least vulnerable to reverse causation — the opposite of what a genuine causal signal should do (Zhou et al., 2022).
Flavonoids. Intake, measured the same way (self-reported food intake run through a food-composition database), shows a borderline overall association (OR 0.90, 95% CI 0.83-0.98) carried entirely by the softest endpoint, “cognitive decline” (0.88, 0.79-0.98), while dementia (0.97, 0.79-1.19) and Alzheimer’s disease (0.90, 0.69-1.17) are both null (Peng et al., 2026). Because the flavonoid figure is computed from the same food-frequency questionnaires used to estimate fruit and vegetable intake, it is not a second, independent line of evidence for a brain benefit — it is the same signal re-expressed in different units, unable to say whether any benefit belongs to the flavonoid, the whole food, or the diet pattern surrounding it Is the Food Category Doing Any Work.
Fish-oil pills do not deliver what fish-eating appears to. Habitual dietary DHA intake carries the strongest observational signal in this space (RR 0.82, 95% CI 0.72-0.93, for dementia), but that is a food-and-years-of-diet exposure, not a supplement dose. The randomized trials that tested DHA and EPA as capsules found no cognitive benefit Multidomain Lifestyle Intervention and Cognitive Decline — the same diet-versus-pill split that shows up across this deliverable. Buying dietary omega-3 in a bottle is not a substitute for the eating pattern the observational signal was actually measuring.
Sleep problems mostly carry a modest signal, and the one large exception is likely the disease talking, not causing. Across eight types of sleep disturbance and four cognitive outcomes in a pooled review of 76 cohort studies, most associations sit in the RR 1.1-1.4 range, and one — restless, movement-disturbed sleep — reaches a notably larger association with vascular dementia (RR 2.53, 95% CI 1.30-4.93) (Zhang et al., 2025).
Sleeping more than eight hours a night carries the largest dementia association in the whole map (RR 1.66 for Alzheimer’s disease), but this arm concentrates in people over 70, correlates with markers of pre-existing brain pathology, and lacks any check for reverse causation — a textbook case for treating a protective- or harmful-looking association as an artifact until it survives a genetic or referent-corrected check The U-Shaped Association Artifact. There is no evidence here that advising a healthy long sleeper to sleep less would help.
The pattern is not uniformly a null, though: insomnia’s association gets larger and more consistent, not smaller, when measured objectively rather than by self-report — the opposite of what a pure reporting artifact should do — which argues that at least this one arm is a real, if modest, signal rather than a measurement ghost.
Which type of exercise a person does matters much less than whether they exercise. An umbrella review of exercise-cognition studies reports mind-body exercise (tai chi, yoga, Pilates) as carrying the largest cognitive effect of any exercise type. That comparison, though, pools separate studies against their own, mostly passive, controls rather than testing mind-body exercise head-to-head against another active routine, and a related, larger umbrella found the same kind of effect shrank to near nothing once publication bias was corrected for (Blomstrand et al., 2023) Mind-Body Exercise and Cognition. Someone already doing aerobic or resistance exercise has no established cognitive reason to add tai chi or yoga on top of it; either may be reasonably chosen for other, separately-evidenced reasons — balance, joints, stress — but not this one.
Air pollution carries a real but small dementia signal, and it is barely something a person controls. A systematic pooling of the primary studies puts the association at a hazard ratio of 1.04 (95% CI 0.99-1.09) per 2 micrograms per cubic meter of fine particulate matter — a confidence interval that just crosses the null, on an effect smaller than the ones reported for education or smoking (Wilker et al., 2023). Most of the studies driving that estimate carried a risk of bias that, if anything, pulled the number toward zero — so the true effect is probably a small positive one, not nothing. But the exposure sits mostly outside individual control: filtration and choice of neighborhood or commute route offer only a thin personal margin, and the bulk of the lever lives at the level of clean-air policy, not personal choice.
Three things the evidence cannot yet tell us.
What the evidence still cannot tell us
Parkinson’s disease has a physical-activity signal, but not yet a sized one. The activity association for Parkinson’s rests on a 2021 umbrella review of the field rather than a purpose-built dose-response meta-analysis (the derivation is in the Parkinson’s section above) (Chen et al., 2021). The direction and reported magnitude hold; a Parkinson’s-specific dose target, absolute-risk figure, or population-attributable fraction cannot be built from evidence never designed to supply one.
The biology that might explain why one lever touches several diseases at once is not evidence this fabric holds. The pattern itself — that the same handful of exposures recurs across dementia, cardiovascular disease, cancer and physical function — is well established across the disease-specific sources above. What is absent is the aging-biology literature (cellular senescence, chronic inflammation, and related mechanisms sometimes grouped as “the hallmarks of aging”) that would explain that pattern mechanistically. Treat any claim about why the levers overlap at that biological level as unstated, not as a confirmed backbone underneath the pattern; the overlap remains actionable without it.
How much total benefit one lever buys, added up across every disease it touches, cannot be computed from what is held here. Each disease’s effect estimate for a given lever comes from a different population, follow-up length and outcome definition; summing or averaging them across dementia, cancer, heart disease and falls would manufacture a combined number no study measured. That is a structural limit on what a review of separate studies can produce, not a gap a bigger meta-analysis eventually closes — closing it would need a single study designed to track the combined outcome directly, and none exists.
What to do
Most people chasing better odds against dementia, cancer and heart disease do not need a different diet or routine for each disease. A short, shared list of changes lowers the risk of several of them at once: stop smoking, keep blood pressure and cholesterol down, manage blood sugar, carry less body fat, stay physically active, and drink less alcohol. Staying active is the single most useful item on that list, because it is the one lever that also keeps the body working day to day — muscle, bone, balance — not only diseases at bay over decades. Nothing sold as a substitute matches it.
For most of these levers the individual payoff is modest, and most of the evidence behind them is observational rather than proof from a trial that changing the behavior changes the disease. The clearest exception is blood pressure: lowering it in a randomized trial reduced the number of people who went on to develop dementia. The benefit to any one person was small, but it was real, and it arrived as a bonus on a treatment already worth doing for the heart — not from a separate brain-health regimen.
The reverse bet has not paid off. A packaged program of brain games, diet coaching and supplements, sold specifically to prevent dementia, can move a cognitive test score in some people, but it has not been shown to keep people from developing the disease — including when the program targets people already at higher risk.
The details many people spend the most energy on — cutting packaged food, adding more fruit and vegetables, chasing flavonoid-rich foods or fish-oil pills, fixing sleep — mostly work, to the extent they work at all, by nudging the same handful of levers already listed, not by adding a new one. They are worth doing as part of a generally sound diet and routine. They are not worth optimizing on their own for brain health.
The ranking shifts with age. A younger adult with low near-term risk still benefits from acting early, because several of these diseases develop over decades and the levers that prevent them work best pulled sooner rather than later. Older adults gain the most from a different set — resistance exercise, adequate protein, and balance training that protect muscle, bone and the ability to avoid a fall. One caution runs the other way: losing weight helps prevent several of these diseases in someone carrying excess fat, but it can harm an already-frail older adult, so it is not a universal recommendation late in life.
None of this adds up to a single number worth optimizing. It adds up to a short list, pulled in the right order, and adjusted as a person ages.
Evidence box
Question ’For an adult deciding how to lower their risk of the diseases of ageing — cardiovascular disease, cancer, dementia, Parkinson”s, frailty/sarcopenia, osteoporosis: which modifiable exposures move which disease, how much, how certainly, do the diseases share a common core of levers or need separate regimens, and how does the ranking shift by age stratum?‘ Evidence included 37 sources — 25 gold, 12 high Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note All sources are gold or high tier. Last updated 2026-09-05 · Independently reviewed: No · Full edit history