Foods to limit or avoid
In modern populations, what you take out of your diet usually moves the dial more than what you put in. Ultra-processed food, sugar-sweetened drinks, processed meat, and excess alcohol carry the clearest harm signals in modern nutrition science — and they are the easiest changes to make.
The reason removal matters so much is that the modern Western diet has accumulated specific harms that no amount of "extra spinach" can offset. Ultra-processed food displaces both calories and bioavailable nutrients. Sugar-sweetened beverages are the single most consistent dietary input on incident cardiovascular disease and type 2 diabetes. Processed meat is a classified human carcinogen. The interventions that produce most of the available longevity dividend cluster around removing or substantially reducing these categories, not adding exotic ones. This article goes through them in order of evidence weight, with the honest summary on the contested debates (seed oils, cheese, saturated fat) at the end.
Ultra-processed food
Evidence rating: Moderate (large, consistent observational dose-response plus supporting controlled-feeding trials; the cohort core carries real residual confounding).
The strongest single dose-response signal in modern nutrition epidemiology. A 2024 umbrella review of 45 pooled analyses — about 9.9 million participants summed across them, with some cohorts counted more than once — linked high ultra-processed-food (UPF) intake to 32 distinct adverse outcomes spanning cardiovascular, metabolic, cancer, and mental-health domains.[1] The headline dose-response is genuinely contested: each 10% rise in UPF as a share of energy intake associates with somewhere between about 3% and 10% higher all-cause mortality — about 3% in one 2025 analysis (RR 1.03),[2] and about 10% in another, with very high disagreement between the studies it pooled (HR 1.10).[3] Both are per-increment estimates.
The strongest causal anchor is controlled feeding. In the 2019 NIH metabolic-ward trial, adults on an ultra-processed diet involuntarily ate about 500 kcal/day more — and gained weight — than on a minimally processed diet matched for calories, sugar, fat, fibre, and macronutrients, despite eating as much as they liked on both.[4] A 2025 randomized crossover at University College London, published in Nature Medicine, pointed the same way — 55 adults with overweight or obesity (50 analysed, 91% women) lost modestly more weight over eight weeks on a minimally processed diet than on an ultra-processed one, both built to the UK Eatwell Guide (Kevin Hall, who led the NIH trial, is a co-author, so the two are not fully independent) — though the result is genuinely contested: three Nature Medicine commentaries noted the two arms were not matched on energy density (1.60 vs 1.25 kcal/g on the trial's own corrected figures — about a quarter more, not double), that the ~1 kg difference is of limited clinical importance, and that several trial authors have obesity-drug industry ties.[5] The mechanism is plural: a faster eating rate that outpaces satiety signalling, high energy density, gut-barrier disruption from some emulsifiers and additives, and missing phytochemicals from intact food structures.
"UPF" is not monolithic. In a pooled analysis of two large US cohorts (the Nurses' Health Study and Health Professionals Follow-up Study, ~114,000 adults, median ~34-year follow-up), the highest UPF intake carried about 4% higher all-cause mortality (HR 1.04) — driven by causes other than cancer and cardiovascular disease, for which no association appeared — and the signal was concentrated in processed meat (HR 1.13 for all-cause mortality, rising to 1.43 for some specific causes of death) and sugar-sweetened and artificially sweetened drinks, while ultra-processed breads and breakfast foods carried a smaller but statistically clear excess (hazard ratio 1.04) — no sub-category in this cohort was protective for all-cause mortality.[6] Within bands of overall diet quality the UPF association in these cohorts largely vanished — but a 2026 analysis of nearly 48,000 US adults, adjusting for a broader diet-quality index, found it only partly attenuated and still present.[7] Whether the NOVA label adds anything beyond the company an ultra-processed food keeps is live rather than settled. UPF intake also tracks with accelerated biological aging: each 10% rise in UPF energy associated with about +2.5 months of accelerated biological aging on PhenoAge — a score built from chronological age plus nine routine blood measurements, not a DNA-methylation clock — in the US National Health and Nutrition Examination Survey (NHANES), falling to about +1.7 months once overall diet quality is accounted for, in a cross-sectional analysis.[8] Because the cohort data are observational and UPF intake correlates with smoking, lower income, and higher body-mass index, residual confounding is real — but the controlled-feeding trials above (the NIH metabolic-ward study, and more tentatively the contested UCL crossover) point to a processing-and-energy-intake pathway that confounding alone does not easily explain.
See Ultra-processed food for the full article — NOVA classification, mechanism, the UCL trial, biological-age effects, and a practical guide to spotting UPF on a label.
Rules of thumb:
- Aim for under 10–20% of daily calories from UPF — a practical target rather than a guideline figure, since no authority has set a numeric threshold. The average US adult is at about 53%, and falling slightly;[9] the UK is similar, while in Mediterranean Europe the figure is far lower.[10]
- Cook from whole or minimally processed ingredients as the default.
- The "kitchen test" on ingredient lists: if the list contains even one substance never or rarely used in home kitchens (modified starches, protein isolates, hydrogenated fats, dough conditioners) or additives whose job is to make the product look or taste better (emulsifiers, flavour enhancers, colours, sweeteners, thickeners), the product is ultra-processed.[11]
Sugar-sweetened beverages
Evidence rating: Moderate (consistent large cohorts with a coherent liquid-fructose mechanism; no randomized hard-outcome trials).
The largest single-food-category mortality signal among beverages.
- Each daily 355 mL (12 oz) serving of sugar-sweetened beverage associates with roughly 7% higher all-cause mortality and 10% higher cardiovascular mortality in two large US cohorts pooled together[12] — figures a 2025 umbrella review of the wider literature closely reproduces.[13]
- One well-known finding is routinely misreported and worth stating correctly: in the Framingham Offspring cohort, the roughly threefold higher rate of incident ischemic stroke over a decade attached to daily artificially sweetened soft drinks, not to sugar-sweetened ones, which showed no association with stroke or dementia. The dementia half of that paper did not survive full adjustment, and a later pooled analysis of six cohorts found no association between artificially sweetened drinks and dementia at all.[14][15] The case against sugared drinks rests on the mortality and cardiometabolic data above, not on this study; the diet-drink signal is covered under Sugar substitutes.
- Soft-drink consumption shows a clear dose-response for incident fatty liver disease (non-alcoholic fatty liver disease, NAFLD, diagnosed by ultrasound — the condition now usually called MASLD): in the Tianjin (TCLSIH) cohort of Chinese adults followed for a median of about four years, risk rose from about 18% higher at one serving/week to roughly 47% higher at ≥4 servings/week (HR 1.47) — consistent with the fructose-driven hepatic lipogenesis mechanism above.[16]
- The category includes sodas, sweetened iced tea, energy drinks, sweetened coffee drinks, fruit juices, and sports drinks (in non-athletes). Fruit juice belongs on the list even at 100%: one pooled analysis put it at about 26% higher all-cause mortality, at least as high as for sugar-sweetened drinks (though only two cohorts fed the juice figure, and a larger Harvard analysis found no juice–mortality link[17]), though in the same data it was not linked to diabetes, high blood pressure or stroke — where sugar-sweetened drinks tracked with about 27% higher type 2 diabetes risk (RR 1.27, 95% CI 1.17–1.38).[18]
- The World Health Organization (WHO) recommends keeping free sugars under 10% of total energy (strong recommendation) and ideally under 5% (~25 g/day for an average adult; conditional); the American Heart Association (AHA)'s daily caps for added sugar are 25 g for women and 36 g for men — derived from a share of a discretionary-calorie allowance rather than from a dose-response threshold.[19] The 2025–2030 US Dietary Guidelines dropped the daily percentage ceiling altogether, stating that "no amount of added sugars or non-nutritive sweeteners is recommended or considered part of a healthy or nutritious diet" and capping any single meal at 10 g of added sugars.[20] The cardiometabolic case rests mainly on the beverage cohort and trial data above; the threshold recommendations themselves are graded moderate-to-very-low.
The mechanism isn't only the sugar load — it's the liquid delivery. Liquid fructose bypasses satiety regulation and arrives at the liver fast enough to overwhelm normal handling, driving hepatic lipogenesis. The same fructose dose in whole fruit, with its fibre matrix slowing absorption, doesn't produce the same metabolic insult — in fact, whole fruit tracks with lower risk where sugar-sweetened drinks track with higher: a separate meta-analysis pooling 27 cohort studies found about 15% lower all-cause mortality per 200 g of fruit a day (RR 0.85).[21]
Diet sodas and other non-sugar sweeteners. Less harmful than the full-sugar versions in the short term, but not biologically inert. The ELSA-Brasil cohort (12,772 adults, ~8 years) found the highest intake of low- and no-calorie sweeteners associated with faster declines in global cognition, memory, and verbal fluency — effects that are real but very small, only in adults under 60, and in a paper corrected twice in under a year.[22] The detail is in Sugar substitutes. Erythritol and xylitol — the dominant sugar alcohols in "keto" and "sugar-free" products — show higher three-year rates of major cardiovascular events at high plasma levels in cohorts of patients undergoing elective cardiac evaluation in the US (Cleveland Clinic) and, for erythritol, Europe (Berlin), with a platelet-reactivity mechanism demonstrated in vitro. The excess was roughly double for erythritol (top versus bottom quarter, HR 1.80 and 2.21 in two separate cohorts) and about 60% for xylitol (top versus bottom third, HR 1.57).[23][24] Plain water, unsweetened tea, sparkling water, and unsweetened coffee are better defaults. See Sugar substitutes for the full discussion.
Processed meat
Evidence rating: Strong for the colorectal-cancer classification (IARC Group 1); Moderate for the cardiometabolic and dementia associations.
Classified by the WHO's cancer agency (the International Agency for Research on Cancer, IARC) as a known human carcinogen (Group 1) — the same category as tobacco and asbestos, though the absolute risk is much smaller.
- Definition: meat preserved by smoking, curing, salting, fermenting, or with added chemical preservatives. Bacon, sausages, hot dogs, ham, salami, deli meats, jerky.
- Each 50 g per day of processed meat associates with an 18% increase in colorectal cancer risk.[25] In absolute terms that moves a lifetime risk of about 3.9% for US adults to roughly 4.6% — the calculation is set out in Red and processed meat.
- About 15% higher type 2 diabetes risk per 50 g/day of processed meat (HR 1.15, 95% CI 1.11–1.20 — the bracketed range is where the true effect most plausibly lies, and because it sits entirely above 1.0 the increase is unlikely to be chance), in a federated meta-analysis of 1.97 million adults across 31 cohorts in 20 countries; the same analysis found about 10% higher risk per 100 g/day of unprocessed red meat (HR 1.10, 1.06–1.15).[26] Per daily serving, processed meat carries about 15% higher all-cause mortality (RR 1.15, 95% CI 1.11–1.19), 15% higher cardiovascular mortality (1.15, 1.07–1.24) and 8% higher cancer mortality (1.08, 1.06–1.11), on curves that are not straight lines; the same analysis found no significant mortality association for unprocessed red meat outside US cohorts, and inside them the two ran at much the same rate.[27] The clearest processed-versus-unprocessed asymmetry comes from the larger cardiovascular pooling, where processed meat carries about 26% higher risk per 50 g/day against 11% per 100 g/day for unprocessed.[28]
- Dementia signal: a 2025 Harvard cohort analysis in Neurology of over 133,000 adults found that eating a quarter-serving a day or more of processed red meat, compared with under a tenth of a serving, associated with 13% higher dementia incidence (HR 1.13, 1.08–1.19) and 14% more self-reported cognitive decline (RR 1.14, 1.04–1.25); each additional daily serving correlated to ~1.6 years of accelerated cognitive aging.[29]
- Processed meat carries roughly 4× the sodium of unprocessed red meat per serving and about 3× the pre-formed nitrosamines, with nitrite modestly higher and nitrate only about 40% higher (4.6 versus 3.3 mg per 50 g) — and it is the sodium alone that epidemiologically explains about two-thirds of the differential coronary-heart-disease risk between processed and unprocessed forms (a modelled estimate). Haem iron and cholesterol are, if anything, lower in processed meat, which is what rules them out as the explanation.[30]
- The carcinogens implicated include N-nitroso compounds (from nitrite preservatives), heterocyclic amines (from high-heat cooking), and polycyclic aromatic hydrocarbons (from smoking). Sodium load contributes separately.
Practical: treat processed meat as an occasional indulgence, not a regular diet item. The World Cancer Research Fund (WCRF) guidance is to consume "very little, if any": there is no established safe threshold, and WCRF states that no level of processed-meat intake is associated with an absence of colorectal-cancer risk.[31]
See Red and processed meat for the full article — the processed-vs-unprocessed distinction, the NutriRECS controversy, the Neu5Gc xenosialitis and TMAO mechanisms, cooking-method mitigation, substitution analysis, and the APOE4 paradox.
Unprocessed red meat
Evidence rating: Moderate — the dose-response is real but weaker than for processed meat, population-dependent (present in US cohorts, often absent in European and Asian ones), and Mendelian-randomization studies — which use genetic variation as a proxy for lifelong intake — have so far found no causal link between unprocessed red meat and heart disease, stroke or type 2 diabetes.[32]
Classified by IARC as probably carcinogenic to humans (Group 2A) — weaker evidence than processed meat but still consistent.[33]
- Each 100 g/day of unprocessed red meat associates with about 11% higher cardiovascular disease (CVD) risk (HR 1.11, 1.05–1.16) in a 2023 meta-analysis of 43 studies pooling ~4.4 million adults,[34] and about 10% higher type 2 diabetes risk per 100 g/day in the 1.97-million-adult federated analysis above.[35] For colorectal cancer, a 2025 meta-analysis of 60 prospective studies put red meat at about 15% higher risk (HR 1.15, 1.10–1.21), against about 21% for processed meat (HR 1.21) — though for colon and rectal cancer taken separately, the paper's own estimates run slightly higher for red meat than for processed.[36] So the processed-versus-unprocessed asymmetry holds for cardiovascular disease and for colorectal cancer overall, not at every site — see the deep-dive for the full picture.
- The cardiovascular signal is partially confounded by saturated fat content, but two distinct human-specific mechanisms operate independently of lipids: Neu5Gc (a non-human sialic acid that triggers chronic low-grade autoimmune-like inflammation, "xenosialitis") and TMAO (a pro-atherogenic metabolite produced when gut microbes convert meat-derived carnitine and choline). See Red and processed meat for the mechanistic detail.
- The NutriRECS controversy (a 2019 five-paper series in Annals of Internal Medicine) argued the underlying evidence is weaker than dietary guidelines suggest.[37] The methodological argument was about applying GRADE to observational nutrition — a framework that structurally guarantees a "low-certainty" verdict in this field. The NutriRECS series' own analysis of dietary patterns still found lower risk with less meat — 13% lower total mortality, 14% lower cardiovascular mortality, 11% lower cancer mortality and 24% lower type 2 diabetes risk, as reported by the Harvard Chan critique of the series.[38] Its cohort analysis, which looked at meat intake directly rather than at whole patterns, described the same direction as "a very small reduction in risk" per three servings a week cut.[39] The Harvard Chan School pushed back hard, and IARC joined thirteen other organisations in a joint statement coordinated by the World Cancer Research Fund rejecting the recommendation.[40] The 2022 Nature Medicine "Burden of Proof" re-analysis of unprocessed red meat dropped the usual assumption that risk rises in a straight line, fitting a flexible curve instead, then graded each association by the most cautious edge of its uncertainty range. On that grading it returned "weak evidence" for colorectal cancer, breast cancer, ischaemic heart disease and type 2 diabetes, and no evidence for either kind of stroke.[41] Its 2025 companion did the same for processed meat, putting the conservative floor at 11% higher type 2 diabetes risk and 7% higher colorectal cancer risk against zero consumption — again "weak" ratings, which the authors nonetheless say justify continuing to recommend limiting intake.[42]
Practical: up to 3 servings per week is a defensible cap — the WCRF's ceiling, which it puts at about 350–500 g cooked weight per week; the upper end, 500 g cooked, is roughly 700–750 g raw. One serving is about 120–170 g cooked. Substitute thoughtfully — the meta-analysis of substitution found lower coronary-heart-disease risk when red meat is replaced by nuts or legumes (both about 16% lower), eggs (14%), poultry (12%) or dairy (10%); swapping in fish was not associated with lower risk, and refined carbohydrates were never tested as a substitute.[43] If you eat red meat, choose lean cuts; cook gently (stew, braise, slow-roast) rather than charring; and co-consume with fibre — a 13-person pilot trial found fibre reduced the abundance of the microbial cutC gene, but did not change the post-beef TMAO response overall, only in the participants who ate meat less than three times a week.[44] The Protein article covers the methionine-balance argument; the full article above covers everything else.
Cheese
Evidence rating: Weak / observational — the stroke-neutral finding rests on cohort data and the lipid evidence on short trials, while the brain-specific concern is an inference rather than a measured cheese effect.
The MIND diet (the Mediterranean–DASH Intervention for Neurodegenerative Delay — a hybrid of the Mediterranean and DASH, or Dietary Approaches to Stop Hypertension, patterns, emphasizing greens, berries, and fish) recommends less than one serving of cheese per week. The wider picture is more nuanced.
- The food-matrix argument: Pooled cohort data show no excess stroke risk with cheese and, above about 25 g a day, a modest inverse association (RR 0.91) — though the flat overall estimate is the more cautious reading, and the authors conclude the mechanism is not yet clear.[45] Randomised trials point the same way on lipids: 135 g a day of hard or semi-hard cheese lowered low-density lipoprotein (LDL) cholesterol compared with the same amount of fat eaten as butter, which is the best direct evidence that the intact lipid-protein matrix blunts what would otherwise be a saturated-fat-driven rise.[46]
- The MIND-specific concern is about the brain, not the heart. The original MIND instrument placed cheese in the "limit to less than one serving per week" list, but not because the cohort found a cheese-specific cognitive signal — the components were chosen "based on comprehensive reviews of the literature on nutrition and the aging brain", and cheese appears in that paper only as a scored item.[47] The saturated-fat-and-inflammation rationale is inference rather than a finding of that cohort.
Reasonable interpretation:
- Modest cheese intake (1–3 servings/week) is compatible with Mediterranean and most longevity-oriented patterns.
- Hard and semi-hard cheeses (parmesan, gouda, gruyère) are what the lipid trials tested; treating processed cheese products as worse is reasonable by analogy, but no study has compared them directly.
- The MIND <1/week guidance is brain-specific and based on observational data — it's not a hard rule against cheese, and a healthy adult eating modest cheese in a Mediterranean pattern is not contradicting the evidence.
- Cheese in volume daily — particularly highly processed cheeses with added oils and salt — is genuinely worth limiting.
Industrial seed oils — the contested debate
Evidence rating: Moderate — biomarker meta-analyses, RCTs, and the Cochrane fat reviews consistently fail to show harm from linoleic acid at dietary intakes, and point the other way.
The "seed oils are toxic" framing has become popular online. The actual evidence is more boring than either camp suggests.
What's true:
- Industrial deep-frying produces oxidized lipid species; the case that they are pro-inflammatory comes mainly from cell and animal work.
- Repeatedly heated oils (commercial fryers running hours daily) are the plausible real exposure issue. In the one small human test, a meal made with fat from a commercial fryer cut blood-vessel dilation from 5.9% to 0.8% within hours — but that was ten men, a single meal, and a measure of blood-vessel function rather than inflammation; no long-term human trial exists.[48]
- Refined oils are calorie-dense and easy to over-consume specifically because they're hidden inside ultra-processed foods.
What's overstated:
- Linoleic acid (the dominant omega-6 in seed oils) at typical dietary intakes is often called "inflammatory", but the cardiovascular data below point the other way (that biomarker pooling measured no inflammatory markers itself). Higher measured linoleic acid in blood or fat tissue tracks with lower cardiovascular risk, not neutral — comparing the top and bottom fifths of the biomarker, cardiovascular mortality was about 22% lower (HR 0.78) across 30 cohorts and 68,659 people, with the authors concluding the data "support a favorable role for LA in CVD prevention". Arachidonic acid, the omega-6 most often named as the inflammatory one, was not associated with harm either.[49]
- A 2025 review pooling 150 cohort and nested case-control publications found that measured omega-6 in the blood tracks with lower cardiovascular disease (RR 0.92) and lower cancer risk (0.88), and that reported dietary intake tracks with lower all-cause mortality (0.88). The dietary estimates for cardiovascular disease and cancer overall were null rather than protective, and one signal ran the other way — a 15% higher risk of endometrial and ovarian cancer — so the overall picture is lower mortality with more omega-6, from observational data that cannot prove benefit, and no sign of the harm the seed-oil critique predicts.[50]
- Decades of pooled cohort data and the Cochrane fat-and-cardiovascular reviews point the other way from the panic: cutting saturated fat reduced cardiovascular events by about 17% (RR 0.83, meaning 17% fewer heart attacks, strokes and revascularisations), with the benefit coming from replacing it with polyunsaturated fat or starchy foods — though death rates themselves barely moved.[51]
Practical synthesis:
- Olive oil, especially extra-virgin (EVOO): best evidence; the preferred default for cooking and dressing.
- Avocado oil and nut oils: fine.
- Canola, sunflower, safflower: neutral; not "toxic" in normal dietary use.
- Avoid the oils used in repeated commercial deep-frying (typical fast-food chains, not your home pan).
The headline isn't "avoid seed oils." It's "avoid ultra-processed foods that contain seed oils alongside refined sugar, refined flour, and additives." The ultra-processed context is what's harmful — the oil molecule by itself is mostly innocent. See Dietary fats for the full lipid-quality picture — trans fats, saturated fat substitution dynamics, PUFAs, MUFAs, eggs, and the omega-6:omega-3 ratio.
Coconut oil is the mirror-image myth — marketed as a health food rather than a toxin. A meta-analysis of clinical trials found coconut oil raises LDL cholesterol by about 10.5 mg/dL (0.27 mmol/L) relative to non-tropical vegetable oils — an average over trials that disagreed considerably (95% CI 3.0–17.9 mg/dL) — and HDL cholesterol by about 4 mg/dL, without any weight or glycemic benefit.[52] The net message is the same even-handed one applied to seed oils, inverted: coconut oil is not a "health food," but it is not uniquely toxic either — use it sparingly and prefer olive oil as the default.
Trans fats
Evidence rating: Strong — consistent cohort data, a clear mechanism, and confirmatory population outcomes after industrial-trans-fat bans.
Industrial trans fats are unambiguously harmful. Each 1% increase in trans-fat energy associates with roughly 6% higher all-cause mortality and 6% higher cardiovascular mortality;[53] the structural geometry of the trans double bond is biologically alien and disrupts endothelial function and lipid homeostasis; in cultured human monocytes and in mice it also shifts global DNA methylation, though no human feeding study has tested a methylation endpoint.[54]
- Between 1990 and 2021, the age-standardised mortality attributable to high trans-fat intake fell by about 69% among adults aged 55 and over (15.5 to 4.8 deaths per 100,000), with the steepest falls in the most developed (high socio-demographic index) regions, where industrial trans fat had been banned.[55]
- As of WHO's 2024 progress report, 53 countries had best-practice trans-fat policies covering 3.7 billion people (46% of the world's population), expected to save ~183,000 lives per year — up from 6% coverage in 2018 — and more than 4 billion people remain unprotected.[56] WHO's current fact sheet attributes more than 278,000 deaths a year worldwide to industrially produced trans fat.[57]
- Mostly eliminated from the food supply in the European Union and the United States since the FDA's 2015 final determination on partially hydrogenated oils (no partially hydrogenated oils could be added to food after June 2018, with the last extended compliance dates running to January 2021) and the EU's 2019/649 regulation (capping trans fat other than that naturally present in animal fat at 2 g per 100 g of fat, not of food, from 2021).
- Still present in some processed foods in countries without similar regulation; check labels for "partially hydrogenated oil" in non-regulated markets.
- Naturally occurring (ruminant) trans fats in dairy and meat — chiefly vaccenic acid and conjugated linoleic acid — are biologically distinct from the industrial form and show no association with coronary heart disease in pooled cohorts at the intakes achievable from dairy and meat — though the estimates are wide enough that this is no evidence of harm rather than proof of none.[58]
Excessive sodium
Evidence rating: Strong for blood pressure (randomized DASH-Sodium data); Moderate for cardiovascular and gastric-cancer outcomes.
- Population-level sodium excess, mostly from processed and restaurant food rather than the salt shaker, drives hypertension.
- The DASH-Sodium trial randomised 412 adults across three sodium levels on two diets. Cutting sodium from high to low lowered systolic blood pressure by about 7 mm Hg across the whole trial on a typical US diet — a substantial fraction of what a single first-line antihypertensive delivers. Combining that cut with the DASH diet itself reached 11.5 mm Hg in hypertensives against the high-sodium control diet; the two levers are sub-additive rather than additive.[59]
- The World Health Organization target is under 2 g sodium per day (about 5 g of salt). The US average is around 3.4 g per day.
- About 70% of dietary sodium in the US comes from salt added during commercial processing and restaurant preparation, not from the salt shaker.[60] Cutting at the shaker is the smallest possible lever; choosing fewer processed foods is the actual one, and for anyone with or at risk of high blood pressure, potassium-enriched salt substitutes are a reasonable swap for home salt unless kidney disease or a potassium-retaining drug rules them out; see Potassium salt substitutes.
The relationship to blood pressure is real and large in salt-sensitive adults (more common with age, in African ancestry, and in people with hypertension or chronic kidney disease) and smaller in salt-resistant individuals.[61] The population-level recommendation skews to the salt-sensitive side because the harm asymmetry favours caution.
The "J-curve" controversy. The PURE study (~102,000 people, 17 countries) reported a J-shaped association, with excess risk appearing only above about 7 g/day of sodium — 6 g/day in people with hypertension — and again below about 3 g/day.[62] The page does not endorse the J-curve, because PURE estimated 24-hour sodium from a single fasting morning urine via the Kawasaki formula — a method known to manufacture a spurious J-shape, partly because the formula itself incorporates age, weight, and creatinine, most of which independently predict mortality. Studies using repeated 24-hour urine collections find a graded, near-linear relationship, and much of the apparent low-sodium "harm" is reverse causation (already-ill people eat less).[63] The very-low-intake harm signal is largely a measurement artefact, not a reason to abandon sodium reduction.
Salt and stomach cancer is a separate, observational harm signal: a dose-response meta-analysis of prospective cohorts found +12% gastric cancer risk per 5 g/day of salt, via mucosal damage, endogenous N-nitroso formation, and synergy with H. pylori.[64] An earlier pooled analysis found a similar direction — 68% higher risk in the highest versus lowest salt-intake category.[65] (Evidence grade: Moderate.) Because the sodium-to-potassium ratio may predict cardiovascular outcomes better than sodium alone — in a trial follow-up that measured both with repeated 24-hour urine collections, so the spot-urine objection above does not apply[66] — favouring potassium-rich whole foods matters as much as cutting salt.
Alcohol
Evidence rating: Strong for cancer (IARC Group 1, no threshold); contested for all-cause mortality.
The minimum-risk dose is at or near zero, particularly for cancer. Alcohol is an IARC Group 1 carcinogen, and the cancer dose-response has no threshold.
- Cancer rises from the first drink. A comprehensive dose-response meta-analysis (572 studies, 486,538 cancer cases) found that even light drinking (≤1 drink/day) raises risk by about 13% for oral and pharyngeal cancer (RR 1.13), about 26% for oesophageal squamous-cell carcinoma (RR 1.26), and about 4% for female breast cancer (RR 1.04); heavy drinking (>50 g/day) raises oral and pharyngeal cancer risk roughly five-fold (RR 5.13) and oesophageal squamous-cell carcinoma nearly five-fold (RR 4.95).[67] For breast cancer specifically, risk rises 7.1% per 10 g/day of ethanol (95% CI 5.5–8.7%) — roughly one daily drink.[68] An estimated 4.1% of all new cancer cases in 2020 were attributable to alcohol, with moderate drinking (<20 g/day) alone accounting for about 103,000 cases.[69]
- The "J-curve" of supposed cardioprotection has largely collapsed. The 2018 Global Burden of Disease analysis found that the level of consumption minimizing total health loss is zero, with all-cause mortality and cancer risk rising monotonically; alcohol was the leading risk factor for death and disability at ages 15–49 globally.[70] The same collaboration softened that in 2022: the harm-minimising level stays near zero under 40, but for older adults the modelled curve becomes J-shaped and varies by region.[71] That "older adults may benefit" element is a relative-risk statement inside an age band, says nothing about any individual's absolute benefit, and is the single most-quoted line in industry communications. Mendelian-randomization studies — which use genetic variation to escape the confounding that inflated the old moderate-drinking "benefit" (sick ex-drinkers misclassified as non-drinkers) — find alcohol "not protective for any disease outcome," with each additional daily drink associated with about 2.65 mmHg higher systolic blood pressure (95% CI 1.40–3.89).[72] A second genetic analysis found roughly 30% higher odds of hypertension and 40% higher odds of coronary artery disease for each standard-deviation increase in predicted intake, with the risk curve steepening rather than flattening at higher consumption.[73]
The all-cause-mortality question remains genuinely contested as of 2026: the US National Academies review concluded with moderate certainty that moderate consumption is associated with lower all-cause and cardiovascular mortality — while still finding that breast-cancer risk rises about 5% for every 10–14 g of alcohol per day (RR 1.05, 95% CI 1.04–1.06), on a dose-response the review found linear rather than J-shaped, with increased risk at every amount. No conclusion in that review reached high certainty, and its non-fatal cardiovascular findings were rated low certainty.[74] A competing US federal review published the same year put meaningful risk above roughly six to nine drinks a week,[75] and WHO Europe's 2023 position holds there is no safe level. The 2025–2030 US Dietary Guidelines, released 7 January 2026, dropped numeric limits entirely in favour of advising people to drink less.[76] The cancer dose-response is not contested. See Alcohol for the full discussion, the Mendelian-randomization evidence, and the WHO Europe 2023 statement.
Refined grains and glycemic load
Evidence rating: Moderate.
The practical message here is "limit refined grains, white bread, sugary cereals, and potato products" — not "avoid carbohydrates." Glycemic load (GL; quantity × quality of carbohydrate) is a more robust predictor than glycemic index (GI) alone.
- High glycemic load is associated with coronary heart disease — about 5% higher risk per 50-unit increment (RR 1.05, 95% CI 1.02–1.08, across 15 cohorts and 438,073 people).[77] The mortality picture is weaker: a dedicated dose-response meta-analysis found no association between glycemic load and all-cause or cardiovascular death.[78]
- The largest analysis — the Richard Doll Consortium mega-cohorts, each over 100,000 people — found a high glycemic index associated with about 8% higher all-cause mortality (RR 1.08) and with diabetes-related cancer, and a high glycemic load with type 2 diabetes and cardiovascular disease (both about 15% higher, RR 1.15).[79] In PURE, one of the cohorts inside that pooling, the association between a high glycemic index and major cardiovascular events was steeper in people who already had cardiovascular disease (HR 1.51, 95% CI 1.25–1.82) than in those who did not (1.21, 1.11–1.34).[80]
See Glycemic index for where the GI/GL evidence is strong versus weak.
Cooking- and storage-derived compounds
Evidence rating: Weak / preliminary (mostly mechanistic and animal data, not hard human outcomes — do not over-read it).
- Acrylamide forms in high-heat starchy foods (fries, crisps, coffee, bread crust); it is an IARC Group 2A probable carcinogen with strong animal genotoxicity. But human epidemiology is largely null: a dose-response meta-analysis found no association between dietary acrylamide and any non-gynaecological site-specific cancer, the one exception being lung cancer in smokers.[81] The gynaecological sites are the genuine loose end — a companion dose-response analysis found little or even inverse association with breast cancer, but increases in ovarian and endometrial cancer risk, concentrated in never-smokers — with one exception that cuts the other way: in premenopausal women, breast-cancer risk rose steadily from about 20 µg/day, which is roughly the average intake.[82]
- Heterocyclic amines (HCAs) and polycyclic aromatic hydrocarbons (PAHs) from charred, grilled, or well-done meat are plausible mechanisms for the red/processed-meat → colorectal link rather than well-established independent exposures. Mitigation: lower-temperature cooking, marinating, and avoiding char and direct flame contact; the formation chemistry and the trials behind these tips are under Red and processed meat.
- Advanced glycation end-products (AGEs) from dry high-heat cooking: low-AGE diets modestly improve insulin resistance — about 1.2 units of HOMA-IR, a standard index where lower is better — and lower LDL and total cholesterol across thirteen small randomised trials, mostly in people with type 2 diabetes or obesity; inflammatory markers were not among the outcomes pooled, and there are no hard-outcome trials.[83] A reasonable reason to favour steaming, boiling, and stewing over frying and grilling — not a proven longevity lever.
Food contaminants
Evidence rating: Moderate. The message is "minimize and diversify," not "avoid" — fish and whole grains are net-beneficial.
- Methylmercury concentrates in large predatory fish (swordfish, shark, king mackerel, marlin, orange roughy, Gulf of Mexico tilefish, bigeye tuna). The benefits of eating fish exceed the potential risks — one to two servings a week is associated with about 36% lower coronary death — and the effects of low-level methylmercury in adults are not clearly established, so the guidance is species selection, not fish avoidance. The most sensitive group is pregnant or breastfeeding women and young children (neurodevelopment), but adults eating five or more servings a week should also favour lower-mercury species.[84]
- Inorganic arsenic in rice (which takes up ~10× more than other grains; brown rice higher than white because arsenic concentrates in the bran) is classified by IARC as a Group 1 human carcinogen, and neither of the two international bodies that set intake limits has been able to define a safe level — though the increment in low-exposure populations is modest. Mitigation: rinse, cook in excess water (about six parts to one) and drain, which removes a substantial share of the inorganic arsenic at the cost of some minerals and, in enriched rice, some added B vitamins; and vary grains. Sensitive group: infants fed rice cereal, for which the US action level is 100 µg/kg.
- Cadmium is most concentrated in shellfish and organ meats, but most dietary exposure comes from staple plant foods — cereals, potatoes and vegetables — with shellfish a minor contributor. Aflatoxins (improperly stored nuts, grains, and corn; IARC Group 1, liver cancer) act synergistically with hepatitis B and C infection.[85] Diversifying sources is the main defence, and for cadmium that means varying the staples, not avoiding the shellfish.
What's not on this list
- Eggs — earlier dietary-cholesterol concerns have eased, though not vanished. Whole eggs at up to about one a day are not associated with cardiovascular disease for most adults.[86] There is a geographic pattern: per ~1 egg/day, US cohorts show a small positive signal (RR 1.08, 95% CI 1.02–1.14), European cohorts are borderline (1.05, 0.98–1.14) and Asian cohorts show no association (0.96, 0.87–1.06)[87] — plausibly the companion foods rather than the egg itself, though the authors read their own results as support for limiting dietary cholesterol. Feeding studies summarised in a 2022 egg meta-analysis put each extra 100 mg/day of dietary cholesterol at roughly +1.9 mg/dL LDL cholesterol and +0.4 mg/dL HDL — the liver compensates by down-regulating its own synthesis[88] — but the ratio of total to HDL cholesterol rises slightly rather than staying put, according to the pooled trial analysis summarised in a 2013 meta-analysis.[89] Egg yolk is among the densest dietary sources of choline, much of it as phosphatidylcholine (a brain phospholipid and acetylcholine precursor), though beef liver carries more per 100 g than a whole egg. In a crossover trial in 18 adults, egg-yolk phospholipid gave a ~4× higher plasma choline response than synthetic choline bitartrate — but at a 3 g dose of choline, roughly twenty times what one egg delivers, and in a trial partly run by an ingredient company (two of its six authors were employees).[90] People with type 2 diabetes may warrant a more conservative cap: in the cohorts restricted to that group, each extra egg a day carried about 25% higher cardiovascular risk (RR 1.25, 95% CI 0.99–1.59) — a signal that stops just short of statistical significance, with considerable disagreement between studies.[91] An earlier meta-analysis also linked eating an egg a day or more to about 42% more incident type 2 diabetes in five Western general-population cohorts (HR 1.42, 95% CI 1.09–1.86) — observational, and its authors caution it may not justify changing dietary advice.[92] See Dietary fats for the geographic detail.
- Saturated fat in moderation from whole foods (yogurt, cheese, eggs, occasional butter) — not the same as saturated fat from processed foods. The "saturated fat causes heart disease" framing has been moderated in one specific way: pooled trials of saturated-fat reduction show no mortality benefit — nine trials in 13,532 people found no effect on cardiovascular death (RR 0.94) or death from any cause (RR 1.01).[93] The operative variable is the substitution effect: replacing saturated fatty acids (SFA) with polyunsaturated vegetable oil cut cardiovascular disease by about 30% in the trials the American Heart Association pooled, while replacing SFA with refined carbohydrates and sugars — unlike the starchy whole foods in some of the Cochrane trials above — did not.[94] Food matrix matters separately. In a cross-sectional NHANES analysis, a doubling of total SFA intake was associated with +0.42 years of GrimAge2 biological aging, with palmitic (+0.55 y) and stearic acid (+0.43 y) carrying the signal — though the same analysis flagged monounsaturated fat slightly more strongly (+0.54 years), both estimates lost statistical significance in the authors' own sensitivity analysis, and no correction was made for testing many fatty acids at once — so read the clock numbers as suggestive rather than as a verdict on any particular food.[95]
- Coffee, tea — generally favourable cohort and biomarker data; not on the harm list.
- Wine specifically — same risk profile as other alcohol; no special health benefit despite the resveratrol marketing.
- Gluten, dairy, legumes, nightshades — for adults without confirmed allergy, celiac disease, or specific sensitivity, no evidence of harm in cohort or trial data — an absence of evidence rather than proof of safety, and for nightshades there is essentially no research in either direction.
The 80/20 framing
The goal isn't perfection — it's the pattern.
- 80%+ of intake from whole or minimally processed foods — vegetables, legumes, fish, nuts, fruit, whole grains, olive oil, fermented dairy.
- Less than 20% from the categories on this list — ultra-processed food, sugar-sweetened beverages, processed and red meat, sweets, alcohol.
Treat 80/20 as a practical target rather than a measured threshold — no analysis has found the benefit levelling off at a fifth of intake. The reason not to push for 95/5 is adherence: a pattern you keep beats a stricter one you abandon. The goal is durable change, not optimization of a single week.
Further reading
- Lane MM et al. Ultra-processed food exposure and adverse health outcomes — umbrella review of epidemiological meta-analyses. BMJ 2024.[96]
- Hall KD et al. Ultra-processed diets cause excess calorie intake and weight gain — NIH metabolic-ward trial. Cell Metab 2019.[97]
- Bouvard V et al. Carcinogenicity of consumption of red and processed meat (IARC). Lancet Oncol 2015.[98]
- Johnston BC et al. Unprocessed red meat and processed meat — NutriRECS dietary guideline recommendations. Ann Intern Med 2019.[99]
- Lescinsky H et al. Burden of Proof — health effects of unprocessed red meat consumption. Nature Medicine 2022.[100]
- Malik VS, Hu FB. Sugar-sweetened beverages and cardiometabolic health — pooled analyses. Circulation 2019.[101]
- Pase MP et al. Sugar- and artificially sweetened beverages and the risks of incident stroke and dementia — Framingham Offspring. Stroke 2017.[102]
- Sacks FM et al. Dietary fats and cardiovascular disease — AHA Presidential Advisory. Circulation 2017.[103]
- Marklund M et al. Biomarkers of dietary omega-6 fatty acids and incident cardiovascular disease and mortality. Circulation 2019.[104]
- Sadeghi R et al. Dietary and circulating omega-6 fatty acids and their impact on cardiovascular disease, cancer risk, and mortality — systematic review and meta-analysis of 150 cohort and nested case-control publications. Journal of Translational Medicine 2025.[105]
- Kishimoto Y et al. Egg consumption and human health — a comprehensive review. Journal of Poultry Science 2026.[106]
- Fang Z et al. Association of ultra-processed food consumption with all-cause and cause-specific mortality — population-based cohort study. BMJ 2024.[107]
- Shi W et al. Red meat consumption, cardiovascular diseases, and diabetes — systematic review and meta-analysis. European Heart Journal 2023.[108]
- Zhao B et al. Associations of dietary cholesterol, serum cholesterol, and egg consumption with mortality. Circulation 2022.[109]
- Sacks FM et al. Effects on blood pressure of reduced dietary sodium and the DASH diet (DASH-Sodium). NEJM 2001.[110]
- de Goede J et al. Dairy consumption and risk of stroke — systematic review and updated dose-response meta-analysis. JAHA 2016.[111]
- Gonçalves NG et al. Association between consumption of low- and no-calorie artificial sweeteners and cognitive decline — an 8-year prospective study (ELSA-Brasil). Neurology 2025.[112]
- Chen H et al. Sweetened beverages and incident all-cause dementia among older adults. JAMA Psychiatry 2025.[113]
- Mousavi SM et al. Consumption of eggs and risk of all-cause and cause-specific mortality — dose-response meta-analysis. Advances in Nutrition 2022.[114]
- Witkowski M et al. The artificial sweetener erythritol and cardiovascular event risk. Nature Medicine 2023.[115]
- Witkowski M et al. Xylitol is prothrombotic and associated with cardiovascular risk. European Heart Journal 2024.[116]
- Bagnardi V et al. Alcohol consumption and site-specific cancer risk — comprehensive dose-response meta-analysis. British Journal of Cancer 2015.[117]
- GBD 2016 Alcohol Collaborators (Griswold MG et al.). Alcohol use and burden for 195 countries and territories, 1990–2016. Lancet 2018.[118]
- Li C et al. Meat consumption and incident type 2 diabetes — federated meta-analysis of 1.97 million adults. Lancet Diabetes & Endocrinology 2024.[119]
- O'Donnell M et al. Urinary sodium and potassium excretion, mortality, and cardiovascular events (PURE). NEJM 2014.[120]
- Neelakantan N et al. The effect of coconut oil consumption on cardiovascular risk factors — meta-analysis of clinical trials. Circulation 2020.[121]
- Jenkins DJA et al. Association of glycaemic index and glycaemic load with type 2 diabetes, cardiovascular disease, cancer, and all-cause mortality — meta-analysis of mega cohorts of more than 100,000 participants. Lancet Diabetes & Endocrinology 2024.[122]
- Fang X et al. Landscape of dietary factors associated with risk of gastric cancer — systematic review and dose-response meta-analysis of prospective cohort studies. European Journal of Cancer 2015.[123]
- Filippini T et al. Dietary acrylamide exposure and risk of site-specific cancer — systematic review and dose-response meta-analysis. Frontiers in Nutrition 2022.[124]
- Sohouli MH et al. The impact of a low advanced glycation end products diet on metabolic risk factors — systematic review and meta-analysis of randomised controlled trials. Advances in Nutrition 2021.[125]
- Cardoso BR et al. Association between ultra-processed food intake and biological ageing in US adults (NHANES 2003–2010). Age and Ageing 2024.[126]
For the deep dives that sit underneath this survey:
- Red and processed meat — Neu5Gc, TMAO, cooking-method mitigation, NutriRECS, dementia, APOE4
- Dietary fats — trans, saturated, PUFA, MUFA, eggs, omega-6:omega-3 ratio, epigenetic-clock signals
- Ultra-processed food — NOVA, the UCL crossover, biological-age effects
- Sweeteners — sugar, fructose, high-fructose corn syrup
- Sugar substitutes — non-sugar sweeteners, erythritol/xylitol, allulose
- Alcohol — Mendelian randomization, J-curve collapse