Ultra-processed food and mortality: what the pooled evidence actually grades

An umbrella review of 45 meta-analyses, nearly ten million people, and one honest asterisk: almost all of it is graded low or very low quality, because it is observational.

mixedStudies disagree, or the mechanism is contested.Revised 02 Oct 2026

Supermarket aisle of packaged foods photographed in flat daylight

The NOVA categories are a definition, not a verdict. Nine and a half million people is a lot of data about association, not a controlled experiment.

The short version

Eat less of it. That is what the pooled evidence supports, with a caveat printed below that matters more than the headline: the certainty ratings attached to most of these findings are low or very low, because the underlying studies are observational cohorts, not experiments.

What the umbrella review found

The 2024 BMJ umbrella review pulled together every meta-analysis it could find on ultra-processed food, defined per the NOVA classification used by the authors as food formulated mainly from ingredients extracted from petroleum or derived from biomass, and their association with health outcomes. Forty-five unique pooled analyses, thirteen of them looking at graded steps of exposure, covering 9,888,373 people.

In 32 of those 45 analyses (71%) the association ran in the same direction: more exposure, worse outcome, across mortality, cardiovascular disease, type 2 diabetes, depression, anxiety, sleep problems and gastrointestinal outcomes.

The authors did something rarer than a plain meta-analysis. They graded each finding, in five evidence classes, and then applied the GRADE framework. Only four associations reached the top class:

cardiovascular-disease-related mortality, risk ratio 1.50 (95% CI 1.37 to 1.63), GRADE very low
type 2 diabetes, per graded step of exposure, risk ratio 1.12 (1.11 to 1.13), GRADE moderate
prevalent anxiety, odds ratio 1.48 (1.37 to 1.59), GRADE low
combined common mental disorders, odds ratio 1.53 (1.43 to 1.63), GRADE low

One class down, “highly suggestive”: all-cause mortality, risk ratio 1.21 (1.15 to 1.27), GRADE low; heart-disease-related mortality, hazard ratio 1.66 (1.51 to 1.84), GRADE low; and depressive outcomes, hazard ratio 1.22 (1.16 to 1.28).

Read the two numbers next to each other. The single finding with the highest certainty, a graded effect on type 2 diabetes, is a ratio of 1.12 per step. The one that sounds most alarming, cardiovascular mortality at 1.50, carries the lowest certainty grade in the whole table. Pooled evidence does not sort that way for a reason: observational data can be consistent and still be confounded, graded down for exactly that.

The mortality meta-analysis

A 2022 systematic review in the American Journal of Epidemiology worked on the mortality question directly: 40 prospective cohorts, 5,750,133 individuals, publications from 1984 to 2021. Comparing the highest intake group against the lowest, ultra-processed food carried a pooled relative risk of 1.29 (1.17 to 1.42) for all-cause mortality.

The useful part of that paper is what sits next to the headline. Stratified, sugar-sweetened beverages gave 1.11 (1.04 to 1.18), artificially sweetened beverages 1.14 (1.05 to 1.22), processed and red meat 1.15 (1.10 to 1.21), while breakfast cereals came out below one, at 0.85 (0.79 to 0.92). A blanket “processed food is worse” rule cannot explain a category that runs the other way. The authors also flag the field’s real methodological weakness: no standardised method for classifying how processed a food is.

A mechanism-adjacent finding

The French NutriNet-Santé cohort gives one of the few findings that comes with a plausible mechanism attached. In 110,260 adults followed from 2009, each ten-point increase in the share of ultra-processed food in the diet was associated with a higher risk of overweight (hazard ratio 1.11, 1.08 to 1.14) and of obesity (1.09, 1.05 to 1.13), and with a small but significant gain in body mass index over follow-up. The association held after adjusting for total energy and nutritional quality of the diet.

That last clause is what makes the finding worth reading twice: adjusting for calories and diet quality did not erase it. The authors are explicit about the limits, namely a self-selected cohort, 78% women, mean age 43, residual confounding, and possible misclassification of processing level.

What the data do not show

No trial has assigned people to eat more or less ultra-processed food and then counted deaths. Not one. Everything above is association, and association in food epidemiology is permanently entangled with income, education, smoking, physical activity, sleep, and the fact that the people who buy the most industrialised food are often the ones with the least time for cooking.

Nobody has shown that one NOVA category, a specific additive or a specific kind of packaging, is the culprit. Nor that the 1.21 all-cause figure means twenty-one extra deaths per hundred people. The category is a definition applied to shopping baskets; the mechanism is still a working hypothesis, not a finding.

Nor is any of this a verdict on the Mediterranean diet, on home cooking, or on the older adult who eats a fortified cereal for breakfast, the one category that moved favourably. It is a statement about an exposure defined by industrial formulation, and it says nothing about any single product on any single shelf.

What to do with it

Treat the effect size as the message, not the mechanism. A 1.12 to 1.21 range on all-cause mortality, graded low to very low, is not a reason to reorganise a diet around labels. It is a reason to notice how much of a day’s intake arrives in a wrapper, and to move some of it: water instead of sweetened drinks, food you can recognise as food in place of the fifth thing that comes off a factory line, and a reading of the boring basics first.