10 Million People, 32 Outcomes, and One Detail Nobody Quotes

A 2024 umbrella review in the BMJ pooled 45 meta-analyses covering roughly 9.9 million people, examining 32 health outcomes associated with ultra-processed food consumption.

The findings were consistent in direction and large in size. All-cause mortality was approximately 21% higher among the highest consumers compared with the lowest. Cardiovascular mortality was around 50% higher. Type 2 diabetes and anxiety outcomes were both graded as supported by convincing evidence.

Then the detail that almost never survives the summary: 84% of those analyses were rated low or very low quality. By the review's own authors, using standard evidence-grading criteria. Only 9% reached the highest tier.

This is not a hostile reading of the paper. It is the paper's own conclusion, and stating it is not the same as dismissing the finding. An umbrella review is a synthesis of syntheses — it inherits every limitation of the meta-analyses beneath it, which in turn inherit the limitations of the observational cohorts beneath them. Grading the result honestly is what a good review does.

There is a second structural problem, in the exposure definition itself. Ultra-processed food is defined by the NOVA classification, which categorises by degree and purpose of industrial processing rather than by nutritional composition. The resulting category holds soft drinks, packaged confectionery and reconstituted meat products — and also wholegrain supermarket bread, plain yoghurt with added cultures, and tinned legumes. Grouping these together and computing a single hazard ratio assumes they behave similarly. They almost certainly do not.

Residual confounding compounds this. High ultra-processed food intake correlates with lower income, lower physical activity, higher smoking rates, shift work, and reduced healthcare access. Adjustment helps but never fully separates a dietary pattern from the life circumstances that produce it.

What remains after all of that is still worth something. The direction is consistent across dozens of independent analyses and many populations, which is not nothing. Mechanistic candidates exist and are reasonable: energy density, hyperpalatability driving passive overconsumption, low fibre, and the displacement of whole foods. A small tightly-controlled feeding trial found people ate roughly 500 more calories per day on an ultra-processed diet matched for presented macronutrients — direct experimental support for at least one part of the story.

The accurate position is that the direction is well supported and the certainty is routinely overstated. "Ultra-processed food is associated with worse outcomes, and reducing it is a reasonable default" is defensible. "Ultra-processed food causes a 21% increase in mortality" is not what this evidence supports.

Why this matters for tracking: the NOVA category is too broad to be a useful personal metric. If you want to track something from this literature that is likely to matter, track the specific mechanisms it implicates — total energy intake, fibre grams, and the proportion of meals you prepared yourself. Those are measurable, actionable, and do not require deciding whether your bread counts.