Poultry, Nuts and a Glass of Wine: How an Algorithm Rewrote the Longevity Menu
Researchers in China used machine learning on dietary data from nearly 192,000 UK Biobank participants to build a 10-component “MYTH” diet score (Machine-learning YouTHful) optimized specifically to predict aging-related death rather than general disease. People in the top quartile of adherence had roughly 21 percent lower aging-related mortality in the UK cohort and about 32 percent lower in a separate US cohort, with the diet also linked to slower biological aging in the lungs, liver, pancreas, kidneys and arteries, and lower risk of 15 age-related diseases. The effects are real at population scale but statistically small per person, and because this is an observational study it shows association, not proof of cause.
For decades, dietary advice has been built by committees deciding which foods are “good” and which are “bad.” A team from China Medical University, Jilin University and Zhejiang University tried something different. They let an algorithm decide.
The idea, published in npj Science of Food, was to stop asking what prevents heart disease or cancer individually and instead ask a single question: which eating pattern best tracks with dying more slowly of aging itself. To answer it, the team pulled records from 191,689 middle-aged and older adults in the UK Biobank, screened 34 food groups against aging-related death, then fed the survivors of that screen into a gradient-boosting machine learning model that ranked each food by how much it helped predict who aged toward death and who did not.
The result is the MYTH diet, a simple 0 to 10 checklist. It rewards poultry, nuts and, surprisingly, refined grains. It penalizes processed meat, red meat, butter and sugary drinks. And it takes a notably grown-up stance on the middle ground, awarding points for moderate legumes, coffee and even moderate alcohol, whether beer or wine, while penalizing both zero and excessive intake. Several foods showed a U or J shaped relationship with mortality, meaning the lowest risk sat at a moderate intake, not at zero and not at maximum.
The payoff scaled with the score. People in the highest adherence quartile had about a fifth lower aging-related mortality than the lowest, and the pattern reproduced in an independent US cohort where it slightly outperformed three established diet indices. The team then went looking for why. Using blood proteins, metabolites and inflammation markers from tens of thousands of participants, they traced part of the benefit to lower inflammation, healthier fat metabolism and a specific immune protein called TNFRSF4. Higher scores were also tied to younger-looking organs, most strongly the lungs and liver.
The big idea is not any single food. It is the method: using population-scale data and machine learning to derive a diet aimed squarely at the biology of aging, then reverse-engineering the molecular machinery behind it. The caution is equally clear. Nobody was actually assigned to eat this way. The signal is a correlation in people who already chose their diets, so the true test, a randomized trial, has not yet been run.
Actionable Insights
The take-home foods are unglamorous and cheap: eat some poultry (about 80 grams a day or more), a small handful of nuts (20 grams or more), and keep legumes, coffee and, if you drink, alcohol to moderate amounts. Cut processed and red meat, butter and sugary drinks toward zero. None of this is exotic, and that is the point.
Now the honest part about magnitude. The headline “21 percent lower risk” describes relative risk, not your personal odds. In this study only about 7 in 100 people died of aging-related causes over roughly 12 years. Shaving that by a fifth moves the top-adherence group to somewhere near 5.5 to 6 in 100. In plain terms, that is on the order of one to two fewer deaths per hundred people over a decade, which is meaningful across a whole country but modest for one individual.
Translated into a standardized effect size (Cohen’s d, where 0.2 is considered small, 0.5 medium, 0.8 large), the top-versus-bottom quartile contrast works out to roughly d = 0.13 in the UK data and about d = 0.21 in the US data. These are small effects. That does not make the advice wrong; small, cheap, safe changes applied for decades add up. But it does mean you should not expect a dramatic personal transformation from a diet score, and you should weight it accordingly against interventions with larger evidence, such as not smoking or staying physically active.
Context and Source
- Open Access Paper: A Machine Learning-Derived Dietary Pattern for Aging" (MYTH Diet)
- Authors: Yating Miao, Zhirong Li, Xinyao Zhang, Zuyun Liu, Yanan Ma.
- Institutions and country: China Medical University (Shenyang), Jilin University (Changchun) and Zhejiang University School of Medicine (Hangzhou), China. Data sources are the UK Biobank (United Kingdom) and NHANES (United States).
- Journal: npj Science of Food (Nature Portfolio).
- Impact evaluation: The impact score of this journal is 7.53 (2024 Journal Impact Factor; the newer 2025 JIF released June 2026 is 9.7, and its CiteScore is 7.6), evaluated against a typical high-end range of 0 to 60+ for top general science journals, therefore this is a Medium impact journal.
