MYTH diet score tracked 21% lower aging-related death risk, level with an older healthy-eating index

Most diet scores start from a rulebook. The Healthy Eating Index, widely studied in aging research, is based on the Dietary Guidelines for Americans. Eat as the guidance advises, and you score well.

Yating Miao, Yanan Ma and colleagues at China Medical University in Shenyang, with collaborators at Jilin University and Zhejiang University School of Medicine, built a score the other way around. They named it the “Machine-learning YouTHful” (MYTH) diet. Writing in npj Science of Food, they used data from 191,689 participants in the UK Biobank, a long-running British health study. The team weighed the link between 34 food groups and aging-related deaths, then let a machine-learning model rank the foods that mattered most.

The 23,443 people who scored 5 or higher out of 10 were about a fifth less likely to die of aging-related causes over some 12 years than those scoring 2 or less. The name promises something new. Open the score up, though, and most of it is advice people have heard for years. Two rules cut against it. And compared with three established dietary indices, the score ran level with the benchmark, the Healthy Eating Index.

The MYTH diet’s ten food rules, and the two that cut against standard advice

Everyone got a point for each food where their intake fell within the target range, so totals ran from 0 to 10. Poultry, nuts and refined grains were classified as encouraged foods, at 80, 20 and 150 grams a day or more. Processed meat was capped at 20 grams a day, while red meat, butter and sugar-sweetened beverages earned the point only at zero; all four were categorized as discouraged foods. Legumes, coffee and alcohol were recommended for moderate consumption, at 30 to 100 grams of legumes and 70 to 200 of coffee a day. Beer and wine were combined into a single category, and the point went to anyone drinking 60 to 400 grams of beer or 250 to 400 of wine a day.

A 2020 review for the Dietary Guidelines Advisory Committee covered 153 studies of diet and death. The eating patterns it tied to longer lives featured vegetables, fruits, legumes, nuts, whole grains, fish, and lean meat or poultry when meat was included. They were relatively low in red and processed meat, high-fat dairy, and refined carbohydrates or sweets. Most of the MYTH rules point the same way. But the score has no place for whole grains, fruit, vegetables or fish.

And it rewards refined grains. The authors note that existing evidence on them remains inconsistent, and suggest they may serve as a key energy source in an overall balanced diet. Their own introduction had faulted older indices because they “do not adequately differentiate whole grains from refined grains”.

The wine range breaks with drinking guidance. At its low end, 250 grams a day is roughly one large glass of red or white wine, which the NHS lists as 250 milliliters at 12% strength, or 3 units. The NHS advises men and women not to drink more than 14 units a week on a regular basis. A glass that size every day comes to about 21 units a week.

How machine learning turned 34 food groups and 13,652 deaths into a score

A subset of 210,744 volunteers used a questionnaire called the Oxford WebQ to record what they had eaten and drunk the previous day, one to five times between 2009 and 2012. The outcome, aging-related death, meant death from any cause except infectious and parasitic diseases, congenital conditions, accidents, other external causes and ill-defined causes. Heart disease, cancer and stroke deaths were all included. Over a median follow-up of 12 years, 13,652 people in the analysis died this way, about 7 in every 100.

Each food group was first tested on its own, adjusting for age, sex, smoking, weight, deprivation and other factors. Eighteen stood out; two that overlapped with the rest, other vegetables and tea, were dropped. A machine-learning model called LightGBM was then used to rank food importance, how much each food helped predict who died, putting poultry first and red meat second. Foods joined the score in that order until the team stopped at wine, the 11th, and each target was set at the intake level associated with the lowest risk. Whole grains missed the cut, even though people eating 44 grams a day or less had a 16% higher risk than those eating about 170.

Miao and Ma’s group calls this among the first efforts to integrate long-term whole-food intake data with machine learning for an aging diet. Letting data pick an eating pattern is not new, though: the 2020 review counted 25 articles that did it with factor and cluster analysis.

Checked in US data, the MYTH score ran level with the Healthy Eating Index

That fifth-lower risk was measured in the same British volunteers whose data picked the foods and set the cut-offs. Tuned on one set of people, a score partly fits their quirks and chance patterns, so it looks best there.

The fairer test used the US National Health and Nutrition Examination Survey, known as NHANES, which collects diet through dietary recall interviews. In the NHANES cohort, top scorers were about a third less likely to die than the lowest scorers, though that estimate was less precise. A pattern that survives a move to another country is less likely to be a British fluke. Both datasets are observational, however, so the replication cannot show that the food itself is responsible.

Nor did the MYTH score clearly beat the benchmark. In UK Biobank, its performance was comparable to that of the Healthy Eating Index. In NHANES, the older index showed a marginally stronger association when both were treated as sliding scales, while the MYTH score edged ahead at the top end.

Blood proteins explained about a quarter of the MYTH diet link, on paper

UK Biobank also measured 2,923 plasma proteins, the ones circulating in blood, in some volunteers. A signature of 50 of them that tracked the score accounted for an estimated 27% of its association with aging-related death. TNFRSF4, an immune co-stimulatory receptor that may influence chronic inflammation, had the biggest single share, about 17%. The protein data also fed clocks estimating how old individual organs look, and higher scores were associated with reduced aging across five organ systems, from the lungs to the arteries. The score also went with lower rates of 15 of the 49 age-related diseases tested, though most of those gaps were small.

But the blood was measured at a single time point, while diet and death risk play out over years. The pathways, the authors write, “should be regarded as hypothesis-generating rather than evidence of causal mechanisms”. They are routes through the numbers; whether food travels them inside the body is untested.

Food recalls, healthy volunteers and smoking blur the MYTH diet evidence

A score about decades of eating rests on a few days of it. About four in ten participants, 40.1%, completed one recall, and the authors write that such recalls may not fully capture long-term habitual intake. UK Biobank volunteers also tend to be healthier and better off than others their age, which the researchers say may lead to underestimation of some associations.

Nor can the team fully rule out residual confounding or reverse causality: hidden differences between high and low scorers, or early illness changing what someone eats. Setting aside those who died within the first two or three years, when undiagnosed disease is most likely to have shaped a diet, left the results largely consistent.

Smoking is the obvious example. The lungs showed the largest organ-specific effect in the study, and emphysema and lung cancer were among the strongest disease links. Yet the smoking adjustment sorted everyone only as never, current, or former smokers, with no allowance for how much they smoked. Smoking causes most cases of chronic obstructive pulmonary disease, a group of lung diseases that includes emphysema. If high scorers smoked less within those categories, some of the lung signal could belong to tobacco.

The researchers say future randomized controlled trials will be essential to establish whether the diet itself changes aging. Until then, the best-supported rules are those the score shares with older advice. The 2020 review concluded that “a high-quality dietary pattern comprised of nutrient-dense foods, regardless of the label, associated with reduced all-cause mortality risk”.

The two rules that part from that advice, refined grains and wine by the large glass, are the ones with the least support outside this study.