Machine-derived diet score tracks lower aging-related mortality—but does not prove a longevity diet

What happened

Researchers developed a machine-learning dietary score that was associated with lower aging-related mortality in a large human cohort. The peer-reviewed study, published August 7 in *npj Science of Food*, analyzed data from 191,689 UK Biobank participants and named the resulting pattern the Machine-learning YouTHful, or MYTH, Diet.

Participants in the highest quarter of the score had a reported 21% lower hazard of aging-related mortality than those in the lowest quarter in the internal validation analysis. The hazard ratio was 0.79, with a 95% confidence interval of 0.75 to 0.84. An external validation analysis produced a hazard ratio of 0.68, with a confidence interval of 0.58 to 0.80.

Those figures describe associations. They do not show that adopting the diet score caused the difference in mortality or that the score can extend an individual’s lifespan.

How the score was built

The investigators first performed a broad analysis of reported food consumption and identified 18 food groups associated with aging-related mortality. A Light Gradient Boosting Machine algorithm then ranked the food variables, which the researchers reduced to a 10-component score ranging from zero to 10.

This was a data-derived observational design rather than a randomized dietary intervention. The primary population came from UK Biobank, a long-running prospective research cohort with linked health records and extensive questionnaire, biomarker and imaging data.

UK Biobank collects dietary information through self-completed food-frequency questions and, for a subset of participants, repeated online 24-hour recalls. Its documentation says these tools can rank people by consumption of major food groups, but brief recalls and questionnaires cannot perfectly represent long-term intake.

The study’s main endpoint was aging-related mortality. Secondary analyses connected higher scores with lower risks for 15 aging-related diseases and with more favorable biological-aging estimates for the lungs, liver and pancreas. Proteomic, metabolic and inflammatory measurements were also used to explore possible mediating pathways, including signals involving TNFRSF4, polyunsaturated fatty acids and lipid-related metabolites.

Why the result is not a prescription

Machine learning can detect combinations that conventional analyses may overlook, but it cannot eliminate the central limitations of observational nutrition research. People who report healthier diets may also differ in smoking, physical activity, education, income, healthcare access, medication use or pre-existing illness. Statistical adjustment can reduce those differences without removing all residual confounding.

Reverse causation is another concern. Illness can change appetite and food choices before a diagnosis or death, making diet appear predictive even when underlying disease influenced both the exposure and the outcome.

UK Biobank also is not demographically representative of the broader population. An official comparison found a healthy-volunteer selection effect: participants tended to be healthier and less socioeconomically deprived than the eligible population. That does not invalidate associations within the cohort, but it limits how confidently absolute risks and dietary rankings can be generalized.

The score was selected using the same broad research ecosystem in which many candidate foods, biomarkers and outcomes were available. Internal and external validation strengthen the result, yet independent replication is still necessary to test whether the exact 10-component structure remains stable across cultures, dietary measurement systems and populations with different baseline health.

Practical meaning

The study provides a potentially useful research tool for comparing dietary patterns with aging-related outcomes. It does not establish a clinically validated “anti-aging diet,” and its organ-aging measures are surrogate estimates rather than direct evidence of preserved function.

A randomized trial would be needed to determine whether deliberately increasing the score changes disease incidence, physical or cognitive function, mortality, or quality of life. Such a trial would also need to assess adherence, nutritional adequacy and possible differences by age, sex, health status and cultural eating patterns.

For now, the MYTH score is best read as a new observational model: informative about which dietary combinations deserve further testing, but not proof that an algorithm has identified a life-extending menu.