A Machine-Learning-Imputed Global Atlas of Ultra-Processed Food Supply Shares and PIF-Based Burden Estimates for Non-Communicable Diseases

No open atlas of ultra-processed food supply exists with documented predictive validity across countries, and the global non-communicable disease burden linked to such foods, estimated under counterfactual exposure scenarios with transparently decomposed uncertainty, has not been quantified within a single framework. We built a machine-learning-imputed atlas of the share of dietary energy from ultra-processed food for 174 countries over 2010 to 2023, mapping 44 published national estimates onto food supply structure via Random Forest regression with leave-one-country-out validation. Five linked evaluations follow: an ecological phenome-wide association study across 27 non-communicable disease outcomes; a potential impact fraction estimation with time-varying exposure and a five-layer sensitivity decomposition; a 60-year generational evaluation of processed macro-ingredient supply; a synthetic-control assessment of sugar-sweetened beverage taxation across 16 countries; and a cross-level comparison of ecological and individual-level effect magnitudes using nationally representative survey data. The supply-side estimate yields a range of 1.6 to 9.5 million disability-adjusted life years in 2021. A credibility-discounted figure places the burden at roughly 3.6 million. Burden growth from 2010 to 2021 was entirely denominator-driven: population ageing and disease prevalence expansion supplied 103% of the increase, while changing supply contributed minus 3 percent. This holds consistently with a 20-year generation-lag between dietary-structure change and population obesity. Leave-region-out cross-validation returns zero generalisability for Latin America and the Caribbean. The denominator-driven pattern does not imply ultra-processed food is harmless; it indicates the processed-food environment was structurally established in most countries by 2010. The atlas, sensitivity framework, and all code are released as public-health infrastructure.