Artificial intelligence detects the dismantling of nutrient co-occurrence architecture by ultra-processing: explainable machine learning and network analysis of 22,481 usda-classified foods
TL;DR
UPFNet is constructed, an artificial-intelligence framework that couples nutrient co-occurrence network analysis with explainable machine learning across 22,481 foods from two USDA composition databases spanning 788,582 measurements across 33 nutrients, with NOVA group assignment via a validated 67-term keyword algorithm.
Abstract
Ultra-processed foods have been linked to 32 adverse health outcomes in an umbrella review of epidemiological meta-analyses spanning evidence published between 2009 and 2023, yet the nutritional mechanism underlying these associations remains poorly characterised at scale. We constructed UPFNet, an artificial-intelligence framework that couples nutrient co-occurrence network analysis with explainable machine learning across 22,481 foods from two USDA composition databases spanning 788,582 measurements across 33 nutrients, with NOVA group assignment via a validated 67-term keyword algorithm. Vitamin C, water, vitamin B12 and protein were depleted by 85.7%, 81.8%, 81.2% and 63.4%, respectively, in NOVA 4 relative to NOVA 1 (all p < 10−9; 27 of 33 nutrients survived Bonferroni correction). The whole-food co-occurrence network held 249 edges, whereas the ultra-processed network retained only 162, a 34.9% collapse (KS = 0.195; p = 3.37 × 10−9). The dominant hub shifted from fibre (betweenness centrality 0.790) to water (0.940). Gradient boosting classification achieved an AUROC of 0.9848 (95% CI 0.9793–0.9903), an accuracy of 94.2%, and a fold coefficient of variation of 0.29%, with mean-decrease-in-impurity attribution identifying energy, water, carbohydrate, sodium and protein as the five dominant discriminative features. Industrial processing fundamentally dismantles the architecture of nutrient co-occurrence, not merely individual nutrient content.
