Abstract
BACKGROUND: Training of machine learning algorithms on dish images collected in other countries requires possible sources of systematic discrepancies, including country-specific food composition databases (FCDBs), to be tackled. The US Nutrition5k project provides for ~5000 dish images and related dish- and ingredient-level information on mass, energy, and macronutrients from the US FCDB. The aim of this study is to (1) identify challenges/solutions in linking the nutritional composition of Italian foods with food images from Nutrition5k and (2) assess potential differences in nutrient content estimated across the Italian and US FCDBs and their determinants. METHODS: After food matching, expert data curation, and handling of missing values, dish-level ingredients from Nutrition5k were integrated with the Italian-FCDB-specific nutritional composition (86 components); dish-specific nutrient content was calculated by summing the corresponding ingredient-specific nutritional values. Measures of agreement/difference were calculated between Italian- and US-FCDB-specific content of energy and macronutrients. Potential determinants of identified differences were investigated with multiple robust regression models. RESULTS: Dishes showed a median mass of 145 g and included three ingredients in median. Energy, proteins, fats, and carbohydrates showed moderate-to-strong agreement between Italian- and US-FCDB-specific content; carbohydrates showed the worst performance, with the Italian FCDB providing smaller median values (median raw difference between the Italian and US FCDBs: -2.10 g). Regression models on dishes suggested a role for mass, number of ingredients, and presence of recreated recipes, alone or jointly with differential use of raw/cooked ingredients across the two FCDBs. CONCLUSIONS: In the era of machine learning approaches for food image recognition, manual data curation in the alignment of FCDBs is worth the effort.