Predicting carbohydrate quality in a global database of packaged foods

Background: Carbohydrates are the major contributor to the energy intake of worldwide population. There is established evidence of links of carbohydrate quality with human health. Knowledge of specific carbohydrate in packaged food, such as added and free sugars, could help further investigate this link, however this information is generally not available. Objective: To develop an algorithm to predict the content of free sugars in a global database of packaged foods and beverages; and test the applicability of the algorithm to assess carbohydrate quality in packaged food products from different countries and monitor the evolution over time. Carbohydrate quality was defined using a 10:1|1:2 ratio for carbohydrate, fibers and free sugar, i.e., for every 10 g of total carbohydrates in a diet or product, there is at least 1 g of dietary fibers, and less than 2 g of free sugars for every 1 g of dietary fibers. Methods: We used a machine learning approach to predict added and free sugars, which enabled us to predict the carbohydrate quality of products from a global database of packaged food. Our predictions were tested by splitting the dataset into training, validation, and test sets, using US data. Results: We were able to predict free sugars and carbohydrate quality for 424,543 products in the U.S. and in 14 countries. The overall mean absolute error on the test set was 0.96 g/100 g of product. The predictions generalized with a high accuracy to non-US countries, and we were able to effectively predict the proportion of products meeting the 10:1|1:2 criteria in the food supply of 15 countries. Conclusion: Our methodology achieved high accuracy and is fully automated; it may be applied to other databases of packaged products and can be easily applied for continuous monitoring of the carbohydrate quality of the global supply of packaged food.