Modelling of prebiotic-supplemented powdered infant formulas quality parameters using machine learning

Abstract

The food industry has made several attempts to develop infant formula (IFs) that fulfill the nutritional needs of the new born including supplementation with prebiotic oligosaccharides. Some processes in IF manufacture, involving heat, can give rise to a loss of nutritive value. Maillard reaction (MR) takes place during IFs processing and storage causing deterioration of proteins. Furosine is an indicator of early stages of MR while hydroxymethylfurfural (HMF) is used to evaluate intermediate and advanced stages. On the other hand, data modelling allows finding valuable relationships between the chemical changes suffered by the samples. Two classification methods based on statistical learning theory, support vector machines (SVM) and random forests (RF) have been applied on spectroscopic data of milk powder for brand identification and component analysis. Therefore the aim of this work has been to use SVM and RF classification methods to study the relationships between 24 commercial prebiotic-supplemented and not IFs considering a set of quality parameters: pH, aw, protein content, carbohydrate composition (reducing sugars and prebiotic content) and MR indicators (furosine and HMF). For this, samples were grouped in different categories according to i) protein source (whey, whey hydrolysates, milk, soy, amino acids), ii) carbohydrate source (lactose, maltodextrins, lactose and maltodextrins) and iii) by age, (starting, follow up) types.Furosine, HMF and prebiotic content ranged from 94 - 1226 mg 100 g-1 protein, 62 -510 μg 100 g-1 and 0.08 – 5 g 100 g-1 product, respectively. Then, samples were classified using SVM and RF. SVM finds patterns in empirical data with regard to sample categories. In RF, multiple decision trees are constructed, outputting the different classes and each node is split using a subset of variables. These models were trained with 70% of the data, 10-fold crossvalidated and then tested with 30% of new data. According to the results obtained, high classification rates of new IFs were reported with these models, being according to protein and carbohydrate source and by age types, 90.0, 90.0 and 90.0% and 80.0, 90.0 and 85.0%, for SVM and RF prediction parameters, respectively. Furosine and reducing sugars were the most influent factors in SVM, while prebiotic content was also influent in RF prediction.These models allowed to classify commercial IFs and could be applied on identification of IFs from unknown origin.