Development and application of chemometrical methods based on 1H-NMR spectra of fruit juices – Structural approach for analysis of wide and unbalanced datasets

Abstract

Analytical verification of the authenticity of fruit juices with respect to organic quality or production as direct juice (NFC) vs. juice from concentrate (FC) is challenging. To address this, the promising combination of 1H-NMR analysis with multivariate classification methods was applied to a sample collection of 506 orange juices, 649 apple juices and 224 tomato juices with the purpose to develop optimized classifying models for a subsequent use in juice monitoring. In a standardized procedure, the treatment of outliers, variable selection, the handling of unbalanced datasets and adequate model validation were addressed. The algorithms LDA, PLSDA, Random Forest, linear SVM, radial SVM and kNN were applied to each classification question. Therefore, standardized guidelines for selecting an optimal sampling method and a final model for the use in routine were developed. The differentiation of the classes orange – organic vs. non-organic, orange – NFC juices vs. FC juices, apple – NFC juices vs. FC juices and tomato – NFC juices vs. FC juices, was achieved each with a mean accuracy of above 95.0% in independent validation. These high performing models were subsequently implemented into an R-Shiny application for routine analysis. The final models for apple - organic vs. non-organic and tomato - organic vs. non-organic performed with accuracies of 82.8% and 89.8%, respectively. Ultimately, through standardization and optimization, it was possible to transparently select valid classification models that are now used in routine control analysis of orange, apple, and tomato juices.