Model calibration and feature selection for orange juice authentication by 1H NMR spectroscopy

A 1H NMR spectroscopic profiling approach has been investigated to discriminate between authentic and adulterated juices. An experimental database of 150 samples of authentic or adulterated orange juices, with a known percentage of clementine juice, was prepared. A repeated stratified cross-validation process was adopted for the validation of PLS regression models and classification rules. The choice of a type of statistical data pre-treatment was discussed. The result was that logarithmic transformation combined with Pareto scaling was the most relevant. The selection of spectral variables has also proven to lead to better results than using the whole spectral range. Various feature selection procedures were compared. The CovSel approach appeared to be the most efficient. However, for a better understanding of the features of atypical profiles, it would be wise to use more than one selection procedure like the ones based on Backward Interval PLS regression approach or on the genetic algorithms.