Towards reliable estimation of an "electronic tongue" predictive ability from PLS regression models in wine analysis

The paper is devoted to an assessment of the predictive power of PLS (partial least squares) models derived from "electronic tongue" data. A multisensor system ("electronic tongue") based on a potentiometric platform was applied to the analysis of wines. Both white and red wine varieties were analyzed employing different sensor arrays. 36 different samples of white wines from New Zealand (Chardonnay, Sauvignon Blanc, Pinot Gris varieties) were analyzed by a number of standard chemical techniques to assess the contents of free and total sulfur dioxide, total acidity, ethanol, pH and some phenolics. Furthermore, 27 samples of red wines produced in Slovakia (Blaufränkisch variety) were assessed by a skilled sensory panel to rate a set of 7 taste descriptors. In addition, all of the wines were analyzed by potentiometric electronic tongue (ET). PLS regression (partial least squares) was used to assess the correlation between ET response, and chemical analytical data, or human perceived sensory characteristics of the wines. Methods that are widely used in the ET literature for estimation of the predictive ability of the PLS models, such as full cross-validation and test set validation with a single random split of samples, were compared with a k-fold random split test set approach. It was shown that the latter does not tend to produce over-optimistic results in small data sets, as are typically available in ET research.