Analytica Chimica Acta 558 (2006) 144–149 Discrimination of wines based on 2D NMR spectra using learning vector quantization neural networks and partial least squares discriminant analysis Saeed Masoum a , Delphine Jouan-Rimbaud Bouveresse b , Joseph Vercauteren c , Mehdi Jalali-Heravi a , Douglas Neil Rutledge b,∗ a Department of Chemistry, Sharif University of Technology, P.O. Box 11365-9516, Tehran, Iran b Laboratoire de Chimie Analytique, UMR 214 INRA/INA P-G, 16, rue Claude Bernard, 75005 Paris, France c Laboratory of Pharmacognosy, Faculty of Pharmacy, 15, Avenue Charles Flahault, University of Montpellier 1, 34093 Montpellier Cedex 5, France Received 21 June 2005; received in revised form 3 November 2005; accepted 7 November 2005 Available online 20 December 2005 Abstract The learning vector quantization (LVQ) neural network is a useful tool for pattern recognition. Based on the network weights obtained from the training set, prediction can be made for the unknown objects. In this paper, discrimination of wines based on 2D NMR spectra is performed using LVQ neural networks with orthogonal signal correction (OSC). OSC has been proposed as a data preprocessing method that removes from X information not correlated to Y. Moreover, the partial least squares discriminant analysis (PLS-DA) method has also been used to treat the same data set. It has been found that the OSC–LVQ neural networks method gives slightly better prediction results than OSC–PLS-DA © 2005 Elsevier B.V. All rights reserved. Keywords: Learning vector quantization (LVQ) neural networks; Partial least squares (PLS) discriminant analysis; Orthogonal signal correction (OSC); Principal component transform; 2D NMR spectra 1. Introduction Pattern recognition techniques have already attracted con- siderable attention for the purpose of classification or discrim- ination [1]. Earlier work from Rosenblat [2] has proposed the perceptron approach for the automatic learning of discriminant functions for pattern classification. In 1986, Rumelhart et al. [3] suggested a feedforward perceptron approach for learning discriminants from a group of samples with known identities of classes, and then classifying unknown input vectors into classes. Lippman [4] also demonstrated the role of multiple- layer perceptrons in forming boundaries between distributions for partitioning samples into various classes. Kohonen et al. [5] proposed the use of learning vector quantization (LVQ) for classifying samples with inherent class intersections. Subse- quently, a number of papers and textbooks [6–8] have discussed ∗ Corresponding author. E-mail address: rutledge@inapg.inra.fr (D.N. Rutledge). the classification of patterns using statistical or neural network approaches. PLS discriminant analysis (PLS-DA) is a linear regression method whereby the multivariate variables corresponding to the observations (spectral descriptors) are related to the class mem- bership for each sample [9]. The method is supervised, i.e. the class membership of the samples is included in the calculation. PLS-DA provides an easily understood graphical means of iden- tifying the spectral regions of difference between the classes and also allows statistical evaluation as to whether the differences between classes are significant. To develop the classification rules for unknown samples in real applications, PLS discrimi- nant analysis has been utilized as supervised learning [10,11]. Nuclear magnetic resonance (NMR) with two dimensions ( 1 H– 13 C) has been applied to the study of wine polyphenol contents. While polyphenols are almost ubiquitous secondary metabolites, their fine structures and their relative ratio closely depend on the producing plant (specific genetic inheritance of the vine). It has thus been established, in 1994, that vine varieties (c´ epages) are possibly authenticated by statistical multiparamet- 0003-2670/$ – see front matter © 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2005.11.015