Residual bilinearization. A useful tool providing PLS and NN with the second-order advantage

GOICOECHEA, HÉCTOR C; AC OLIVIERI,; MJ CULZONI,

Campinas

Congreso; 10th Internacional Conference on Chemometrics in Analytical Chemistry (CAC 2006); 2006

Partial least-squares (PLS) regression based on unfolded second- or higher-order data can be combined with residual bilinearization (RBL) to yield a flexible higher-order multivariate method showing the second-order advantage. Due to its latent variable structure, it can deal with data sets deviating from trilinearity, or showing linear dependency in one or more modes. The technique is known from ca. 15 years ago,1 but only recently captured the attention of analytical chemists.2,3 In this report, the results are presented for several experimental data sets, prepared in order to show the great potentiality of the PLS/RBL technique, in all cases in the presence of an interfering background requiring strict adherence to the second-order advantage. They were designed to present the following challenges to second-order algorithms: 1) variable analyte-background interactions, 2) linear dependency due to pH changes in one mode, 3) linear dependency due to a reaction kinetics in one mode, and 4) fluorescence inner filter effect. In all of these cases, deviations from the ideal trilinearity are likely to occur due to changes in component profiles from sample to sample.             Data set 1 involves two pharmaceuticals which are determined in complex biomedical samples (i.e., salicylate and tetracycline in human serum), where interactions between the analytes and the background matrix modifies the analyte fluorescent spectra, causing sample-to-sample spectral variations. PLS/RBL outperformed parallel factor analysis (PARAFAC) in the analysis of the less sensitive analyte tetracycline.             Data set 2  includes absorbance-pH measurements made by creating a double pH gradient within a flow injection system. These data were applied to resolve mixtures of four dyes employed in the manufacturing of juice powders. Linear dependency occurs in the pH mode, because of the relationship between analyte species at different pH values, and also in the spectral mode, due to severe collinearity between acid and basic absorption spectra. Only PLS/RBL (and its classical variant of bilinear least-squares, BLLS/RBL) could be used to resolve the analytical problem.             In Data set 3,  absorbance-time data were studied for the reaction between the antibiotic amoxicillin and copper(II) ions, in order to determine the former in human urine samples. Linear dependency in both the time and spectral modes precluded the use of multivariate curve resolution (MCR), while PLS/RBL provided better results than PARAFAC.             Data set 4 consists of fluorescent excitation-emission matrices measured for a mixture of fungicides, to be determined in water in the presence of other potentially interfering agrochemicals. Inner filter effect from one of the analytes over the other one precluded the use of PARAFAC, MCR and BLLS/RBL, leaving PLS/RBL as the only second-order tool to be successfully employed if the second-order advantage is required.             Finally, developments concerning the achievement of the second-order advantage from non-linear second-order signals will be discussed. A recent algorithm based on the combination of artificial neural networks (ANN) with RBL may be able to provide the second-order advantage in this instance. Acknowledgments. CONICET (Consejo Nacional de Investigaciones Cientificas y Tecnicas, Project No. PIP 5303), ANPCyT (Agencia Nacional de Promocion Cientifica y Tecnologica, Project No. PICT ), University of Rosario, and University of El Litoral are gratefully acknowledged for financial support.