D-optimal designs and N-way techniques to determine sulfathiazole in milk by molecular fluorescence spectroscopy

Abstract

The present work proposes an analytical procedure to determine sulfathiazole in milk by using molecular fluorescence spectroscopy. For this sulfonamide the European Union in Regulation 37/2010 has established a maximum residue limit in milk of 100 μg kg −1. The study includes the effect of six factors on the recovery of sulfathiazole. The factors are: (i) The one related to the matrix depending on the heat treatment of the milk (UHT, pasteurized); (ii) Those related to the protein precipitation step, namely the ratio between the volume of trichloroacetic acid (TCA) and milk, centrifugation speed and temperature; (iii) Those affecting the derivatization reaction: derivatization time and volume of fluorescamine. To do this, two chemometric tools are used together: a D-optimal design for studying the effect of the factors on the recovery of sulfathiazole, considerably reducing the number of needed experiments; and the second-order property of the PARAFAC (Parallel Factor Analysis) decomposition that avoids the need of fitting a new calibration model each time that the experimental conditions change. It has been found that the type of milk, the TCA:milk ratio and the volume of fluorescamine have significant effect on the response. The rest of factors and interactions are not significant. The best recovery is obtained with UHT milk, 4:6 rate for TCA:milk volumes and 40 μL of fluorescamine. In UHT milk, the mean recovery ( n = 5) in the optimal conditions is 88.7% (RSD = 12.4%). As some non-linear behaviour may occur when using fluorescence spectroscopy, the calibration model that relates the fluorescence spectra with the concentration is computed by a partial least squares regression and a multi-layer feed-forward neural network. In both cases, the proposed procedures have been validated according to Decision 2002/657/EC, concluding that the two are accurate although the calibration model built with the neural network has better figures of merit, the decision limit (CC α) for x 0 = 100 μg L −1 is 103.3 μg L −1 and the detection capability (CC β) is 106.5 μg L −1, with the probabilities of false noncompliance ( α) and false compliance ( β) equal to 5%.