Multivariate detection capability using a neural classifier for nonselective signals Luis A. Sarabia a, * , M. Cruz Ortiz b , M. Julia Arcos b , M. Sagrario Sa ´nchez a , Ana Herrero b , Silvia Sanllorente b a Department of Mathematics and Computer Science, Faculty of Sciences, University of Burgos, Pza. Misael Ban ˜uelos s/n, 09001 Burgos, Spain b Department of Chemistry, Faculty of Sciences, University of Burgos, Pza. Misael Ban ˜uelos s/n, 09001 Burgos, Spain Received 24 April 2001; accepted 5 October 2001 Abstract A new methodology is proposed based on a neural network to determine the detection capability of an analytical procedure, in complex matrices, with the evaluation of the probability of false detection, a, and false nondetection, b, according to the ISO norms. This methodology is designed for first or greater order signals for which there is currently no procedure with these characteristics, which makes it difficult to use these signals in analytical procedures standardized according to the ISO norm. The procedure consists of: (i) an experimental design suited to the increase in analyte to be detected from a threshold level; (ii) a homogenisation of the multivariate signals by a Piecewise Direct Standardization (PDS) transformation; (iii) the training of a neural network with stochastic learning, Genetic Inside Neural Network (GINN), which optimises a and b directly. The procedure was applied to the polarographic determination of Tl(I)/Pb(II) mixtures and indomethacin/tenoxicam mixtures. In the first case one can assure the detection of 1 mM (threshold: 12 mM) with a and b less than 5% for both metals. While for tenoxicam it is possible to detect less than 10% of 12 mM (threshold) with a < 10% and b < 5%, for indomethacin one can assure less than 10% of 86 mM (threshold) with a and b less than 5%. D 2002 Elsevier Science B.V. All rights reserved. Keywords: Chemometrics; Neural networks; Genetic algorithms; False detection; False nondetection; Minimum detectable net concentration; Multivariate/multicomponent analysis; Polarography; IUPAC; ISO 1. Introduction The study of the detection capability of an analyt- ical procedure should not be limited to the detection limit, which consists of verifying the null hypothesis: ‘‘There is no analyte in the sample’’ (H 0 : c = c 0 = 0), because in practice it may be more important to verify the null hypothesis: ‘‘The concentration of analyte in the sample is less than c 0 ’’ (H 0 : c V c 0 a 0). For example, laboratories which depend on the adminis- tration routinely apply official methods of analysis to decide whether the quantity of the analyte in question surpasses or not a threshold level, c 0 , established by legal norms. This is a detectability problem from the point of view of the ISO norms, but not a problem of 0169-7439/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved. PII:S0169-7439(01)00202-7 * Corresponding author. Tel.: +34-947-25-8829; fax: +34-947- 25-8831. E-mail address: lsarabia@ubu.es (L.A. Sarabia). www.elsevier.com/locate/chemometrics Chemometrics and Intelligent Laboratory Systems 61 (2002) 89 – 104