A chemometric approach to the environmental problem of predicting toxicity in contaminated sediments

Sediments can act as sinks of multiple chemicals that accumulate over time and could represent a potentially significant hazard to the ecosystem and human health. The assessment of sediment quality, which is a complex task, usually involves, in its initial steps, the measurement of sediment contamination, followed by tests for assessing toxicity. This work presents a chemometric approach to tackle the challenging issue of linking concentrations of chemicals to the potential for observing toxicity in sediments. Various methods were applied to large databases of field-collected chemical and biological effects data, including linear and quadratic discriminant analysis (LDA and QDA), partial least squares-discriminant analysis (PLS-DA), extended canonical variates analysis (ECVA), classification and regression trees (CART) and counter-propagation artificial neural networks (CP-ANN). LDA, QDA, PLS-DA and ECVA showed very similar and satisfactory performances, but the best results were obtained with CP-ANNs and CART. In any case, the developed models for predicting toxicity improved the classification performance compared to previous approaches, giving non-error rates (NERs) in the range of 76.0-97.4. Moreover, the exploration of the internal structure of the models, jointly with the application of variable selection techniques, allowed the study of the importance of the 16 chemical contaminants considered, emphasizing Cu, followed by Zn, as the most discriminating variables for predicting toxicity in the developed models. Copyright (C) 2009 John Wiley & Sons, Ltd.