PLS-DA wavenumber selection for the categorization of medicine samples based on multiple criteria

ATR-FTIR spectra may include a large number of noisy and correlated wavenumbers that tend to affect and reduce the performance of exploratory and classificatory multivariate techniques. We propose a method based on Partial Least Square Discriminant Analysis (PLS-DA) for identifying the relevant subsets of ATR-FTIR wavenumbers aimed at classifying Viagra and Cialis into authentic or fraudulent categories. In our propositions, the PLS-DA is applied to ATR-FTIR data, and four indices aimed at evaluating wavenumber importance are derived from PLS-DA parameters. Next, an iterative wavenumber elimination and classification procedure integrating PLS-DA and the proposed indices is carried out: the wavenumber with the smallest index is removed, and a new classification is performed using the remaining wavenumbers. The classification performance is assessed through multiple criteria, i.e., sensitivity, specificity and percent of wavenumbers retained; the recommended wavenumber subset is chosen based on the distance between the candidate subsets and a hypothetical ideal solution. The proposed method significantly reduced the percent of wavenumbers to be assessed, and slightly improved classification performance for Viagra and Cialis data when compared to classification on all the original wavenumbers.