A Feature Selection Strategy for the Development of a New Drug Sensing System

Abstract

In order to efficiently detect four drug precursor molecules in presence of interfering species and background air, using a EC-QCLPAS sensor operating in the mid-infrared region, a complex strategy of spectral response simulation has been developed. In this context, spectra of gases from literature databases have been collected, denoised by means of the Wavelet Transform and mixed together according to a concentration matrix, which was specifically designed to represent a comprehensive combination of possible realistic cases. To scale database spectra to the appropriate concentration levels, an ad-hoc algorithm based on a sigmoidal transfer function has been used. In this way the baseline shape and intensity is preserved. Afterwards, a preliminary wavelength selection has been carried out to exclude noisy regions. The optimal range has finally been defined by maximizing the classification efficiency for all the target gases by means of Partial Least Squares-Discriminant Analysis.