Cluster analysis evaluating the automated detection of drugs of abuse with a new hollow fiber based quantum cascade laser infrared spectrometer

Abstract

In this study we are assessing the effect of spectra preprocessing on the efficiency of a Principal Component Analysis (PCA) application designed to screen for stimulant and hallucinogenic amphetamines, as well as for ephedrines, which are their main precursors. The application has been developed for a new portable hollow fiber infrared spectrometer equipped with a quantum cascade laser (QCL) source of radiation. The challenge is that the class identity recognition must be done based only on the absorptions present in the very narrow spectral window of the QCL (UT8, 1405–1150 cm−1). The 2D score plots obtained with the first three principal components (PCs) indicate that spectra preprocessing increases significantly the distances between the clusters formed by the targeted classes of compounds, but also lowers the density of these clusters. Taking into account the legal implications of misclassifications, the potential overlap of the various clusters has been estimated by calculating their density distributions with a normal kernel function. The most effective detection may be obtained by using a 3D score plot built with the first three PCs. If so, even classes of drugs of abuse with extremely similar molecular structures, like stimulant amphetamines and ephedrines, may be successfully discriminated.