An essential factor influencing the efficiency of the predictive models built with principal component analysis (PCA) is the quality of the data clustering revealed by the score plots. The sensitivity and selectivity of the class assignment are strongly influenced by the relative position of the clusters and by their dispersion. We are proposing a set of indicators inspired from analytical geometry that may be used for an objective quantitative assessment of the data clustering quality as well as a global clustering quality coefficient (GCQC) that is a measure of the overall predictive power of the PCA models. The use of these indicators for evaluating the efficiency of the PCA class assignment is illustrated by a comparative study performed for the identification of the preprocessing function that is generating the most efficient PCA system screening for amphetamines based on their GC-FTIR spectra. The GCQC ranking of the tested feature weights is explained based on estimated density distributions and validated by using quadratic discriminant analysis (QDA).

No abstract

We are presenting a pattern recognition analysis assessing the class identity recognition efficiency of a portable laser infrared sensor detecting controlled phenethylamines, i.e. the stimulant and hallucinogenic amphetamines, as well as ephedrines, which are their main precursors. The training set consists of laser infrared spectra of the later compounds and of negatives, which are randomly selected non-amphetamines. The spectra have been recorded in the spectral domain 1405 -1150 cm -1 , preprocessed with a w TE 2 Fisher discriminating function, and then subjected to Principal Component Analysis (PCA). The PCA scores have been used in order to build several pattern recognition systems designed to recognize the class identity of the targeted compounds, i.e. Cluster Analysis and Naive Bayesian Classifier. The detection efficiency obtained for these two systems is presented and discussed in detail.

No abstract

No abstract