Automated identification of novel amphetamines using a pure neural network and neural networks coupled with principal component analysis

A pure neural network (NN) and several neural networks coupled with principal component analysis (PC-NN) have been developed in order to identify illicit amphetamines necessary in the investigation of drugs of abuse for epidemiological, clinical, and forensic purposes. The NN system has as input variables 260 spectral data, representing absorption intensities measured for each normalized infrared spectrum at 260 wavenumbers 10 cm−1 apart. In the case of PC-NN systems, the original spectral data (absorption intensities) have been compressed with the principal component analysis method (PCA), the scores of the principal components (PCs) being the inputs of these systems. We have built nine PC-NN systems, which have a different number of input variables: 3PCs, 4PCs, 5PCs, 6PCs, 7PCs, 8PCs, 9PCs, 10PCs and 15PCs. All systems are specialized to distinguish between stimulant amphetamines (class code M), hallucinogenic amphetamines (class code T) and nonamphetamines (class code N). We are now presenting a comparison of the validation results obtained for the NN system and for the best PC-NN system based on the scores of the first nine PCs (9PC-NN). The NN system correctly classifies all the positive samples, as opposed to the 9PC-NN system, which is characterized by a true positive rate (TP) of 90.91%. The true negative rate (TN) obtained for the first system (83.33%) is slightly higher than in the case of the later system (82.71%). Thus, the NN system is more sensitive and selective than the 9PC-NN system. We are also presenting a spectroscopic analysis of the false negative samples obtained in the case of 9PC-NN system.