In a previous study {Takagi, T. et al., Chem. Pharm. Bull., 52(12), 1427-1432 (2004)}, we applied a slightly revised neural Independent Component Analysis (ICA) for profiling illegally distributed methamphetamine. Using ICA and an hourglass type Hierarchical Neural Network (HNN), we obtained better classification results than by using Principal Component Analysis (PCA), CATegorical PCA (CATPCA) and the MultiDimensional Scaling method (MDS). The HNN is a nonlinear machine learning method, and the ICA applied in that study exhibited nonlinear characteristics. The results indicated that nonlinear analysis is more efficient than linear analysis for profiling confiscated methamphetamine. Consequently, in this study, we applied Self-Organizing Maps (SOMs) to impurity profiling of methamphetamine.
While SOM is currently a frequently employed nonlinear classification method, the ordinary SOM uses only that information contained by the winner neuron for classification and the information of other grid points is neglected. We therefore attempted to simultaneously utilize the information of loser neurons in order to avoid information loss. First, we visualized the resultant reference vectors using a contour map of each sample. Although considerable information can be visually compared using the SOM contour maps, metric comparisons are difficult. We therefore used MDS to construct a similarity matrix using the data of the resultant reference vectors to visualize metric data. To assess the results, we assumed that there are four synthetic routes (Nagai, Leuckart, Emde and reductive amination methods), and that each of these can be identified by comparing route-specific impurities.
View full abstract