Chemometric classification of traditional Chinese medicines by their geographical origins using near-infrared reflectance spectra.

Some raw materials that have different places of production for the plant sources of the drugs Astragalus membranaceus and ginseng have been studied, based on their near-infrared reflectance spectra. The experimentally recorded spectra represent heavily ill-posed and highly correlative data sets. Three related methods, i.e. the Fisher linear discriminant analysis (FLDA), the ridge-type linear discriminant analysis (RLDA) and a newly proposed penalized ridge-type linear discriminant analysis (PRLDA), have been investigated. FLDA over-fits for the training objects of the two data sets to a high extent and is unstable for the predictive objects of the two data sets. RLDA shows obvious improvement in terms of over-fitting and unstability, but the stability for the predictive objects of the two data sets is too sensitive to their ridge-type penalized weights, tending to produce erroneous discrimination results. The proposed PRLDA can circumvent the two aforementioned problems with a large domain of penalized weights for correct discriminant analysis of the two data sets studied. The combination of the PRLDA method and near infrared reflectance spectroscopy can be adapted for the discrimination of the production places of plant sources of these drugs.