Penalized logistic regression for classification and feature selection with its application to detection of two official species of Ganoderma

Abstract

Two species of Ganoderma, Ganoderma lucidum (G. lucidum) and Ganoderma sinense (G. sinense) have been widely used as traditional Chinese herbal medicine for their high medicinal value. Recent studies show that the two species differ in levels of their main active compounds triterpenoids though both have antitumoral effects. An effective and simple analytical method using attenuated total reflection Fourier transform infrared (ATR-FTIR) spectroscopy to discriminate between the two species is of essential importance for its quality assurance and medicinal value estimation. In this study three penalized logistic regression models, weighted least absolute shrinkage and selection operator (Lasso), elastic net and weighted fusion, using ATR-FTIR spectroscopy have been explored for the purpose of classification and interpretation. The weighted fusion model incorporating spectral correlation structure allowed an automatic selection of a small number of spectral bands and achieved an excellent overall classification accuracy of 99% in discriminating spectra of G. lucidum from that of G. sinense. Its classification performance was superior to that of the weighted Lasso model and elastic net model. The automatic selection of informative spectral features results in substantial reduction in model complexity and improvement of classification accuracy, and it is particularly helpful for the quantitative interpretations of the major chemical constituents of Ganoderma regarding its anti-cancer effects.