Support Vector Regression and Radial Basis Function Neural Networks Applied to Semi-quantitative Prediction of Rhubarbs

Methods for building near-infrared spectrometry (NIRS) calibration models and for predicting active constituents of rhubarb samples using principal components analysis (PCA) and support vector regression (SVR) were investigated. Principal component analysis was used to reduce the number of spectral variables. Radial basis function neural networks (RBFNNs), ridge regression RBFNNs (RRRBFNNs), and SVR were used to model and predict four classes of active constituents (anthraquinones, anthraquinone glucosides, stilbene glucosides, tannins and tannin derivatives) in rhubarb using the principal component scores of the first-derivative spectra. The results show that the prediction accuracy by SVR is better than the accuracy obtained from the RBFNNs and RRRBFNNs. Therefore, SVR is a promising method for semiquantitative prediction of active constituents in Chinese herbal medicine.