Successive projections algorithm combined with uninformative variable elimination for spectral variable selection

Abstracts E–SPA method, successive projections algorithm (SPA) combined with uninformative variable elimination (UVE) is proposed as a novel variable selection approach for multivariate calibration. UVE is used to select informative variables, and SPA is followed to select variables with minimum redundant information from the informative variables. The proposed method was applied to near-infrared (NIR) reflectance data for analysis of nicotine in tobacco lamina and NIR transmission data for active pharmaceutical ingredient (API) in single tablet. On the aspect of elimination of uninformative variables, the effect of UVE using first derivative spectra was better than that of using raw spectra. In terms of variable selection, fewer variables with better performance were selected by UVE—SPA method than by direct SPA method. For NIR spectral analysis of nicotine in tobacco lamina, the number of variables selected from 3001 spectral variables reduced from 48 by direct SPA to 35 by UVE–SPA, and the root mean squared error of prediction set (RMSEP) of the corresponding MLR models decreased from 0.174 (%, mg/mg) to 0.160. For NIR spectral analysis of API in each tablet, the number of variables selected from 650 spectral variables reduced from 46 by direct SPA to 17 by UVE–SPA, and RMSEP of the corresponding multiple linear regression (MLR) models decreased from 0.842 (%, mg/mg) to 0.473. MLR model using variables selected by UVE–SPA had better prediction performance than full-spectrum partial least-squares (PLS) model, and comparable to PLS model of UVE.