Comparison of different variable selection methods conducted on NIR transmission measurements on intact tablets

Near infrared (NIR) transmission spectroscopy is a promising method for fast quantitative measurements on pharmaceutical tablets, but there are still some problems to overcome in order to incorporate the technique as a control tool in tablet production. The main problem is the limited precision for multivariate calibrations based on NIR transmission data. The precision is affected by several factors, where one of the most important is which variable to include in the multivariate calibration model. In this work, four different methods for variable selection in partial least square (PLS) regression were studied and compared to a calibration made with manually selected wavelengths. The methods used were genetic algorithm (GA), iterative PLS (IPLS), uninformative variable elimination by PLS (UVE-PLS) and interactive variable selection for PLS (IVS-PLS). All methods improved the predictive abilities of the model compared to the model where the wavelengths were selected manually. For the data set used in this work, IVS-PLS and GA achieved the best results with improvements in prediction error by 20%, but further measurements and investigations have to be made before any general conclusion can be drawn. (C) 2003 Published by Elsevier B.V.