Rapid and reliable spectral variable selection for statistical calibrations based on PLS-regression vector choices

The optimum number of spectral variables necessary for analytical spectroscopy is still a subject of debate. For sensor applications using miniaturized instrumentation, a small set of significant wavelengths with robust predictive performance is especially appreciated. A fast procedure is proposed, based on pairwise selection of spectral variables suggested by the weights of the optimum PLS-regression vector. The performance of multiple linear regression models based on such choices was similar to, and sometimes improved on full spectrum based modeling. Several examples from clinical studies of blood substrate assays using attenuated total reflection infrared spectroscopy of biofluids are presented.