Estimation of signal backgrounds on multivariate loadings improves model generation in face of complex variation in backgrounds and constituents

The present study is designed to understand further implications of using multivariate loadings for the correction of background signal, which has previously been shown to be highly reproducible even for very low quality signals. Singular value decomposition (SVD)-based background correction was compared with the traditional per-signal paradigm for a biomedical dataset to generate qualitative and quantitative models. The qualitative effect on a principal component analysis model and the quantitative effect on a partial least square regression model were assessed for these background correction methods. The chosen quantitative parameter was the concentration of a pathologically relevant protein modification, pentosidine. Of the approaches tested, the SVD-based paradigm provided the regression model with the highest correlations, highest accuracy (lowest standard error of prediction) and repeatability (lowest sampling error). Contrasted against the traditional approaches, it was determined that the improved accuracy and repeatability of the SVD-based approach arises from its ability to simultaneously handle very complex background shapes alongside the complex variation in biochemical species that resulted in Raman signals with incompatible baseline regions. A better understanding of the interaction of SVD-based baseline correction, and data will give the reader more insight into the potential applicability of the procedure for other datasets. Copyright (C) 2012 John Wiley & Sons, Ltd.