How to pre-process Raman spectra for reliable and stable models? Original Research Article

Raman spectroscopy in combination with chemometrics is gaining more and more importance for answering biological questions. This results from the fact that Raman spectroscopy is non-invasive, marker-free and water is not corrupting Raman spectra significantly. However, Raman spectra contain despite Raman fingerprint information other contributions like fluorescence background, Gaussian noise, cosmic spikes and other effects dependent on experimental parameters, which have to be removed prior to the analysis, in order to ensure that the analysis is based on the Raman measurements and not on other effects. Here we present a comprehensive study of the influence of pre-processing procedures on statistical models. We will show that a large amount of possible and physically meaningful pre-processing procedures leads to bad results. Furthermore a method based on genetic algorithms (GAs) is introduced, which chooses the spectral pre-processing according to the carried out analysis task without trying all possible pre-processing approaches (grid-search). This was demonstrated for the two most common tasks, namely for a multivariate calibration model and for two classification models. However, the presented approach can be applied in general, if there is a computational measure, which can be optimized. The suggested GA procedure results in models, which have a higher precision and are more stable against corrupting effects.