| Abstract [eng] |
Fourier transform infrared (FT-IR) spectroscopy is a versatile, non-destructive and label-free vibrational spectroscopy method with great potential in biomedical research and diagnosis. One of the advantages of this method is that many data analysis methods used in spectral analysis provide the means to detect the most important spectral regions that relate to biological mechanisms under investigation. However, the validity of these spectral markers is not well understood because of the complexity of vibrational spectra of biological samples, as well as the fact that usually there are no reference biomarkers to compare to. Furthermore, the effect spectral preprocessing has on spectral marker extraction has not yet been investigated. The aim of this study is to evaluate the validity of FT-IR spectral biomarkers obtained by various statistical and machine learning methods and how they depend on spectral preprocessing. For this, two methods of spectral marker identification were tested: forward feature selection (FFS) wrapped with a linear discriminant classification model and extraction of loadings from a principal component analysis - linear discriminant classification (PCA-LDC) model. Two experimental FT-IR spectral datasets were used as examples, one containing spectra from pathological microorganisms and the another containing spectra from extracellular vesicles extracted from human blood of healthy individuals and myocardial infarction patients. Spectral markers extracted from spectra preprocessed by various means were compared and evaluated by their similarity, stability and spectral interpretation. The results show that spectral preprocessing has a major impact on spectral marker extraction results. Despite poorer stability and similarity, the FFS method is considered overall more reliable than PCA-LDC on the grounds of clearer spectral interpretation. The best FT-IR spectral preprocessing procedure for spectral marker extraction is believed to be spectral truncation to the fingerprint region + second order differentiation. |
