CC BY
13B-PS-6
comparative study of multivariate analysis methods of blood raman spectra classification was performed
L. Bratchenko1, I. Bratchenko1, A. Lykina1, M. Komarova1, D. Artemyev1, O. Myakinin1, A. Moryatov2, I. Davydkin3, S. Kozlov2, V. Zakharov1
1Samara University, Department of Laser and Biotechnical Systems, Samara, Russian Federation
2Samara State Medical University, Department of Oncology, Samara, Russian Federation 3Samara State Medical University, Department and Clinic of Hospital Therapy, Samara, Russian Federation
Pathosis of the human body leads to changes in body fluids biochemical composition. Currently used biochemical analysis of body fluids are notable for low-informative value to identify certain localization of pathologies. An alternative to biochemical methods is analysis by optical methods. Raman spectroscopy allows for the evaluation of blood characteristics at the molecular level. Raman blood spectra are characterized by multicollinearity feature and presence of contain autofluorescent background and noises of different nature. Selection of methods for experimental data processing of blood spectra is crucial for obtaining statistically reliable information about a pathological process in the body. Therefore, in this paper we examine various approaches to multidimensional analysis of blood samples of various size and perform statistical processing of experimental data from Raman scattering of blood by Factor analysis, Logistic regression, Discriminant analysis, Classification tree, Projection to latent structures discriminant analysis (PLS-DA) and Soft independent modelling of class analogies (SIMCA) to discriminate blood samples according to the pathology type. The analysis of the discussed multivariate methods for processing blood spectra obtained by cost-effective Raman setup in a clinical setting demonstrates that 1) the PLS-DA method (sensitivity 0.75, specificity 0.81) turned out to be the most optimal approach to blood samples classification by cancer localization; 2) the most optimal approach for blood samples classification by the presence of hyperproteinemia is the logistic regression method (sensitivity 0.89, specificity 0.99). In general, the selected multivariate methods may be a reliable tool for analyzing the body fluids spectral characteristics.
