Desenvolvimento de técnicas de classificação supervisionada para dados químicos multivariados

Abstract

This dissertation is composed by a theoretical contribution about the development of supervised classification techniques for application using multivariate chemical data. For this, chemometric techniques based on quadratic discriminant analysis (QDA) and support vector machines (SVM) were built combined with principal component analysis (PCA), successive projections algorithm (SPA) and genetic algorithm (GA) for supervised classification using data reduction and feature selection. These techniques were employed in analyzing first-order data, composed by attenuated total reflectance Fourier transform infrared spectroscopy (ATRFTIR) and mass spectra obtained from liquid chromatography time of flight (LC/TOF) and surface-enhanced laser desorption/ionization time of flight (SELDI/TOF). ATR-FTIR data were used to differentiate two classes of fungus of Cryptococcus gene, whereas the mass spectra data was used to identify ovarian and prostate cancer in blood serum. In addition, new twodimensional discriminant analysis techniques based on principal component analysis linear discriminant analysis (2D-PCA-LDA), quadratic discriminant analysis (2D-PCA-QDA) and support vectors machine (2D-PCA-SVM) were developed for applications in second-order chemical data composed by excitation-emission matrices (EEM) molecular fluorescence of simulated and real samples. The results show that the developed techniques had better classification performance for both first and second-order data, with classification rates, sensitivity and specificity reaching values between 90 to 100%. Also, the developed twodimensional techniques had overall performance superior than traditional multivariate classification methods using unfolded data, showing its potential to other future analytical applications.