Development of Radial Basis Function Cascade Correlation Networks and Applications of Chemometric Techniques for Hyphenated Chromatography- Mass Spectrometry Analysis

A cascade correlation learning architecture has been devised for radial basis function neural networks. Cascade correlation furnishes incremental learning networks. The proposed algorithm was applied to three different datasets: a synthetic dataset and two chemical datasets. The synthetic dataset was used to test the novelty detection ability of the proposed network. In the chemical datasets, the growth regions of Italian olive oils were identified by their fatty acid profiles; mass spectra of polychlorobiphenyl compounds were classified by chlorine number. The prediction results by bootstrap Latin partition indicate the proposed neural network is useful for pattern recognition.;A discriminant based charge deconvolution analysis pipeline is proposed. The molecular weight determination (MoWeD) charge deconvolution method was applied directly to the discrimination rules obtained by the fuzzy rule-building expert system (FuRES) pattern classifier. This approach was demonstrated with synthetic electrospray ionization mass spectra. Identification of the tentative protein biomarkers by bacterial cell extracts of Salmonella enterica serovar typhimurium strains A1 and A19 by liquid chromatography--electrospray ionization-mass spectrometry (LC--ESI-MS) was also demonstrated. The data analysis time was reduced by applying this approach. Furthermore, this method was less affected by noise and baseline drift.;The gasoline and kerosene collected from different locations in the United States were identified by gas chromatography/mass spectrometry (GC/MS) followed by chemometric analysis. Classifications based on twoway profile and target component ratio were compared. The projected difference resolution (PDR) mapping was applied to measure the differences among the ignitable liquid (IL) samples by their GC/MS profiles quantitatively. FuRESs were applied to classify individual ILs. The FuRES models yielded correct classification rates greater than 90% for discriminating between samples. PDR mapping, a new method for characterizing complex data sets, was consistent with the FuRES classification result.