Overlapped moving windows followed by principal component analysis to extract information from chromatograms and application to classification analysis

Abstract

Variable generation from chromatograms is conveniently accomplished using unsupervised rather than manual techniques. With unsupervised techniques, there is no need for selecting a few peaks for manual integration and valuable information is quickly and efficiently collected. The generation of variables can be performed by using either peak searching or moving window (MW) strategies. With a MW approach, the peaks are ignored and many variables, only part of them carrying information, are generated. Thus, variable generation by MWs should be followed by data compression to generate the variables to be further used for classification or quantitation purposes. In this work, unsupervised processing of chromatograms was performed by MWs followed by principal component analysis (MW-PCA). The principal components were selected and used as input variables for class prediction by linear discriminant analysis (LDA). Both simulated and real chromatograms were used to study the influence of both the window width and the degree of overlapping of consecutive windows on the quality of the predictions (estimated as the percentage of correctly classified chromatograms by leave-one-out). Hence, the quality of the variables generated by MW-PCA was inferred. For this purpose, a series of chromatograms containing a few peaks and belonging to four classes or categories were generated. Extra virgin olive oils of four pure cultivars were also used. In this case, a gas chromatograph provided with a flame ionization detector was used to obtain the chromatograms of fatty acid methyl esters. Windows within a wide range of moderate widths provided similar performances, and the quality of the predictions always increased with the degree of overlapping of consecutive windows. Thus, for MW-PCA variable generation, moderately wide and largely overlapped windows are recommended.