Body odor classification by selecting optimal peaks of chemical compounds in GC–MS spectra using filtering approaches

Present study deals with the operation of filter based approaches for selection of an optimal subset of peaks of chemical compounds in gas chromatography (GC)-mass spectrometry (MS) spectra with the objective to robust classification modeling of human body odor. Particularly, we have employed four filter based approaches including CFS, Linear-correlation, Rank-correlation, and Relief, in the selection of significant peaks and compared their performance. Selected subsets have been validated for qualitative and quantitative classification of human body odor samples in principal component (PC) space. Filter schemes were validated by analyzing sixteen decisive odor data sets obtained through characterization of body odor samples by GC–MS in four different experiments. Every feature filtering method results in an optimal subset of peaks for each data set. Efficiency of a particular subset of peaks has been evaluated by using them in PC analysis, and thereafter on the basis of visual discrimination as well inter-class separation (b) and intra-class (a) separation in PC space. Few methods result in a common subset of peaks for some data sets, though the maximum value of b and a minimum value of a has been obtained for discrimination amongst body odor samples by using selected subsets of peaks compare to all peaks in spectra. Best human body odor class discrimination outcomes have been achieved by using peaks of chemical compounds selected by Relief and CFS filters.