Computerized delimitation of odorant areas in gas-chromatography-olfactometry by kernel density estimation: Data processing on French white wines

{GC-O using the detection frequency method gives a list of odor events (OEs) where each OE is described by a linear retention index (LRI) and by the aromatic descriptor given by a human assessor. The aim of the experimenter is to gather OEs in a total olfactogram on which he tries to delimit odorant areas (OAs), then to compute each detection frequency. This paper proposes a computerized mathematical method based on kernel density estimation that makes up the total olfactogram as continuous and differentiable function from the OEs LRI only. The corresponding curve looks like a chromatogram, the peaks of which are potential OAs. The limits of an OA are the LRI of the two minima surrounding the peak. The method was applied on a big data set 18 white wines, 17 assessors, 13,037 OEs. A previous manual delimitation made by the experimenter was used as benchmark to test the quality of the rendition by the computed delimitation. A contingency table containing the numbers of OEs that belonged to both benchmark OAs and computed OAs was built. This table enabled to assess the quality of the global rendition (Tschuprow's T coefficients) and the quality of individual rendition of each benchmark OA. In order to define a suitable range of application, the kernel-based method was tested on sub-sets from the global dataset, by randomly drawing n wines out of 18 and p assessors out of 17. The method gave very satisfying results for at least n = 9 wines