Modelling and Classification of GC/IMS Breath Gas Measurements for Lozenges of Different Flavours

The composition of exhaled breath contains information about diet, oral hygiene and other environmental influences as well as on the state of health and on medications. Therefore, rapid and sensitive breath analysis would be a helpful tool, for example, for medical diagnosis or therapy control. Ion mobility spectrometry coupled with gas-chromatographic pre-separation (GC/IMS) meet those requirements and can be used to distinguish between healthy and diseased persons or to detect drug usage, for example, based on characteristic exhaled metabolites. So far, the detection of peaks in IMS measurements and the assignment of compounds is done manually and an automated procedure is urgently needed. In this article, we analyse breath gas measurements by GC/IMS from a volunteer having consumed lozenges of 12 different flavours. The IMS measurements are modelled along drift time with an additive model of unimodal regressions to describe each peak. The regressions are afterwards combined across all spectra and all datasets to determine typical peak locations and the respective heights of the peaks in each measurement are inferred. The obtained matrix of peak intensities is then used to classify the measurements into the 12 flavour groups using support vector machines. Since the true class labels are known, we can assess the mis-classification rate using cross-validation.