Detection of soft-refined oils in extra virgin olive oil using data fusion approaches for LC-MS, GC-IMS and FGC-Enose techniques: The winning synergy of GC-IMS and FGC-Enose

Extra virgin olive oil (EVOO) is frequently adulterated by mixing it with soft refined oils (SROO). The differentiation of EVOO from its blends with SROO is not possible with the most common approaches, and, for this reason, the discriminating power of liquid chromatography-high resolution mass spectrometry (LC-MS), gas chromatography ion mobility spectrometry (GC-IMS) and flash gas-chromatography electronic nose (FGCEnose) was examined previously. Here, the combination of the above-mentioned techniques for an improvement in classification power of the methods is explored. A total of 43 commercial EVOOs and 18 illegal mixtures of SROO with EVOO were previously analysed by LC-(+/-)MS, GC-IMS and FGC-Enose. Low-level and mid-level data fusion of the four datasets were performed. The merged unique fingerprints were submitted to partial least squared discriminant analysis (PLS-DA), and the extrapolated most informative variables were used to build support vector machine (SVM) classifiers. Statistical indicators were calculated and compared to find out the best classifier. The results of PLS-DA-SVM strategies on the combination of datasets demonstrated that, after low-level data fusion, the discriminatory capability of the two merged GC-based techniques was remarkably improved as compared to the individual techniques. This indicates that merging the datasets before PLS-DA better retrieves the most informative variables and, thus, enhances group separation and classification of unknowns. The combination of LC(+/-)MS datasets, both by mid-and low-level data fusion, did not show significant enhancement in terms of discrimination of EVOO from SROO as compared to the individual LC(+)MS matrix. The low-level combination of the four datasets (LC(+/-)MS, GC-IMS, FGC-Enose) was successful, although this laborious option is not a viable path in industry quality assurance. This study primarily provides new paths for the authentication of EVOO, taking advantage of merging multimodal LC-(+/-)MS, GC-IMS and FGC-Enose data, with consequent improvement in the performances of the classification models. The most promising results were achieved by the low-level data fusion of GC-IMS and FGC-Enose data.