Data fusion of headspace gas-chromatography ion mobility spectrometry and flash gas-chromatography electronic nose volatile fingerprints to estimate the commercial categories of virgin olive oils

In recent years, different gas-chromatographic methods based on the determination of volatile compounds, combined with chemometrics, have been proposed as methods to support the olive oil Panel test in classifying samples into commercial categories (EV, extra virgin olive oil; V, virgin olive oil; L, lampante olive oil). A valid strategy is to merge the outcomes of different analytical sources, applying a data fusion. This approach may be useful to improve the efficiency of prediction and robustness of a model compared to the results obtained by individual screening methods. In this analysis, inputs obtained by HS-GC-IMS and FGC E-nose analyses of 246olive oil samples were elaborated to classify samples according to commercial categories (EV, V, or L). PLSDA models based on three (EV, V, and L) or two classes (EV vs noEV, L vs noL, EV vs V, and L vs V) were developed. Furthermore, two different data fusion strategies (low and mid fusion level) were tested. In the lowlevel fusion, data from the two sources were concatenated directly, while in the mid-level fusion, features extracted separately from each source were combined into a common data matrix. Regardless of the single data set or data fusion approach, the strategy based on two class PLS-DA models showed the best results in which the percentages obtained in test set validation (TSV) ranged from 77.8% to 86.7% (FGC E-nose) and from 75% to 89.6% (HS-GC-IMS). A clear increase of the percentage of correctly classified samples was reached adopting the data fusion strategy, especially for class V (low level data fusion: +16.6%; mid level data fusion: +12.5%) and EV (+12.0% for both data fusion levels). Comparing the two strategies, mid level data fusion showed the most effective performance for both techniques, HS-GC-IMS (8.3 +/- 6.4%) and FCG-E-nose (8.7 +/- 4.8%), compared to the low fusion level, in which average percentage increases of 5.3 +/- 2.7% and 6.4 +/- 5.6% were reported with respect to the results of HS-GC-IMS and FGC E-nose models, respectively. The highest increases were achieved for L vs V models for both data fusion strategies. These promising results suggest that the data fusion approach can be an option to enhance the predictive efficiency in classifying olive oil samples into three commercial categories, providing a more reliable method to support the Panel test compared to the use of the single techniques.