Ana María Jiménez Carvelo

Vegetable oils are a widely consumed commodity, with virgin olive oil being the highest-priced, which makes it highly susceptible to fraudulent practices, such as mislabeling and adulteration. While conventional standard analytical methods for olive oil authentication are reliable, their practical application is restricted by high resource consumption and environmental concerns. To address these shortcomings, we proposed a rapid, cost-effective, and sustainable methodology for olive oil purity assessment, using 1H low-field nuclear magnetic resonance (1H LF-NMR) analytical fingerprinting merged with machine learning. 785 edible vegetable oil samples were analyzed in this study, including different commercial categories of virgin olive oils, various seed oils, and binary blends (extra virgin olive oils mixed with sunflower oils and pomace olive oils). The analytical strategy involved a two-stage evaluation: a preliminary screening to discriminate commercial virgin olive oils from other edible vegetable oils, followed by a fatty acid (FA) profile evaluation to verify compliance with EU regulations. In both cases, the partial least-squares (PLS) method was implemented as a benchmark, while support vector machine (SVM) was used to evaluate potential improvements through nonlinear strategies. For the first time, classification thresholds are selected based on a precision–sensitivity curve proposed in this study. For both classification and quantitation outcomes, a strategy for defining appropriate uncertainty bands was proposed to identify inconclusive samples. Moreover, distribution-free homoscedastic prediction intervals were incorporated to estimate the quantitation uncertainty. The SVM classification model achieved 99% accuracy, outperforming PLS. SVM also demonstrated a better quantitation performance. Furthermore, the whiteness evaluation according to white analytical chemistry (WAC) principles confirmed that the proposed methodology is truly sustainable and efficient. Findings suggest that this holistic approach provides a robust tool for routine quality control, enabling fast screening and reliable quantitation, with uncertainty assessed through homoscedastic prediction intervals to support evidence-based decision-making.

The utilisation of artificial intelligence (AI)-based methodologies has increased considerably for the extraction of analytical information from complex spectroscopic fingerprints. In this study, low-field nuclear magnetic resonance (LF-NMR) spectroscopy was combined with chemometric and machine learning approaches for the quantification of oleic acid in sunflower oils. A repeatability study demonstrated high instrumental stability and day-to-day reproducibility of the LF-NMR measurements, with a mean Weighted Similarity Index (WSI) of 0.954. A total of 100 laboratory-prepared binary mixtures of conventional and high-oleic sunflower oils were used to develop Partial Least Squares Regression (PLS-R) and Support Vector Regression (SVR) models. The reference oleic acid concentrations were obtained using a previously developed PLS-R model based on spatially offset Raman spectroscopy (SORS) fingerprints. Exploratory principal component analysis (PCA) confirmed that the 1H NMR fingerprints contained sufficient analytical information for predictive modelling. Although both models demonstrated excellent predictive performance and good generalisation ability, the SVR model outperformed PLS-R, providing lower prediction errors, improved statistical performance, and higher residual predictive deviation (RPD = 19.24 vs. 11.13) and range error ratio (RER = 46.71 vs. 26.93) values for an independent set of commercial sunflower oil samples analysed by gas chromatography–flame ionisation detection (GC-FID). Furthermore, a strong agreement was observed between the LF-NMR predictions and the reference values obtained using SORS and GC-FID. These findings demonstrate that LF-NMR fingerprinting combined with chemometric and machine learning approaches provides a rapid, robust, non-destructive, and cost-effective alternative for oleic acid quantification, with potential for quality control in the vegetable oil industry.