Description
Bee pollen is a natural product rich in many nutritional compounds, but it could be contaminated with pyrrolizidine alkaloids (PAs) that can cause acute and chronic toxicity in humans. These toxins has been detected by time-consuming and expensive analytical chemistry techniques. Therefore, NIRS could be used as an alternative for the detection of PAs contamination. For this purpose, NIR spectral data were collected in 60 dehydrated ground bee pollen samples, by using a Foss DS2500 scanning monochromator (Foss NIR-System, Hillerød, DK) and a portable NIR system (PoliSPECNIR, ITPhotonics, Breganze, IT) and PAs were analyzed by LC-MS/MS. Based on EFSA suggestions, samples were grouped as PAs amount (µg kg-1 bee pollen): absence (N-class), < 84 (L-class), > 84 (H-class). For both spectrometers, a canonical discriminant analysis (CDA) was carried out following a feature selection of the significant NIR spectral variables based on ANOVA procedure. The D2-Mahalanobis sorted by the CDA were used to build a confusion matrix to test the reliability of the models. To determine the confusion matrix, each sample was reassigned to a PA-class according to the shortest D2-Mahalanobis distance; and the discriminative performance was evaluated by using the Matthews correlation coefficient (MCC). The CDA showed very similar high discriminative capacity (Wilks’λ < 0.22; p < 0.001) for both NIR systems, expecially in the identification of samples of the N-class (MCC = 0.78) even if 14.3% of samples were recognized as a false negative. A lower discriminant performance was observed for L- and H-class with a low value of MCC for the portable compared to the Foss system (0.58 vs. 0.70). Moreover, 7.1% of H-samples were misclassified as N-class both in Foss and PoliSPECNIR. Findings suggest that the absence (N-class) of PAs could be reliable detected by NIR spectroscopy but discriminative models should be implemented analyzing additional contaminated bee pollens.
| Consider for full paper in JNIRS | No, thank you |
|---|
