Comparison of Multiple NIR Spectrometers for Detecting Low-Concentration Nitrogen-Based Adulteration in Protein Powders

Abstract

Protein adulteration is a common fraud in the food industry due to the high price of protein sources and their limited availability. Total nitrogen determination is the standard analytical technique for quality control, which is incapable of distinguishing between protein nitrogen and nitrogen from non-protein sources. Three benchtops and one handheld near-infrared spectrometer (NIRS) with different signal processing techniques (grating, Fourier transform, and MEM-micro-electro-mechanical system) were compared with detect adulteration in protein powders at low concentration levels. Whey, beef, and pea protein powders were mixed with a different combination and concentration of high nitrogen content compounds-namely melamine, urea, taurine, and glycine-resulting in a total of 819 samples. NIRS, combined with chemometric tools and various spectral preprocessing techniques, was used to predict adulterant concentrations, while the limit of detection (LOD) and limit of quantification (LOQ) were also assessed to further evaluate instrument performance. Out of all devices and measurement methods compared, the most accurate predictive models were built based on the dataset acquired with a grating benchtop spectrophotometer, reaching R²P values of 0.96 and proximating the 0.1% LOD for melamine and urea. Results imply the possibility of using NIRS combined with chemometrics as a generalized quality control tool for protein powders.

Keywords: chemometrics; food fraud; handheld NIRS; melamine; near-infrared spectroscopy; quality control; whey protein.

Figure 1

Average raw spectra of pure protein powders (A) and each pure adulterant recorded with the FT-NIR (MPA) instrument (B). Vertical markers refer to some of the more significant characteristic absorbance peaks of a given adulterant.

Figure 2

Average raw spectra of samples with the highest singular adulterant concentration (A) and their second derivatives (B) recorded with the FT-NIR (MPA) device. Vertical markers indicate previously identified prominent absorbance peaks visible in either of the panels.

Figure 3

PCA score plot for the MPA dataset (A) and the NIR-S-G1 (bag) (B) dataset, using the entire available spectral region for both devices. Savitzky–Golay filter with 2nd-order polynomial and 21 smoothing points and SNV were applied as pretreatments.

Figure 4

LDA classification score plots with confidence intervals (circles) and average accuracy results for benchtop sub-datasets. “Rec.” = average correct recognition of the calibration; “Pred.” = average correct prediction of the cross-validated model; “Pret.” = applied pretreatments; “SG 21” = Savitzky–Golay filter with 21 smoothing points.

Figure 5

LDA classification score plots with confidence intervals (circles) and average accuracy results for handheld sub-datasets. “Rec.” = average correct recognition of the calibration; “Pred.” = average correct prediction of the cross-validated model; “Pret.” = applied pretreatments; “SG 21” = Savitzky–Golay filter with 21 smoothing points.

Figure 6

Test-set validation to predict protein powder content based on data collected with the NIRS6500 instrument. A = Y-fit plot with whey protein as sub-dataset, B = Y-fit plot with beef protein as sub-dataset, and C = Y-fit plot with pea protein as sub-dataset; “D”, “E” and “F” refer to respective regression vectors.

Figure 7

Test-set validation to predict protein powder content based on data collected with the NIR-S-G1 handheld instrument while scanning through a plastic bag. A = Y-fit plot with whey protein as sub-dataset, B = Y-fit plot with beef protein as sub-dataset, and C = Y-fit plot with pea protein as sub-dataset; “D”, “E” and “F” refer to respective regression vectors.

Figure 8

Method of preparing adulterated protein powder samples and spectral datasets with multiple devices.