- Email: [email protected]
Methods for detection of pork adulteration in veal product based on FT-NIR spectroscopy for laboratory, industrial and on-site analysis
Accepted Manuscript Methods for detection of pork adulteration in veal product based on FT-NIR spectroscopy for laboratory, industrial and on-site analysis Matthias Schmutzler, Anel Beganovic, Gerhard Böhler, Christian W. Huck PII:
S0956-7135(15)00236-4
DOI:
10.1016/j.foodcont.2015.04.019
Reference:
JFCO 4416
To appear in:
Food Control
Received Date: 3 February 2015 Revised Date:
10 April 2015
Accepted Date: 14 April 2015
Please cite this article as: Schmutzler M., Beganovic A., Böhler G. & Huck C.W., Methods for detection of pork adulteration in veal product based on FT-NIR spectroscopy for laboratory, industrial and on-site analysis, Food Control (2015), doi: 10.1016/j.foodcont.2015.04.019. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT Methods for detection of pork adulteration in veal product based on FT-NIR spectroscopy for laboratory, industrial and on-site analysis
Matthias Schmutzler, Anel Beganovic, Gerhard Böhler, Christian W. Huck*
RI PT
Institute of Analytical Chemistry & Radiochemistry, CCB – Center for Chemistry and Biomedicine, Leopold-Franzens-University, Innrain 80-82, 6020 Innsbruck, Austria
*corresponding author Phone: +43 512 507 57304
SC
Fax: +43 512 507 57399
M AN U
eMail: [email protected]
Keywords: FT-NIR spectroscopy; pork; veal; handheld spectrometer; PCA-SVM
Abstract
AC C
EP
TE D
Three different methods for near infrared (NIR) based multivariate analyses were developed to reveal deliberate adulteration or accidental contamination of a pure veal product with pork and pork fat. More precise, methods for laboratory use of high performance Fourier transform-NIR (FT-NIR) desktop devices, methods suitable for industrial purpose like in- and on-line application with a fibre optic probe and methods applying a handheld spectrometer ready for on-site analyses were established. The methods were developed for the detection of pork adulteration in the meat and fat part of veal sausages. Therefore sausages were self-made based on a commercial veal product. Adulterations up to 50 % (in 10 % steps) with pork and pork fat were analysed, respectively. Principal component analyses (PCA) were developed for every setup with previous data pre-treatment steps including wavelength selection, scattering corrections and derivatives of the spectral data. PCA scores were used as input data for support vector machines (SVM) classification and validation. Advantages and disadvantages of the equipment were discussed and the limits of detection regarding the setups were determined. Measurements were also carried out directly through a polymer packaging of the samples and compared to measurements through quartz cuvettes. Meat and fat adulteration could be detected up to the lowest level of contamination (10 %) applying the laboratory setup and the industrial fibre optics setup, regarding measurements through quartz and polymer packaging. Analyses with the on-site setup led to successful separation up to the lowest degree of contamination (10 %, measurement through quartz cuvettes) regarding meat adulteration and up to 20 % and 40 % contamination regarding the fat adulteration performing measurements through quartz cuvettes and through polymer packaging, respectively.
1
ACCEPTED MANUSCRIPT 1. Introduction
RI PT
In the food sector it is essential that consumers can rely on descriptive data, ingredients and information provided by the producers according to their products. Market confidence is necessary, especially within the high-range price segment, to make it possible for producers and distributors to sell their premium products at fair prices. Thus food controls are imperative to check the apparent authenticity, to strengthen the trust of the consumer and to protect from harmful frauds. Therefore, and in respect of many other topics concerning food safety, many regulations were adopted according to the European Commission's food safety policy (European Commission, 2015; European Food Safety Authority, 2015). The latter topics are mainly summarized in Regulation (EC) No 178/2002 (the “General Food Law”) of the European Union (European Union law, 2002). However, there is no definition of “food fraud” in EU legislation. Thus, national laws of each EU member state have to be taken into account in case of the intention of producers to obtain an undue benefit.
M AN U
SC
Complexity of legislation and regulation often results in confusion and distrust by the consumer. Followed by adverse events like the horse meat scandal in 2013 (European Commission, 2013; Premanandh, 2013), that made headline news across Europe and even further afield, demonstrates the problem of meat-based product safety and clearly underlines the necessity of new and feasible analytical methods like presented in this work. In view of a rapidly increasing human population and a global demand for food resources for at least the next four decades, concerns regarding food quality are well justified (Godfray, et al., 2010) not only in the European Union, but all over the world.
AC C
EP
TE D
Representing one piece in the puzzle of comprehensive product quality control a quick analytical method to check the animal species of the meat used as raw material in processed meat products is a matter of particular interest. Such a method serves to verify the authenticity of the product. Furthermore, detection of deliberate adulteration or accidental contamination with cheaper or unwelcome meat becomes feasible. Thus, socio-religious, safety and confidence issues can be faced. Easy handling, high speed, cost efficiency and the possibility to make the measurement an automated repetitive task are of vital importance for an analytical method used in the food sector due to the multitude of samples and the need of a high throughput analysis. Especially for meat and meat-based products NIR spectroscopy has been proven to be a reliable and easy to handle tool to determine several parameters like fat (Windham, Lawrence, & Feldner, 2003; Wold, O'Farrell, Høy, & Tschudi, 2011), protein and water content (Lanza, 1983; Tøgersen, Isaksson, Nilsen, Bakker, & Hildrum, 1999), to name but a few. Moreover, the methods presented in this work could be a complement or even an alternative to PCRbased DNA analyses like reviewed by Rolf Meyer and Urs Candrian (Meyer & Candrian, 1996) or Chandrika Murugaiah et al. (Murugaiah, et al., 2009), with the additional advantages of the NIR technology. Thus, a non-destructive character, simple or no sample preparation, easy handling, quick analyses, simultaneous determination of different analytes and the possibility to use optical fibres are some key advantages of the NIR technology (Huang, Yu, Xu, & Ying, 2008; McClure, 1994; O’Brien, Hulse, Pfeifer, & Siesler, 2013; Ozaki, McClure, & Christy, 2006; Reich, 2005; Siesler, Ozaki, Kawata, & Heise, 2008; Tsenkova, et al., 2006). Moreover, with no need for any chemicals and a very low effort of energy NIR spectroscopy is real “green science”. On a second thought, the mere
2
ACCEPTED MANUSCRIPT awareness of the availability of such an analysis could restrain producers from deliberate adulterating food products.
RI PT
In respect of a fast technological progress in the field of NIR based analytical equipment, methods for detection of adulterated or contaminated food and feed products will be of particular interest for controlling reasons across the board (Huck, 2014). Wireless NIR devices (Hattori, Tajiri, Yonai, & Otsuka, 2013; Sorak, et al., 2012; Yao, Guo, Sheng, Zhang, & Zhu, 2014) as well as a continuous advancing miniaturization process (dos Santos, Lopo, Páscoa, & Lopes, 2013; Muehlemann, Haensse, & Wolf, 2008) will make NIR spectroscopy an indispensable analytical tool for food control in the near future.
AC C
EP
TE D
M AN U
SC
The aim of this work was to develop and to compare independent NIR based methods for detection of pork adulteration in a veal sausage suited for three different fields of use: First for laboratory application, second for industrial measurements and third for on-site analyses.
3
ACCEPTED MANUSCRIPT 2. Materials and methods
SC
RI PT
This work is focused on the comparison of three different NIR based analytical measurement setups, which are dedicated to reveal adulteration of processed meat products with cheap and unwelcome meat from other animal species. In this connection, we selected a commercial sausage of the assortment of a local supermarket chain as an example. Mentioned sausage is made from pure veal and additional veal fat. Sausages for analyses were adulterated with pork during manufacturing as described in section 2.1. Since most of the processed meat products also have additional fat, adulteration with pork fat was taken into account. All sausages were self-made in strict accordance with the producer´s guidance to keep as close as possible to a realistic scenario. The only step renounced was the sausage casing due to the high diversity of materials used for this purpose. Moreover it is recommended to perform the measurement on the cut face of the sausage or to use the inner material for analyses. Reference samples without adulteration and samples with increasing level of adulteration were prepared as listed in table 1. Attention should be paid to the fact, that all percentage values are related to the adulterated part only (meat part and fat part, respectively), whereas all remaining ingredients correspond to the original recipe.
EP
2.1. Samples
TE D
M AN U
All analyses were performed using three different NIR spectroscopic measurement devices: A desktop setup, suitable for laboratory use and ready for highest performance analyses. Instruments alike are used in modern state-owned laboratories and private institutes as well. The same spectrometer was equipped with a light-fibre coupled probe for analyses. With use of the light-fibre it is possible to separate the spectrometer and the location of measurement over distances up to several meters. Thus industrial installations with high flexibility and fully automatic operation are possible. Last but not least a portable handheld NIR spectrometer was used for quick and easy analyses. Such versatile spectrometers are ready to use for goods inward inspection on the one hand and spot tests directly in the shops on the other hand. Employing a portable spectrometer comes along with striking benefits but also several drawbacks. In this work, convenience using a handheld spectrometer will be faced with its lack of precision and reproducibility.
AC C
In this work, a pure veal sausage product, produced and distributed by the Spar Holding AG (Salzburg, Austria) named “Edelbrater” was used as a basis for analyses. The goal of the experiments was to detect adulterations with pork and pork fat in different percentage rates. Therefore, sausages (without sausage casing) were self-made according to the original “Edelbrater”-recipe. The sausages consist of 76.0 g veal, 14.4 g veal fat and 9.6 g water, various spices and other typical ingredients per 100 g. The ratio between meat and fat was kept constant for every sample. Also the amounts of all other ingredients were kept constant. The composition of the meat-part was changed in 10 % steps from original (100 % veal) to 50 % veal and 50 % pork. Samples for the detection of the adulteration with pork fat were made in the same manner changing the fat-part in 10 % steps. Resulting in a total of 12 basic samples listed in table 1. We only used original ingredients provided by the producer. All samples were vacuum-packed in a double layer polymer packaging (first layer: high density polyethylene, second layer: polycaprolactam) and stored in the freezer at -40 °C. In advance of the spectroscopic measurements, samples were equilibrated for 4 h at 22 °C in the absence of air. A number of sub-samples were taken from each basic sample depending on the measurement device 4
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
and setup as follows: A total of 72 sub-samples were prepared for measurements with the laboratory setup in quartz cuvettes. These amount of samples consists of 30 samples (6 sub-samples each adulteration level) with adulterated fat part, 30 samples (6 sub-samples each adulteration level) with adulterated meat part and 12 samples as genuine reference samples with no adulteration. Applying the industrial setup a total of 84 samples were prepared. Composed of 12 main-samples in polymer packaging (two genuine samples and one sample each adulteration level for meat and fat adulteration, respectively) and 72 sub-samples prepared in quartz cuvettes in an analogous manner as described before. All samples prepared for the industrial setup were also analysed with the on-site setup (handheld spectrometer).
5
ACCEPTED MANUSCRIPT Table 1: Composition of self-made samples for the detection of adulterations in a veal sausage with pork and pork fat, respectively.
Pork 0 10 20 30 40 50
Veal fat 100 90 80 70 60 50
AC C
EP
TE D
M AN U
SC
Veal 100 90 80 70 60 50
Adulteration with pork fat (%) Total meat is constant
6
Pork fat 0 10 20 30 40 50
RI PT
Adulteration with pork (%) Total fat is constant
ACCEPTED MANUSCRIPT 2.2. Instrumentation
AC C
EP
TE D
M AN U
SC
RI PT
Technical characteristics of the instrumental equipment used in this work are summarised in table 2.
7
ACCEPTED MANUSCRIPT Table 2: Technical characteristics of the equipment. Industrial setup NIRFlex N-500, Cell: Fiber Optic Solids
On-site setup microPHAZIR GP 4.0
BUCHI AG, Flawil, Switzerland
Light source Dispersive element
Tungsten light bulb TeO2
Tungsten light bulb TeO2
Detector type
InGaAs (temperature controlled) 12,500 - 4,000
InGaAs (temperature controlled) 12,500 - 4,000
Thermo Fisher Scientific Inc., Massachusetts, USA Tungsten light bulb Programmable microdiffraction grating InGaAs (cooled)
6,028 - 5,480 (meat) 6,028 - 5,480 (fat) 8 16
6,028 - 5,480 (meat) 5,784 - 5,736 (fat) 8 16
6,037 - 5,550 (meat) 6,037 - 5,576 (fat) 21 21
32 Approx. 60 seconds
32 Approx. 40 seconds 2.0 mm diameter (source beam) and 3.5 mm diameter (collector), total length 2m
6 <3 seconds
AC C
EP
TE D 8
6,267 - 4,173
SC
Available wavenumber range (cm-1) Wavenumber range used for analysis (cm-1) Best resolution (cm-1) Resolution for analysis (cm-1) Number of scans Measurement time Fibre optic
M AN U
Manufacturer
RI PT
Laboratory setup NIRFlex N-500, Cell: Solids, add-on: XL & Spinner BUCHI AG, Flawil, Switzerland
ACCEPTED MANUSCRIPT 2.2.1. Laboratory setup
RI PT
As a desktop instrument the Buchi NIRFlex N-500 FT-NIR spectrometer and the Solids measuring cell were utilized. The XL add-on in combination with the Spinner add-on allowed the rotation of the sample during the measurement. Samples were placed in quartz cuvettes with 34 mm diameter (Hellma GmbH & Co. KG, Müllheim, Germany). The NIR spectra were recorded in the diffusereflection mode with the NIRWare 1.4.3010 software package (BUCHI AG, Flawil, Switzerland). All spectra were recorded with an absolute wavenumber accuracy of ±2 cm−1 and a relative reproducibility of 2.0 cm−1. For details in resolution, wavelength region and number of scans please refer to table 2. Internal and external reference measurements were repeated every 2 sub-samples (external reference against a Spectralon assembled reference cap). 2.2.2. Industrial setup
2.2.3. On-site setup
TE D
M AN U
SC
In order to simulate a practical industrial approach, the Buchi NIRFlex N-500 FT-NIR spectrometer was utilized with the Fiber Optic Solids measuring cell. This cell is equipped with a bifurcated fibre optic probe (for details please refer to table 2). To avoid direct reflection intake, the fibre optic probe was placed at a 70° angle to the level. Spectra were recorded in diffuse reflection mode. To obtain comparable spectra with both the laboratory and the industrial setup, instrument settings, spectral resolution, spectral range, temperature and number of scans were chosen in an equal manner as described in table 2. Internal and external reference measurements were repeated every 2 subsamples. 6 sub-samples were prepared in quartz cuvettes for every adulteration level. All subsamples were measured at 6 different random spots through the bottom of the quartz cuvettes. Thus, 36 spectra for each adulteration level were recorded with the fibre optic setup. In addition 6 measurements on different random spots were performed directly through the polymer packaging of the main samples for all adulteration levels.
AC C
EP
To demonstrate the possibility of rapid on-site measurements a handheld FT-NIR spectrometer (microPHAZIR GP 4.0, Thermo Fisher Scientific Inc., Massachusetts, USA) was used as a third alternative. Diffuse reflectance spectra were recorded in the wavelength ranges described in table 2. With a weight of only 1.25 kg, a common tungsten light bulb as the light source and a measuring time less than 3 seconds, the handheld instrument can be readily utilized for on-site analyses. 36 spectra were recorded for each adulteration level through quartz cuvettes as described in section 2.2.2. by placing the cuvettes directly on top of the handheld spectrometer. Additional measurements through polymer packaging were carried out by pressing the handheld spectrometer in contact with the polymer packs. 6 measurements on different random spots were performed for each adulteration level. The temperature was kept constant at 22 °C room temperature during all measurements. 2.3. Multivariate data analysis All data pre-treatments, PCA models and SVM classifications were performed using the software The Unscrambler X Ver. 10.2 (CAMO Software, Oslo, Norway). 9
ACCEPTED MANUSCRIPT
RI PT
Due to the fact that different instrument setups and configurations have specific influences on the spectral data, it is necessary to develop a multivariate model for data analysis adjusted for the given situation. Also different external factors can cause the need of special data pre-treatment. Therefore, five different methods for analyses have been developed in this work (one method for each instrumental setup and two additional methods for the measurements through the polymer packaging) including data pre-treatment, wavelength selection and PCA settings. The optimization of the methods was based on the best clustering (no overlap, narrow clusters and good detachment). Systematic cross validation with 6 samples each segment was performed for all PCA methods.
SC
A set of common mathematical and scatter correction pre-treatments and combinations thereof were used as part of the optimization process: standard normal variate (SNV) transformation, multiplicative scatter correction (MSC), 1st / 2nd / 3rd derivative (Savitzky-Golay method), detrending, base line correction methods and normalizations. For detailed information about the used pretreatments and wavelengths of the input data for each PCA model please refer to the corresponding section 3.1 and 3.2.
TE D
M AN U
To compare the three different instrumental setups, PCA clustering models were optimized in order to achieve the following objectives: Differentiation between the genuine product and one adulteration level: genuine vs. 50 % adulteration, genuine vs. 40 % adulteration etc. down to genuine vs. 10 % adulteration. In addition to the six mentioned models, one PCA model was developed to achieve identification of all adulteration levels and the genuine product in a single two-dimensional PCA scores plot. The latter is considered a rather scientific objective since the percentage of adulteration of a specific food product will be of similar magnitude in a real scenario. However, such a PCA represents the most challenging one and can give an idea of the robustness and accuracy of the corresponding method.
AC C
EP
In order to quantify the results in a numeric way, several classical chemometric methods can be used such as the following regression and classification techniques: PCA-linear discriminant analysis (PCALDA), PCA-quadratic discriminant analysis (PCA-QDA) (Wu, Massart, & De Jong, 1997), soft independent modelling of class analogy (SIMCA) (Kowalski, 1984), partial least squares (PLS) (Kowalski, 1984; Martens, 1989), etc. In this work another method: SVM, was chosen (Cristianini & Shawe-Taylor, 2000). Based on a supervised machine learning procedure, this technique is perfectly suited for multivariate data like NIR spectra (Boser, Guyon, & Vapnik, 1992; Vapnik, 1998). Furthermore SVM has been proven to be an accurate and reliable method, often superior to other classification methods in the field of spectroscopic analyses (Balabin & Lomakina, 2011; Chauchard, Cogdill, Roussel, Roger, & Bellon-Maurel, 2004; Pierna, Baeten, Renier, Cogdill, & Dardenne, 2004). PCA scores were used as input data for SVM classification and validation. In order to optimize the predictive performance of the classification process, the critical parameters C (trade-off regularization between minimum error and maximum margin) and γ (Gaussian function width) were calculated applying a “grid search” algorithm like it was proposed by Chih-Chung Chang & Chih-Jen Lin (Chih-Chung & Chih-Jen, 2001). Radial basis functions were used as kernel types for all classifications in this work. The performance of the classification is expressed in percentage of correctly classified samples. Thus SVM classification serves as an objective computational evaluation of the achieved PCA results.
10
ACCEPTED MANUSCRIPT 3. Results 3.1. Detection of meat adulteration
M AN U
SC
RI PT
After completion of the optimization process of the spectral data it was found the 2nd derivative (Savitzky-Golay method, 2nd polynomial order, 3 smoothing points) and the combination of the latter with SNV transformation to be the best data pre-treatments for the laboratory setup and the other setups, respectively. Reduction of the wavenumber region used for the analyses improved clustering significantly. Thus, original spectra enclosing 1,501 data points (wavenumbers from 10,000 - 4,000 cm-1) could be narrowed to a wavenumber region from 6,028 - 5,480 cm-1 enclosing only 548 data points for the laboratory and the industrial setups. Spectra derived from the on-site instrument were narrowed to a similar region from 6,037 - 5,550 cm-1. As a consequence, redundant and deceptive influences could be eliminated from the spectra, whereas pertinent information was highlighted. Increasing contamination of the samples results in an increase of the amplitudes of the 2nd derivative in the mentioned wavenumber regions. Close connection between the signal intensity and the level of adulteration at 5,940 cm-1 (1,683 nm), 5,908 cm-1 (1,693 nm), 5,892 cm-1 (1,697 nm), 5,868 cm-1 ( 1,704 nm), 5,776 cm-1 (1,731 nm), 5,756 cm-1 (1,737 nm), 5,668 cm-1 (1,764 nm), 5,648 cm-1 (1,770 nm) and 5,492 cm-1 (1,821 nm) are shown in figure 1.
λ=1737 nm
λ=1697 nm
0.001
λ=1770 nm
λ=1817 nm
λ=1704 nm λ=1693 nm
-0.001
AC C
-0.002
-0.003
λ=1821 nm
TE D
0
λ=1764 nm
EP
Intensity (arbitrary units)
λ=1683 nm
6000
5900
adulteration (%)
λ=1731 nm
5800 5700 Wavenumber (cm-1) 50
40
30
20
5600 10
5500
genuine
Figure 1: 2nd derivative spectra (6,028 – 5,480 cm-1) measured with the laboratory setup as a function of the adulteration of the veal product with pork. Adulteration levels from genuine (no adulteration) up to 50 % (in 10 % steps).
The genuine product could be successfully identified by all three experimental setups performing measurements through quartz cuvettes. Adequate PCA clustering and SVM classification accuracy of 11
ACCEPTED MANUSCRIPT 100 % was reached. Furthermore, every product with pork adulteration between 20 % and 50 % corresponding to 80 % and 50 % veal content, respectively, was successfully identified/classified in 10 % steps by applying the different instrumental alternative with the polymer packaging.
RI PT
Sausages with the lowest percentage of adulteration consist of 90 % veal and 10 % pork according to the sample set. These samples could be successfully identified with the laboratory setup and the industrial setup using quartz cuvettes and also performing measurements directly through the polymer packaging. In the latter case the accuracy of classification decreased by 8.3 % (validation samples) due to the influences of the polymer packaging. Nonetheless, the resulting classification accuracy of 91.7 % was proven to be successful.
AC C
EP
TE D
M AN U
SC
With the handheld device it was possible to discriminate unadulterated samples from samples with an adulteration level of 10 % using quartz cuvettes. However it was not possible to differentiate theses samples performing the measurement directly through the polymer packaging due to losses in signal intensity and resulting overlap of the PCA clusters. SVM classification accuracy declined to 83.3 %. The results are summarized in table 3.
12
ACCEPTED MANUSCRIPT Table 3: SVM classification and PCA discrimination objectives for the detection of meat adulteration by the laboratory, industrial and on-site instrumental setups. Industrial setup
On-site setup
through quartz
through plastic package
through quartz
through plastic package
PCA clustering
yes
yes
yes
yes
yes
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering
100
100
100
100
yes
yes
100
100
100
100
Option
Genuine vs. 30 % adulteration
Genuine vs. 20 % adulteration
Genuine vs. 10 % adulteration
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%)
AC C
EP
Classification of all adulteration levels in one PCA/SVM
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering
13
yes 100 100
100
100
100
100
100
100
yes
yes
yes
100
100
100
100
100
100
yes
yes
yes
yes
100
100
100
100
100
100
100
M AN U
Genuine vs. 40 % adulteration
100
yes
yes
yes
yes
yes
100
100
100
100
100
100
100
100
100
100
yes
yes
yes
yes
no
100
100
91.7
100
83.3
100
100
91.7
100
83.3
yes
yes
no
no
no
94.4
91.7
58.3
77.8
41.7
91.7
86.1
58.3
75.0
36.1
TE D
Genuine vs. 50 % adulteration
RI PT
Laboratory setup through quartz
SC
Detection of meat adulteration
ACCEPTED MANUSCRIPT Clustering was performed using PCA methods with a maximum of 9 principal components, however only the first two principal components were used for interpretation regarding the laboratory setup and the industrial setup. For the data of the handheld spectrometer it was necessary to use up to 4 principal components for interpretation.
2.0e-4
SC
RI PT
Figure 2 illustrates the differentiations between the genuine product and all adulteration levels with the industrial setup, thus 5 independent PCA scores plots are depicted for comparison. Clear differentiations could be achieved for all levels. Spectra were recorded through quartz cuvettes with the fibre optic probe. One score represents 6 measurements of one sub-sample. 6 sub-samples were analysed for each adulteration level.
adulteration 50% vs. genuine
0
M AN U
-2.0e-4 2.0e-4 1.0e-4
adulteration 40% vs. genuine
0
adulteration 30% vs. genuine
TE D
2.0e-4 0 -2.0e-4 4.0e-4 2.0e-4 0
AC C
-2.0e-4
adulteration 20% vs. genuine
EP
PC-2 (1%)
-1.0e-4 4.0e-4
1.0e-4
adulteration 10% vs. genuine
0
-1.0e-4
-3.0e-3
-2.0e-3
-1.0e-3
1.0e-3
0
2.0e-3
3.0e-3
4.0e-3
PC-1 (97%) adulteration (%)
50
40
30
20
10
genuine
Figure 2: Industrial setup: PCA scores plots for the detection of adulterated meat in veal sausages. All adulteration levels and the genuine product could be successfully discriminated.
14
ACCEPTED MANUSCRIPT
RI PT
As mentioned already, the final objective (to cluster all adulteration levels in one single PCA) embodies a rather scientific challenge, but it can be understood as a proof for quality and limitations of each method. The final objective could be achieved applying the laboratory setup (shown in figure 3) and the industrial setup for measurements through quartz cuvettes with a classification accuracy of 91.7 and 86.1 % (validation samples), respectively. However, measurements with the fibre optic probe directly through the polymer packaging obtained results showing some overlap of the 40 %, 30 %, and 20 % adulteration levels and led to broadened clusters in general. The spectra measured with the handheld spectrometer could not be represented in a satisfactory manner in a single PCA plot for all degrees of contamination.
SC
0.0003
0.0001
M AN U
PC-2 (2%)
0.0002
0 -0.0001 -0.0002
-0.003
-0.002
-0.001
0
0.001
0.002
TE D
PC-1 (92%) adulteration (%)
50
40
30
20
10
genuine
EP
Figure 3: Laboratory setup: PCA scores plot of all adulteration levels and the genuine product for pork detection in veal sausages. Adulteration of the product from genuine up to 50 % in 10 % steps.
AC C
3.2. Detection of fat adulteration
For this purpose spectral data, pre-treated with SNV transformation in combination with the 1st derivative (Savitzky-Golay method, 2nd polynomial order, 3 smoothing points), were used for clustering with PCA for all three instrumental setups. Best results were achieved using three different wavenumber regions for the three different setups: 6,028 - 5,480 cm-1 for the laboratory setup, 5,784 - 5,736 cm-1 for industrial setup and measurement through quartz cuvettes, 6,028 - 5,480 cm-1 for measurements through the polymer packaging and the region 6,037 - 5,576 cm-1 for all measurements with the handheld spectrometer. Results of the detection of the fat adulteration are summarized in table 4.
15
ACCEPTED MANUSCRIPT Table 4: SVM classification and PCA discrimination objectives for the detection of fat adulteration by the laboratory, industrial and on-site instrumental setups. Industrial setup
On-site setup
through quartz
through plastic package
through quartz
through plastic package
PCA clustering
yes
yes
yes
yes
yes
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering
100
100
100
100
yes
yes
100
100
100
100
Option
Genuine vs. 30 % adulteration
Genuine vs. 20 % adulteration
Genuine vs. 10 % adulteration
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%)
AC C
EP
Classification of all adulteration levels in one PCA/SVM
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering
16
yes 100 100
100
100
100
100
100
91.7
yes
yes
yes
100
100
100
100
100
91.7
yes
yes
yes
no
100
100
100
83.3
100
100
83.3
M AN U
Genuine vs. 40 % adulteration
TE D
Genuine vs. 50 % adulteration
RI PT
Laboratory setup through quartz
SC
Detection of fat adulteration
100
yes
yes
yes
yes
no
100
100
100
100
83.3
100
100
100
100
75.0
yes
yes
yes
no
no
100
100
100
83.3
75.0
100
100
100
83.3
75.0
no
no
no
no
no
88.9
69.4
50.0
66.7
-
88.9
61.1
50.0
63.9
-
ACCEPTED MANUSCRIPT All PCA models were performed with a maximum of 9 principal components, however only the first 3 principal components were used for interpretation.
RI PT
Differentiation between the genuine product and samples with adulteration levels from 20 % up to 50 % was possible applying PCA models with one and two principal components regarding the laboratory and the industrial setups. Three principal components were necessary for several models applying the on-site setup. Samples with 10 % adulteration of the fat part could be separated from the genuine products applying the laboratory and the industrial setups. In that case measurements through polymer packaging also allow successful clustering. Thus classification could be performed with 100 % correct discrimination of all validation samples.
0
-0.01
AC C
-0.02
EP
PC-2 (5%)
0.01
TE D
M AN U
SC
Using the handheld spectrometer, it was not possible to reach sufficient clustering of the samples with 10 % adulteration through quartz cuvettes. A classification accuracy of 83.3 % was not satisfactory. Thus for analyses of fat adulteration, carried out with the on-site setup, the limit of the method was reached between the adulteration levels of 10 % and 20 %. Figure 4 illustrates the discrimination of the genuine product and samples with an adulteration level of 20 % for the measurements through quartz cuvettes. Performing measurements through the polymer packaging results in a shift of the limit of detection to even higher contamination. Thus samples with 30 % (or lower) adulteration could not be classified with more than 83.3 % correct discrimination by SVM.
-0.03 -0.1
0
0.1 PC-1 (94%)
adulteration (%)
20
genuine
Figure 4: On-site setup: PCA plot of samples with 20 % adulteration of the fat part and the genuine product. Measurements through quartz cuvettes.
Detection of contamination with pork fat in the fat part of the veal sausages requires a more sensitive method than detection of meat contaminations, due to the fact that the fat part (14.4 % of 17
ACCEPTED MANUSCRIPT the sausage) is smaller than the meat part (76.0 % of the sausage). Thus 10 % adulteration means that only 1.4 % of the total sausage was modified.
AC C
EP
TE D
M AN U
SC
RI PT
However, it was not possible to perform sufficient PCA clustering of all fat adulteration levels and the genuine product in one single analysis for any instrumental setup and option. Therefore classification accuracy of 88.9 % correct discrimination of validation samples (laboratory setup) could not be exceeded.
18
ACCEPTED MANUSCRIPT 4. Discussion
RI PT
In this work we developed methods for NIR spectroscopy, based on multivariate analyses to reveal deliberate adulteration or accidental contamination with pork in a pure veal product with regard to practical applications in the field of food and feed control. More precise, methods for laboratory use, industrial purposes and on-site analyses were established. On the one hand for the detection of pork adulteration in the meat part of sausages, on the other hand for pork fat contamination of the fat part of the sausages. For this purpose sausages were chosen as samples because of their similarity in composition and processing of the meat and fat parts to many other products on the market like burger, ravioli stuffing, ground meat and more. 4.1.1. Laboratory setup
M AN U
SC
The laboratory setup embodies the most accurate measurement device with the highest reproducibility of the measurements performed in this study. However, it is the most inflexible instrument due to the restriction to use quartz cuvettes filled with the samples. Therefore sample preparation was necessary. All samples have to be measured one by one with cleaning steps beforehand. In comparison to the other experimental modes, the largest measurement times of approx. 60 seconds were required.
TE D
Features of the laboratory setup are: best technical characteristics like a temperature controlled detector, widest spectral range, excellent wavenumber accuracy and relative reproducibility (for details please refer to section 2.2.). Due to the use of the spinner add-on (rotation of the quartz cuvettes during the measurement) an excellent averaging process was performed on the early state of scanning the samples. Rotating the sample while measurement is generally known to be particularly advantageous using non-destructive techniques like NIR spectroscopy (Schmutzler & Huck, 2014; Xiaobo, Jiewen, Xingyi, & Yanxiao, 2007).
AC C
EP
Applying the laboratory setup for the measurements provided the highest resolution spectra with minimal noise or perturbation compared to all other instrumental setups. Thus, flexibility and possibilities in the process of developing methods for multivariate data analysis are increased. Additionally, all adulteration levels could be detected for meat and fat without reaching the limit of detection. Further studies have to be performed applying the laboratory setup to investigate the limit of detection for meat adulteration in processed food and feed products with samples of lesser distinctions. 4.1.2. Industrial setup
We used the same basis instrument and operational settings as for the laboratory setup, to highlight influences of the fibre optic probe. Hyphenating NIR spectrometers with fibre optic cables opens up a huge field of possible applications for in-, on- and at-line measurements. Process control in real-time with no need for sample preparation comes along with such setups (Blanco & Villarroya, 2002; Fernández-Novales, López, Sánchez, García, & Morales, 2008; Wolfbeis, 2000). Fully automatic measurements are possible resulting in the highest number of measurements per day compared to the other instrumental setups. Moreover, measurements are possible in a hostile environment with extreme temperature, pressure, corrosiveness, humidity or vibrations which are adverse to the conditions required for common optical spectrometers. Problems caused by such 19
ACCEPTED MANUSCRIPT conditions can be easily solved by separating the locations of measurement and the spectrometer over large distances by the use of a light fibre for interconnection (Rohe, Becker, Kölle, Eisenreich, & Eyerer, 1999; Vojinović, Cabral, & Fonseca, 2006). It is obvious, that such devices are the most expensive ones compared to all other instruments presented in this work due to the additional fibre optic probe. Installation and (re)calibration of fibre optic NIR devices should be managed by experts and can cause additional expense.
M AN U
SC
RI PT
An increase of intrinsic scattering effects of the fiber optic cable accompanied with losses of intensity due to a small cross-section of the light collector for intake of diffuse reflected radiation are disadvantages of setups like these. All measurements done in this work applying the industrial device were carried out by hand without rotating the sample. Thus, no averaging process while measuring the samples was possible. As a consequence, reproducibility is poorer and spectra are weaker in intensity compared to the laboratory setup. However even measurements through the polymer packaging of the samples were of high quality. Resulting in successful clustering down to the lowest adulteration levels by means of both, meat and fat adulteration. In summary, the industrial device provides a high-quality standard combined with peerless flexibility in terms of use. 4.1.3. On-site setup
TE D
Mobility is of course an important advantage of the on-site setup. In addition to that, lowest cost of all devices, easy use and operation are features that can be essential for non-specialized users. Organisations in the field of consumer protection or end-product control can extend their capabilities utilising a handheld spectrometer. A typical measurement can be done with a minimum of time exposure of less than 3 seconds, which is also of advantage.
AC C
EP
With regard to technical characteristics, the handheld spectrometer cannot cope with the laboratory and industrial setup, resulting in spectra of lower spectral resolution, signal intensity, signal-to-noise ratio and wavelength range (see section 2.2.). These disadvantages limit the possibilities of multivariate analyses and result in lower reproducibility and higher limits of detection. Also perturbations of any kind taking effect on the spectral data can superimpose information of interest if there is a lack of intensity. In conformity with the latter topic, measurements through the polymer packaging for example came along with significant influence on the spectra. Thus the limit of detection of the on-site setup was higher than for all other devices. Analysing meat and fat adulteration, the limit of detection of the handheld device was found between 20 % to 10 % contamination. However for a realistic scenario the accuracy of the handheld spectrometer should be absolutely satisfactory. As we expect, real adulterations of food or feed products will be of a higher percentage than 20 % of the meat and fat part. On closer examination of the percentage values of correct classification applying SVM, the disturbing influence of the polymer packaging can be pointed out. Furthermore limitations of the handheld device are clearly indicated due to decreasing classification accuracy over the course of all adulteration levels starting with a value of 91.7 % for 50 % adulteration and ending up at 75 % for the lowest degree of contamination. 4.2. Differences between the detection of meat and fat adulteration
20
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
As the results show, the analysis of meat adulteration was successful for lower levels of contamination compared to the attempts to reveal fat adulteration. This could be a consequence of two reasons: First, the meat part of the sausages was 76.0 % of the total weight whereas the fat part was only 14.4 % - that means 10 % adulteration results in either 7.6 % or 1.4 % for meat and fat, respectively. Thus, the method for revealing pork fat as part of the sausage has to be of higher sensitivity. Second, substitution of veal with pork results always in a minor change of the total fat content of the whole sausage, due to the fact that pork (muscles only) contains 1.86 g fat (average) with a variation of 1.00 to 2.80 g per 100 g of weight, while veal (muscles only) contains 0.81 g fat (average) (Scherz & Senser, 1994). As a consequence, changes in the meat composition are accompanied by changes in the total fat content. Therefore, in case of meat adulteration PCA clustering is based on two aspects: the protein profile and the total fat content. In case of fat adulteration only one aspect is changed: the fatty acids profile. However, total fat content and protein profile remains constant.
21
ACCEPTED MANUSCRIPT 5. Conclusions This work illustrates the suitability of NIR technology for adulteration detection in sausages. However, Results show different advantages and disadvantages for the three instrumental setups applied in this work. The question, which one is the best can´t be answered in general. It is mainly depending on the field of application, required precision, requested number of measurements per day, financial effort and other reasons like expertise of the user.
M AN U
SC
RI PT
Meat and fat adulteration could be revealed down to the lowest level of contamination (10 %) applying the laboratory setup and the industrial fibre optics setup. Furthermore, it was possible to measure directly through the polymer packaging of the samples. Analyses with the on-site setup (using a handheld NIR spectrometer) also led to successful separation between the contaminated samples and the genuine products down to the lowest degree of meat adulteration (10 %). However, measurements through the polymer packaging were possible down to a contamination level of 20 %. Measuring fat adulteration, the limits of the handheld spectrometer came to light more clearly. Contamination levels less than 20 % could not be separated from the unadulterated samples in a satisfactory manner. Nevertheless 20 % adulteration means only 2.8 % of the sausage in total was modified (adulteration indication concerns the fat part only).
AC C
EP
TE D
Processing the spectroscopic data, wavelength selection turned out as a critical task to eliminate perturbations of the spectral information. SVM classification has proven to be an optimal technique to evaluate PCA results in a numeric way and to be utilized for automated discrimination between genuine and adulterated samples.
22
ACCEPTED MANUSCRIPT Acknowledgements
AC C
EP
TE D
M AN U
SC
RI PT
The authors want to thank the Interreg IV initiative of the European Union (project “OriginAlp”) for financial support and the whole team of the project “OriginAlp” for good advice and support for all reasons. Special thanks goes to AMTirol, Ing. Alexander Walser and DI. Wendelin Juen. Last but not least we want to thank the Spar Holding AG (Salzburg, Austria) for supplying us with all the ingredients used for producing the “Edelbrater” veal sausage, the original recipe and for their confidence.
23
ACCEPTED MANUSCRIPT References
AC C
EP
TE D
M AN U
SC
RI PT
Balabin, R. M., & Lomakina, E. I. (2011). Support vector machine regression (SVR/LS-SVM)—an alternative to neural networks (ANN) for analytical chemistry. Comparison of nonlinear methods on near infrared (NIR) spectroscopy data. Analyst, 136(8), 1703-1712. Blanco, M., & Villarroya, I. (2002). NIR spectroscopy: a rapid-response analytical tool. TrAC Trends in Analytical Chemistry, 21(4), 240-250. Boser, B. E., Guyon, I. M., & Vapnik, V. N. (1992). A training algorithm for optimal margin classifiers. In Proceedings of the fifth annual workshop on Computational learning theory (pp. 144-152): ACM. Chauchard, F., Cogdill, R., Roussel, S., Roger, J., & Bellon-Maurel, V. (2004). Application of LS-SVM to non-linear phenomena in NIR spectroscopy: development of a robust and portable sensor for acidity prediction in grapes. Chemometrics and Intelligent Laboratory Systems, 71(2), 141150. Chih-Chung, C., & Chih-Jen, L. (2001). LIBSVM: A Library for Support Vector Machines. Retrieved from http://www.csie.ntu.edu.tw/~cjlin (Accessed 04.01.15). Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines and other kernel-based learning methods: Cambridge university press. dos Santos, C. A. T., Lopo, M., Páscoa, R. N., & Lopes, J. A. (2013). A Review on the applications of portable near-infrared spectrometers in the agro-food industry. Applied Spectroscopy, 67(11), 1215-1233. European Commission. (2013). Horse meat - questions and answers. Retrieved from http://ec.europa.eu/food/safety/official_controls/food_fraud/horse_meat/q-ans_en.htm (Accessed 28.03.15). European Commission. (2015). Food and Feed Safety. Retrieved from http://ec.europa.eu/food/food/index_en.htm (Accessed 28.03.15). European Food Safety Authority. (2015). About EFSA. Retrieved from http://www.efsa.europa.eu/en/aboutefsa.htm (Accessed 28.03.15). European Union law. (2002). Access to European Union law EUR-Lex - 32002R0178 - EN. Retrieved from http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32002R0178 (Accessed 20.03.15). Fernández-Novales, J., López, M.-I., Sánchez, M.-T., García, J.-A., & Morales, J. (2008). A feasibility study on the use of a miniature fiber optic NIR spectrometer for the prediction of volumic mass and reducing sugars in white wine fermentations. Journal of Food Engineering, 89(3), 325-329. Godfray, H. C. J., Beddington, J. R., Crute, I. R., Haddad, L., Lawrence, D., Muir, J. F., Pretty, J., Robinson, S., Thomas, S. M., & Toulmin, C. (2010). Food security: the challenge of feeding 9 billion people. Science, 327(5967), 812-818. Hattori, Y., Tajiri, Y., Yonai, T., & Otsuka, M. (2013). Real-time monitoring of pharmaceutical powder blending using a wireless near infrared spectrophotometer. NIR news, 24(4), 9-11. Huang, H., Yu, H., Xu, H., & Ying, Y. (2008). Near infrared spectroscopy for on/in-line monitoring of quality in foods and beverages: A review. Journal of Food Engineering, 87(3), 303-313. Huck, C. W. (2014). Advances of vibrational spectroscopic methods in phytomics and bioanalysis. Journal of Pharmaceutical and Bomedical Analysis, 87, 26-35. Kowalski, B. R. (1984). Chemometrics: mathematics and statistics in chemistry: Springer. Lanza, E. (1983). Determination of moisture, protein, fat, and calories in raw pork and beef by near infrared spectroscopy. Journal of Food Science, 48(2), 471-474. Martens, H. (1989). Multivariate calibration: John Wiley & Sons. McClure, W. F. (1994). Near-infrared spectroscopy. The giant is running strong. Analytical Chemistry, 66(1), 43A-53A. Meyer, R., & Candrian, U. (1996). PCR-based DNA analysis for the identification and characterization of food components. LWT-Food Science and Technology, 29(1), 1-9. 24
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
Muehlemann, T., Haensse, D., & Wolf, M. (2008). Wireless miniaturized in-vivo near infrared imaging. Optics Express, 16(14), 10323-10330. Murugaiah, C., Noor, Z. M., Mastakim, M., Bilung, L. M., Selamat, J., & Radu, S. (2009). Meat species identification and Halal authentication analysis using mitochondrial DNA. Meat Science, 83(1), 57-61. O’Brien, N., Hulse, C. A., Pfeifer, F., & Siesler, H. W. (2013). Near infrared spectroscopic authentication of seafood. Journal of Near Infrared Spectroscopy, 21, 299-305. Ozaki, Y., McClure, W. F., & Christy, A. A. (2006). Near-infrared Spectroscopy in Food Science and Technology: John Wiley & Sons. Pierna, J. A., Baeten, V., Renier, A. M., Cogdill, R. P., & Dardenne, P. (2004). Combination of support vector machines (SVM) and near-infrared (NIR) imaging spectroscopy for the detection of meat and bone meal (MBM) in compound feeds. Journal of Chemometrics, 18, 341-349. Premanandh, J. (2013). Horse meat scandal–A wake-up call for regulatory authorities. Food Control, 34(2), 568-569. Reich, G. (2005). Near-infrared spectroscopy and imaging: basic principles and pharmaceutical applications. Advanced drug delivery reviews, 57(8), 1109-1143. Rohe, T., Becker, W., Kölle, S., Eisenreich, N., & Eyerer, P. (1999). Near infrared (NIR) spectroscopy for in-line monitoring of polymer extrusion processes. Talanta, 50(2), 283-290. Scherz, H., & Senser, F. (1994). Food Composition and Nutrition Tables: Medpharm GmbH Scientific Publishers. Schmutzler, M., & Huck, C. W. (2014). Automatic sample rotation for simultaneous determination of geographical origin and quality characteristics of apples based on near infrared spectroscopy (NIRS). Vibrational Spectroscopy, 72, 97-104. Siesler, H. W., Ozaki, Y., Kawata, S., & Heise, H. M. (2008). Near-infrared spectroscopy: principles, instruments, applications: John Wiley & Sons. Sorak, D., Herberholz, L., Iwascek, S., Altinpinar, S., Pfeifer, F., & Siesler, H. W. (2012). New developments and applications of handheld Raman, mid-infrared, and near-infrared spectrometers. Applied Spectroscopy Reviews, 47(2), 83-115. Tøgersen, G., Isaksson, T., Nilsen, B., Bakker, E., & Hildrum, K. (1999). On-line NIR analysis of fat, water and protein in industrial scale ground meat batches. Meat Science, 51(1), 97-102. Tsenkova, R., Atanassova, S., Morita, H., Ikuta, K., Toyoda, K., Iordanova, I. K., & Hakogi, E. (2006). Near infrared spectra of cow's milk for milk quality evaluation: disease diagnosis and pathogen identification. Journal of Near Infrared Spectroscopy, 14(6), 363-370. Vapnik, V. (1998). Statistical learning theory (Vol. 2): Wiley New York. Vojinović, V., Cabral, J., & Fonseca, L. (2006). Real-time bioprocess monitoring: Part I: In situ sensors. Sensors and Actuators B: Chemical, 114(2), 1083-1091. Windham, W., Lawrence, K., & Feldner, P. (2003). Prediction of fat content in poultry meat by nearinfrared transmission analysis. The Journal of Applied Poultry Research, 12(1), 69-73. Wold, J., O'Farrell, M., Høy, M., & Tschudi, J. (2011). On-line determination and control of fat content in batches of beef trimmings by NIR imaging spectroscopy. Meat Science, 89(3), 317-324. Wolfbeis, O. S. (2000). Fiber-optic chemical sensors and biosensors. Analytical Chemistry, 72(12), 8190. Wu, W., Massart, D., & De Jong, S. (1997). Kernel-PCA algorithms for wide data Part II: Fast crossvalidation and application in classification of NIR data. Chemometrics and Intelligent Laboratory Systems, 37(2), 271-280. Xiaobo, Z., Jiewen, Z., Xingyi, H., & Yanxiao, L. (2007). Use of FT-NIR spectrometry in non-invasive measurements of soluble solid contents (SSC) of ‘Fuji’apple based on different PLS models. Chemometrics and Intelligent Laboratory Systems, 87(1), 43-51. Yao, P., Guo, W., Sheng, X., Zhang, D., & Zhu, X. (2014). A portable multi-channel wireless NIRS device for muscle activity real-time monitoring. In Engineering in Medicine and Biology Society (EMBC), 36th annual international conference of the IEEE (pp. 3719-3722): IEEE. 25
ACCEPTED MANUSCRIPT
EP
TE D
M AN U
SC
RI PT
Adulteration or contamination of veal sausages with unwelcome meat was detected. We compared FT-NIR devices and methods for food authentication control. Instruments for laboratory, industrial and on-site (handheld) analyses were used Principal component analysis and support vector machines classification were utilized.
AC C
-
S0956-7135(15)00236-4
DOI:
10.1016/j.foodcont.2015.04.019
Reference:
JFCO 4416
To appear in:
Food Control
Received Date: 3 February 2015 Revised Date:
10 April 2015
Accepted Date: 14 April 2015
Please cite this article as: Schmutzler M., Beganovic A., Böhler G. & Huck C.W., Methods for detection of pork adulteration in veal product based on FT-NIR spectroscopy for laboratory, industrial and on-site analysis, Food Control (2015), doi: 10.1016/j.foodcont.2015.04.019. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT Methods for detection of pork adulteration in veal product based on FT-NIR spectroscopy for laboratory, industrial and on-site analysis
Matthias Schmutzler, Anel Beganovic, Gerhard Böhler, Christian W. Huck*
RI PT
Institute of Analytical Chemistry & Radiochemistry, CCB – Center for Chemistry and Biomedicine, Leopold-Franzens-University, Innrain 80-82, 6020 Innsbruck, Austria
*corresponding author Phone: +43 512 507 57304
SC
Fax: +43 512 507 57399
M AN U
eMail: [email protected]
Keywords: FT-NIR spectroscopy; pork; veal; handheld spectrometer; PCA-SVM
Abstract
AC C
EP
TE D
Three different methods for near infrared (NIR) based multivariate analyses were developed to reveal deliberate adulteration or accidental contamination of a pure veal product with pork and pork fat. More precise, methods for laboratory use of high performance Fourier transform-NIR (FT-NIR) desktop devices, methods suitable for industrial purpose like in- and on-line application with a fibre optic probe and methods applying a handheld spectrometer ready for on-site analyses were established. The methods were developed for the detection of pork adulteration in the meat and fat part of veal sausages. Therefore sausages were self-made based on a commercial veal product. Adulterations up to 50 % (in 10 % steps) with pork and pork fat were analysed, respectively. Principal component analyses (PCA) were developed for every setup with previous data pre-treatment steps including wavelength selection, scattering corrections and derivatives of the spectral data. PCA scores were used as input data for support vector machines (SVM) classification and validation. Advantages and disadvantages of the equipment were discussed and the limits of detection regarding the setups were determined. Measurements were also carried out directly through a polymer packaging of the samples and compared to measurements through quartz cuvettes. Meat and fat adulteration could be detected up to the lowest level of contamination (10 %) applying the laboratory setup and the industrial fibre optics setup, regarding measurements through quartz and polymer packaging. Analyses with the on-site setup led to successful separation up to the lowest degree of contamination (10 %, measurement through quartz cuvettes) regarding meat adulteration and up to 20 % and 40 % contamination regarding the fat adulteration performing measurements through quartz cuvettes and through polymer packaging, respectively.
1
ACCEPTED MANUSCRIPT 1. Introduction
RI PT
In the food sector it is essential that consumers can rely on descriptive data, ingredients and information provided by the producers according to their products. Market confidence is necessary, especially within the high-range price segment, to make it possible for producers and distributors to sell their premium products at fair prices. Thus food controls are imperative to check the apparent authenticity, to strengthen the trust of the consumer and to protect from harmful frauds. Therefore, and in respect of many other topics concerning food safety, many regulations were adopted according to the European Commission's food safety policy (European Commission, 2015; European Food Safety Authority, 2015). The latter topics are mainly summarized in Regulation (EC) No 178/2002 (the “General Food Law”) of the European Union (European Union law, 2002). However, there is no definition of “food fraud” in EU legislation. Thus, national laws of each EU member state have to be taken into account in case of the intention of producers to obtain an undue benefit.
M AN U
SC
Complexity of legislation and regulation often results in confusion and distrust by the consumer. Followed by adverse events like the horse meat scandal in 2013 (European Commission, 2013; Premanandh, 2013), that made headline news across Europe and even further afield, demonstrates the problem of meat-based product safety and clearly underlines the necessity of new and feasible analytical methods like presented in this work. In view of a rapidly increasing human population and a global demand for food resources for at least the next four decades, concerns regarding food quality are well justified (Godfray, et al., 2010) not only in the European Union, but all over the world.
AC C
EP
TE D
Representing one piece in the puzzle of comprehensive product quality control a quick analytical method to check the animal species of the meat used as raw material in processed meat products is a matter of particular interest. Such a method serves to verify the authenticity of the product. Furthermore, detection of deliberate adulteration or accidental contamination with cheaper or unwelcome meat becomes feasible. Thus, socio-religious, safety and confidence issues can be faced. Easy handling, high speed, cost efficiency and the possibility to make the measurement an automated repetitive task are of vital importance for an analytical method used in the food sector due to the multitude of samples and the need of a high throughput analysis. Especially for meat and meat-based products NIR spectroscopy has been proven to be a reliable and easy to handle tool to determine several parameters like fat (Windham, Lawrence, & Feldner, 2003; Wold, O'Farrell, Høy, & Tschudi, 2011), protein and water content (Lanza, 1983; Tøgersen, Isaksson, Nilsen, Bakker, & Hildrum, 1999), to name but a few. Moreover, the methods presented in this work could be a complement or even an alternative to PCRbased DNA analyses like reviewed by Rolf Meyer and Urs Candrian (Meyer & Candrian, 1996) or Chandrika Murugaiah et al. (Murugaiah, et al., 2009), with the additional advantages of the NIR technology. Thus, a non-destructive character, simple or no sample preparation, easy handling, quick analyses, simultaneous determination of different analytes and the possibility to use optical fibres are some key advantages of the NIR technology (Huang, Yu, Xu, & Ying, 2008; McClure, 1994; O’Brien, Hulse, Pfeifer, & Siesler, 2013; Ozaki, McClure, & Christy, 2006; Reich, 2005; Siesler, Ozaki, Kawata, & Heise, 2008; Tsenkova, et al., 2006). Moreover, with no need for any chemicals and a very low effort of energy NIR spectroscopy is real “green science”. On a second thought, the mere
2
ACCEPTED MANUSCRIPT awareness of the availability of such an analysis could restrain producers from deliberate adulterating food products.
RI PT
In respect of a fast technological progress in the field of NIR based analytical equipment, methods for detection of adulterated or contaminated food and feed products will be of particular interest for controlling reasons across the board (Huck, 2014). Wireless NIR devices (Hattori, Tajiri, Yonai, & Otsuka, 2013; Sorak, et al., 2012; Yao, Guo, Sheng, Zhang, & Zhu, 2014) as well as a continuous advancing miniaturization process (dos Santos, Lopo, Páscoa, & Lopes, 2013; Muehlemann, Haensse, & Wolf, 2008) will make NIR spectroscopy an indispensable analytical tool for food control in the near future.
AC C
EP
TE D
M AN U
SC
The aim of this work was to develop and to compare independent NIR based methods for detection of pork adulteration in a veal sausage suited for three different fields of use: First for laboratory application, second for industrial measurements and third for on-site analyses.
3
ACCEPTED MANUSCRIPT 2. Materials and methods
SC
RI PT
This work is focused on the comparison of three different NIR based analytical measurement setups, which are dedicated to reveal adulteration of processed meat products with cheap and unwelcome meat from other animal species. In this connection, we selected a commercial sausage of the assortment of a local supermarket chain as an example. Mentioned sausage is made from pure veal and additional veal fat. Sausages for analyses were adulterated with pork during manufacturing as described in section 2.1. Since most of the processed meat products also have additional fat, adulteration with pork fat was taken into account. All sausages were self-made in strict accordance with the producer´s guidance to keep as close as possible to a realistic scenario. The only step renounced was the sausage casing due to the high diversity of materials used for this purpose. Moreover it is recommended to perform the measurement on the cut face of the sausage or to use the inner material for analyses. Reference samples without adulteration and samples with increasing level of adulteration were prepared as listed in table 1. Attention should be paid to the fact, that all percentage values are related to the adulterated part only (meat part and fat part, respectively), whereas all remaining ingredients correspond to the original recipe.
EP
2.1. Samples
TE D
M AN U
All analyses were performed using three different NIR spectroscopic measurement devices: A desktop setup, suitable for laboratory use and ready for highest performance analyses. Instruments alike are used in modern state-owned laboratories and private institutes as well. The same spectrometer was equipped with a light-fibre coupled probe for analyses. With use of the light-fibre it is possible to separate the spectrometer and the location of measurement over distances up to several meters. Thus industrial installations with high flexibility and fully automatic operation are possible. Last but not least a portable handheld NIR spectrometer was used for quick and easy analyses. Such versatile spectrometers are ready to use for goods inward inspection on the one hand and spot tests directly in the shops on the other hand. Employing a portable spectrometer comes along with striking benefits but also several drawbacks. In this work, convenience using a handheld spectrometer will be faced with its lack of precision and reproducibility.
AC C
In this work, a pure veal sausage product, produced and distributed by the Spar Holding AG (Salzburg, Austria) named “Edelbrater” was used as a basis for analyses. The goal of the experiments was to detect adulterations with pork and pork fat in different percentage rates. Therefore, sausages (without sausage casing) were self-made according to the original “Edelbrater”-recipe. The sausages consist of 76.0 g veal, 14.4 g veal fat and 9.6 g water, various spices and other typical ingredients per 100 g. The ratio between meat and fat was kept constant for every sample. Also the amounts of all other ingredients were kept constant. The composition of the meat-part was changed in 10 % steps from original (100 % veal) to 50 % veal and 50 % pork. Samples for the detection of the adulteration with pork fat were made in the same manner changing the fat-part in 10 % steps. Resulting in a total of 12 basic samples listed in table 1. We only used original ingredients provided by the producer. All samples were vacuum-packed in a double layer polymer packaging (first layer: high density polyethylene, second layer: polycaprolactam) and stored in the freezer at -40 °C. In advance of the spectroscopic measurements, samples were equilibrated for 4 h at 22 °C in the absence of air. A number of sub-samples were taken from each basic sample depending on the measurement device 4
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
and setup as follows: A total of 72 sub-samples were prepared for measurements with the laboratory setup in quartz cuvettes. These amount of samples consists of 30 samples (6 sub-samples each adulteration level) with adulterated fat part, 30 samples (6 sub-samples each adulteration level) with adulterated meat part and 12 samples as genuine reference samples with no adulteration. Applying the industrial setup a total of 84 samples were prepared. Composed of 12 main-samples in polymer packaging (two genuine samples and one sample each adulteration level for meat and fat adulteration, respectively) and 72 sub-samples prepared in quartz cuvettes in an analogous manner as described before. All samples prepared for the industrial setup were also analysed with the on-site setup (handheld spectrometer).
5
ACCEPTED MANUSCRIPT Table 1: Composition of self-made samples for the detection of adulterations in a veal sausage with pork and pork fat, respectively.
Pork 0 10 20 30 40 50
Veal fat 100 90 80 70 60 50
AC C
EP
TE D
M AN U
SC
Veal 100 90 80 70 60 50
Adulteration with pork fat (%) Total meat is constant
6
Pork fat 0 10 20 30 40 50
RI PT
Adulteration with pork (%) Total fat is constant
ACCEPTED MANUSCRIPT 2.2. Instrumentation
AC C
EP
TE D
M AN U
SC
RI PT
Technical characteristics of the instrumental equipment used in this work are summarised in table 2.
7
ACCEPTED MANUSCRIPT Table 2: Technical characteristics of the equipment. Industrial setup NIRFlex N-500, Cell: Fiber Optic Solids
On-site setup microPHAZIR GP 4.0
BUCHI AG, Flawil, Switzerland
Light source Dispersive element
Tungsten light bulb TeO2
Tungsten light bulb TeO2
Detector type
InGaAs (temperature controlled) 12,500 - 4,000
InGaAs (temperature controlled) 12,500 - 4,000
Thermo Fisher Scientific Inc., Massachusetts, USA Tungsten light bulb Programmable microdiffraction grating InGaAs (cooled)
6,028 - 5,480 (meat) 6,028 - 5,480 (fat) 8 16
6,028 - 5,480 (meat) 5,784 - 5,736 (fat) 8 16
6,037 - 5,550 (meat) 6,037 - 5,576 (fat) 21 21
32 Approx. 60 seconds
32 Approx. 40 seconds 2.0 mm diameter (source beam) and 3.5 mm diameter (collector), total length 2m
6 <3 seconds
AC C
EP
TE D 8
6,267 - 4,173
SC
Available wavenumber range (cm-1) Wavenumber range used for analysis (cm-1) Best resolution (cm-1) Resolution for analysis (cm-1) Number of scans Measurement time Fibre optic
M AN U
Manufacturer
RI PT
Laboratory setup NIRFlex N-500, Cell: Solids, add-on: XL & Spinner BUCHI AG, Flawil, Switzerland
ACCEPTED MANUSCRIPT 2.2.1. Laboratory setup
RI PT
As a desktop instrument the Buchi NIRFlex N-500 FT-NIR spectrometer and the Solids measuring cell were utilized. The XL add-on in combination with the Spinner add-on allowed the rotation of the sample during the measurement. Samples were placed in quartz cuvettes with 34 mm diameter (Hellma GmbH & Co. KG, Müllheim, Germany). The NIR spectra were recorded in the diffusereflection mode with the NIRWare 1.4.3010 software package (BUCHI AG, Flawil, Switzerland). All spectra were recorded with an absolute wavenumber accuracy of ±2 cm−1 and a relative reproducibility of 2.0 cm−1. For details in resolution, wavelength region and number of scans please refer to table 2. Internal and external reference measurements were repeated every 2 sub-samples (external reference against a Spectralon assembled reference cap). 2.2.2. Industrial setup
2.2.3. On-site setup
TE D
M AN U
SC
In order to simulate a practical industrial approach, the Buchi NIRFlex N-500 FT-NIR spectrometer was utilized with the Fiber Optic Solids measuring cell. This cell is equipped with a bifurcated fibre optic probe (for details please refer to table 2). To avoid direct reflection intake, the fibre optic probe was placed at a 70° angle to the level. Spectra were recorded in diffuse reflection mode. To obtain comparable spectra with both the laboratory and the industrial setup, instrument settings, spectral resolution, spectral range, temperature and number of scans were chosen in an equal manner as described in table 2. Internal and external reference measurements were repeated every 2 subsamples. 6 sub-samples were prepared in quartz cuvettes for every adulteration level. All subsamples were measured at 6 different random spots through the bottom of the quartz cuvettes. Thus, 36 spectra for each adulteration level were recorded with the fibre optic setup. In addition 6 measurements on different random spots were performed directly through the polymer packaging of the main samples for all adulteration levels.
AC C
EP
To demonstrate the possibility of rapid on-site measurements a handheld FT-NIR spectrometer (microPHAZIR GP 4.0, Thermo Fisher Scientific Inc., Massachusetts, USA) was used as a third alternative. Diffuse reflectance spectra were recorded in the wavelength ranges described in table 2. With a weight of only 1.25 kg, a common tungsten light bulb as the light source and a measuring time less than 3 seconds, the handheld instrument can be readily utilized for on-site analyses. 36 spectra were recorded for each adulteration level through quartz cuvettes as described in section 2.2.2. by placing the cuvettes directly on top of the handheld spectrometer. Additional measurements through polymer packaging were carried out by pressing the handheld spectrometer in contact with the polymer packs. 6 measurements on different random spots were performed for each adulteration level. The temperature was kept constant at 22 °C room temperature during all measurements. 2.3. Multivariate data analysis All data pre-treatments, PCA models and SVM classifications were performed using the software The Unscrambler X Ver. 10.2 (CAMO Software, Oslo, Norway). 9
ACCEPTED MANUSCRIPT
RI PT
Due to the fact that different instrument setups and configurations have specific influences on the spectral data, it is necessary to develop a multivariate model for data analysis adjusted for the given situation. Also different external factors can cause the need of special data pre-treatment. Therefore, five different methods for analyses have been developed in this work (one method for each instrumental setup and two additional methods for the measurements through the polymer packaging) including data pre-treatment, wavelength selection and PCA settings. The optimization of the methods was based on the best clustering (no overlap, narrow clusters and good detachment). Systematic cross validation with 6 samples each segment was performed for all PCA methods.
SC
A set of common mathematical and scatter correction pre-treatments and combinations thereof were used as part of the optimization process: standard normal variate (SNV) transformation, multiplicative scatter correction (MSC), 1st / 2nd / 3rd derivative (Savitzky-Golay method), detrending, base line correction methods and normalizations. For detailed information about the used pretreatments and wavelengths of the input data for each PCA model please refer to the corresponding section 3.1 and 3.2.
TE D
M AN U
To compare the three different instrumental setups, PCA clustering models were optimized in order to achieve the following objectives: Differentiation between the genuine product and one adulteration level: genuine vs. 50 % adulteration, genuine vs. 40 % adulteration etc. down to genuine vs. 10 % adulteration. In addition to the six mentioned models, one PCA model was developed to achieve identification of all adulteration levels and the genuine product in a single two-dimensional PCA scores plot. The latter is considered a rather scientific objective since the percentage of adulteration of a specific food product will be of similar magnitude in a real scenario. However, such a PCA represents the most challenging one and can give an idea of the robustness and accuracy of the corresponding method.
AC C
EP
In order to quantify the results in a numeric way, several classical chemometric methods can be used such as the following regression and classification techniques: PCA-linear discriminant analysis (PCALDA), PCA-quadratic discriminant analysis (PCA-QDA) (Wu, Massart, & De Jong, 1997), soft independent modelling of class analogy (SIMCA) (Kowalski, 1984), partial least squares (PLS) (Kowalski, 1984; Martens, 1989), etc. In this work another method: SVM, was chosen (Cristianini & Shawe-Taylor, 2000). Based on a supervised machine learning procedure, this technique is perfectly suited for multivariate data like NIR spectra (Boser, Guyon, & Vapnik, 1992; Vapnik, 1998). Furthermore SVM has been proven to be an accurate and reliable method, often superior to other classification methods in the field of spectroscopic analyses (Balabin & Lomakina, 2011; Chauchard, Cogdill, Roussel, Roger, & Bellon-Maurel, 2004; Pierna, Baeten, Renier, Cogdill, & Dardenne, 2004). PCA scores were used as input data for SVM classification and validation. In order to optimize the predictive performance of the classification process, the critical parameters C (trade-off regularization between minimum error and maximum margin) and γ (Gaussian function width) were calculated applying a “grid search” algorithm like it was proposed by Chih-Chung Chang & Chih-Jen Lin (Chih-Chung & Chih-Jen, 2001). Radial basis functions were used as kernel types for all classifications in this work. The performance of the classification is expressed in percentage of correctly classified samples. Thus SVM classification serves as an objective computational evaluation of the achieved PCA results.
10
ACCEPTED MANUSCRIPT 3. Results 3.1. Detection of meat adulteration
M AN U
SC
RI PT
After completion of the optimization process of the spectral data it was found the 2nd derivative (Savitzky-Golay method, 2nd polynomial order, 3 smoothing points) and the combination of the latter with SNV transformation to be the best data pre-treatments for the laboratory setup and the other setups, respectively. Reduction of the wavenumber region used for the analyses improved clustering significantly. Thus, original spectra enclosing 1,501 data points (wavenumbers from 10,000 - 4,000 cm-1) could be narrowed to a wavenumber region from 6,028 - 5,480 cm-1 enclosing only 548 data points for the laboratory and the industrial setups. Spectra derived from the on-site instrument were narrowed to a similar region from 6,037 - 5,550 cm-1. As a consequence, redundant and deceptive influences could be eliminated from the spectra, whereas pertinent information was highlighted. Increasing contamination of the samples results in an increase of the amplitudes of the 2nd derivative in the mentioned wavenumber regions. Close connection between the signal intensity and the level of adulteration at 5,940 cm-1 (1,683 nm), 5,908 cm-1 (1,693 nm), 5,892 cm-1 (1,697 nm), 5,868 cm-1 ( 1,704 nm), 5,776 cm-1 (1,731 nm), 5,756 cm-1 (1,737 nm), 5,668 cm-1 (1,764 nm), 5,648 cm-1 (1,770 nm) and 5,492 cm-1 (1,821 nm) are shown in figure 1.
λ=1737 nm
λ=1697 nm
0.001
λ=1770 nm
λ=1817 nm
λ=1704 nm λ=1693 nm
-0.001
AC C
-0.002
-0.003
λ=1821 nm
TE D
0
λ=1764 nm
EP
Intensity (arbitrary units)
λ=1683 nm
6000
5900
adulteration (%)
λ=1731 nm
5800 5700 Wavenumber (cm-1) 50
40
30
20
5600 10
5500
genuine
Figure 1: 2nd derivative spectra (6,028 – 5,480 cm-1) measured with the laboratory setup as a function of the adulteration of the veal product with pork. Adulteration levels from genuine (no adulteration) up to 50 % (in 10 % steps).
The genuine product could be successfully identified by all three experimental setups performing measurements through quartz cuvettes. Adequate PCA clustering and SVM classification accuracy of 11
ACCEPTED MANUSCRIPT 100 % was reached. Furthermore, every product with pork adulteration between 20 % and 50 % corresponding to 80 % and 50 % veal content, respectively, was successfully identified/classified in 10 % steps by applying the different instrumental alternative with the polymer packaging.
RI PT
Sausages with the lowest percentage of adulteration consist of 90 % veal and 10 % pork according to the sample set. These samples could be successfully identified with the laboratory setup and the industrial setup using quartz cuvettes and also performing measurements directly through the polymer packaging. In the latter case the accuracy of classification decreased by 8.3 % (validation samples) due to the influences of the polymer packaging. Nonetheless, the resulting classification accuracy of 91.7 % was proven to be successful.
AC C
EP
TE D
M AN U
SC
With the handheld device it was possible to discriminate unadulterated samples from samples with an adulteration level of 10 % using quartz cuvettes. However it was not possible to differentiate theses samples performing the measurement directly through the polymer packaging due to losses in signal intensity and resulting overlap of the PCA clusters. SVM classification accuracy declined to 83.3 %. The results are summarized in table 3.
12
ACCEPTED MANUSCRIPT Table 3: SVM classification and PCA discrimination objectives for the detection of meat adulteration by the laboratory, industrial and on-site instrumental setups. Industrial setup
On-site setup
through quartz
through plastic package
through quartz
through plastic package
PCA clustering
yes
yes
yes
yes
yes
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering
100
100
100
100
yes
yes
100
100
100
100
Option
Genuine vs. 30 % adulteration
Genuine vs. 20 % adulteration
Genuine vs. 10 % adulteration
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%)
AC C
EP
Classification of all adulteration levels in one PCA/SVM
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering
13
yes 100 100
100
100
100
100
100
100
yes
yes
yes
100
100
100
100
100
100
yes
yes
yes
yes
100
100
100
100
100
100
100
M AN U
Genuine vs. 40 % adulteration
100
yes
yes
yes
yes
yes
100
100
100
100
100
100
100
100
100
100
yes
yes
yes
yes
no
100
100
91.7
100
83.3
100
100
91.7
100
83.3
yes
yes
no
no
no
94.4
91.7
58.3
77.8
41.7
91.7
86.1
58.3
75.0
36.1
TE D
Genuine vs. 50 % adulteration
RI PT
Laboratory setup through quartz
SC
Detection of meat adulteration
ACCEPTED MANUSCRIPT Clustering was performed using PCA methods with a maximum of 9 principal components, however only the first two principal components were used for interpretation regarding the laboratory setup and the industrial setup. For the data of the handheld spectrometer it was necessary to use up to 4 principal components for interpretation.
2.0e-4
SC
RI PT
Figure 2 illustrates the differentiations between the genuine product and all adulteration levels with the industrial setup, thus 5 independent PCA scores plots are depicted for comparison. Clear differentiations could be achieved for all levels. Spectra were recorded through quartz cuvettes with the fibre optic probe. One score represents 6 measurements of one sub-sample. 6 sub-samples were analysed for each adulteration level.
adulteration 50% vs. genuine
0
M AN U
-2.0e-4 2.0e-4 1.0e-4
adulteration 40% vs. genuine
0
adulteration 30% vs. genuine
TE D
2.0e-4 0 -2.0e-4 4.0e-4 2.0e-4 0
AC C
-2.0e-4
adulteration 20% vs. genuine
EP
PC-2 (1%)
-1.0e-4 4.0e-4
1.0e-4
adulteration 10% vs. genuine
0
-1.0e-4
-3.0e-3
-2.0e-3
-1.0e-3
1.0e-3
0
2.0e-3
3.0e-3
4.0e-3
PC-1 (97%) adulteration (%)
50
40
30
20
10
genuine
Figure 2: Industrial setup: PCA scores plots for the detection of adulterated meat in veal sausages. All adulteration levels and the genuine product could be successfully discriminated.
14
ACCEPTED MANUSCRIPT
RI PT
As mentioned already, the final objective (to cluster all adulteration levels in one single PCA) embodies a rather scientific challenge, but it can be understood as a proof for quality and limitations of each method. The final objective could be achieved applying the laboratory setup (shown in figure 3) and the industrial setup for measurements through quartz cuvettes with a classification accuracy of 91.7 and 86.1 % (validation samples), respectively. However, measurements with the fibre optic probe directly through the polymer packaging obtained results showing some overlap of the 40 %, 30 %, and 20 % adulteration levels and led to broadened clusters in general. The spectra measured with the handheld spectrometer could not be represented in a satisfactory manner in a single PCA plot for all degrees of contamination.
SC
0.0003
0.0001
M AN U
PC-2 (2%)
0.0002
0 -0.0001 -0.0002
-0.003
-0.002
-0.001
0
0.001
0.002
TE D
PC-1 (92%) adulteration (%)
50
40
30
20
10
genuine
EP
Figure 3: Laboratory setup: PCA scores plot of all adulteration levels and the genuine product for pork detection in veal sausages. Adulteration of the product from genuine up to 50 % in 10 % steps.
AC C
3.2. Detection of fat adulteration
For this purpose spectral data, pre-treated with SNV transformation in combination with the 1st derivative (Savitzky-Golay method, 2nd polynomial order, 3 smoothing points), were used for clustering with PCA for all three instrumental setups. Best results were achieved using three different wavenumber regions for the three different setups: 6,028 - 5,480 cm-1 for the laboratory setup, 5,784 - 5,736 cm-1 for industrial setup and measurement through quartz cuvettes, 6,028 - 5,480 cm-1 for measurements through the polymer packaging and the region 6,037 - 5,576 cm-1 for all measurements with the handheld spectrometer. Results of the detection of the fat adulteration are summarized in table 4.
15
ACCEPTED MANUSCRIPT Table 4: SVM classification and PCA discrimination objectives for the detection of fat adulteration by the laboratory, industrial and on-site instrumental setups. Industrial setup
On-site setup
through quartz
through plastic package
through quartz
through plastic package
PCA clustering
yes
yes
yes
yes
yes
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering
100
100
100
100
yes
yes
100
100
100
100
Option
Genuine vs. 30 % adulteration
Genuine vs. 20 % adulteration
Genuine vs. 10 % adulteration
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%)
AC C
EP
Classification of all adulteration levels in one PCA/SVM
SVM correct discrimination of calibration samples (%) SVM correct discrimination of validation samples (%) PCA clustering
16
yes 100 100
100
100
100
100
100
91.7
yes
yes
yes
100
100
100
100
100
91.7
yes
yes
yes
no
100
100
100
83.3
100
100
83.3
M AN U
Genuine vs. 40 % adulteration
TE D
Genuine vs. 50 % adulteration
RI PT
Laboratory setup through quartz
SC
Detection of fat adulteration
100
yes
yes
yes
yes
no
100
100
100
100
83.3
100
100
100
100
75.0
yes
yes
yes
no
no
100
100
100
83.3
75.0
100
100
100
83.3
75.0
no
no
no
no
no
88.9
69.4
50.0
66.7
-
88.9
61.1
50.0
63.9
-
ACCEPTED MANUSCRIPT All PCA models were performed with a maximum of 9 principal components, however only the first 3 principal components were used for interpretation.
RI PT
Differentiation between the genuine product and samples with adulteration levels from 20 % up to 50 % was possible applying PCA models with one and two principal components regarding the laboratory and the industrial setups. Three principal components were necessary for several models applying the on-site setup. Samples with 10 % adulteration of the fat part could be separated from the genuine products applying the laboratory and the industrial setups. In that case measurements through polymer packaging also allow successful clustering. Thus classification could be performed with 100 % correct discrimination of all validation samples.
0
-0.01
AC C
-0.02
EP
PC-2 (5%)
0.01
TE D
M AN U
SC
Using the handheld spectrometer, it was not possible to reach sufficient clustering of the samples with 10 % adulteration through quartz cuvettes. A classification accuracy of 83.3 % was not satisfactory. Thus for analyses of fat adulteration, carried out with the on-site setup, the limit of the method was reached between the adulteration levels of 10 % and 20 %. Figure 4 illustrates the discrimination of the genuine product and samples with an adulteration level of 20 % for the measurements through quartz cuvettes. Performing measurements through the polymer packaging results in a shift of the limit of detection to even higher contamination. Thus samples with 30 % (or lower) adulteration could not be classified with more than 83.3 % correct discrimination by SVM.
-0.03 -0.1
0
0.1 PC-1 (94%)
adulteration (%)
20
genuine
Figure 4: On-site setup: PCA plot of samples with 20 % adulteration of the fat part and the genuine product. Measurements through quartz cuvettes.
Detection of contamination with pork fat in the fat part of the veal sausages requires a more sensitive method than detection of meat contaminations, due to the fact that the fat part (14.4 % of 17
ACCEPTED MANUSCRIPT the sausage) is smaller than the meat part (76.0 % of the sausage). Thus 10 % adulteration means that only 1.4 % of the total sausage was modified.
AC C
EP
TE D
M AN U
SC
RI PT
However, it was not possible to perform sufficient PCA clustering of all fat adulteration levels and the genuine product in one single analysis for any instrumental setup and option. Therefore classification accuracy of 88.9 % correct discrimination of validation samples (laboratory setup) could not be exceeded.
18
ACCEPTED MANUSCRIPT 4. Discussion
RI PT
In this work we developed methods for NIR spectroscopy, based on multivariate analyses to reveal deliberate adulteration or accidental contamination with pork in a pure veal product with regard to practical applications in the field of food and feed control. More precise, methods for laboratory use, industrial purposes and on-site analyses were established. On the one hand for the detection of pork adulteration in the meat part of sausages, on the other hand for pork fat contamination of the fat part of the sausages. For this purpose sausages were chosen as samples because of their similarity in composition and processing of the meat and fat parts to many other products on the market like burger, ravioli stuffing, ground meat and more. 4.1.1. Laboratory setup
M AN U
SC
The laboratory setup embodies the most accurate measurement device with the highest reproducibility of the measurements performed in this study. However, it is the most inflexible instrument due to the restriction to use quartz cuvettes filled with the samples. Therefore sample preparation was necessary. All samples have to be measured one by one with cleaning steps beforehand. In comparison to the other experimental modes, the largest measurement times of approx. 60 seconds were required.
TE D
Features of the laboratory setup are: best technical characteristics like a temperature controlled detector, widest spectral range, excellent wavenumber accuracy and relative reproducibility (for details please refer to section 2.2.). Due to the use of the spinner add-on (rotation of the quartz cuvettes during the measurement) an excellent averaging process was performed on the early state of scanning the samples. Rotating the sample while measurement is generally known to be particularly advantageous using non-destructive techniques like NIR spectroscopy (Schmutzler & Huck, 2014; Xiaobo, Jiewen, Xingyi, & Yanxiao, 2007).
AC C
EP
Applying the laboratory setup for the measurements provided the highest resolution spectra with minimal noise or perturbation compared to all other instrumental setups. Thus, flexibility and possibilities in the process of developing methods for multivariate data analysis are increased. Additionally, all adulteration levels could be detected for meat and fat without reaching the limit of detection. Further studies have to be performed applying the laboratory setup to investigate the limit of detection for meat adulteration in processed food and feed products with samples of lesser distinctions. 4.1.2. Industrial setup
We used the same basis instrument and operational settings as for the laboratory setup, to highlight influences of the fibre optic probe. Hyphenating NIR spectrometers with fibre optic cables opens up a huge field of possible applications for in-, on- and at-line measurements. Process control in real-time with no need for sample preparation comes along with such setups (Blanco & Villarroya, 2002; Fernández-Novales, López, Sánchez, García, & Morales, 2008; Wolfbeis, 2000). Fully automatic measurements are possible resulting in the highest number of measurements per day compared to the other instrumental setups. Moreover, measurements are possible in a hostile environment with extreme temperature, pressure, corrosiveness, humidity or vibrations which are adverse to the conditions required for common optical spectrometers. Problems caused by such 19
ACCEPTED MANUSCRIPT conditions can be easily solved by separating the locations of measurement and the spectrometer over large distances by the use of a light fibre for interconnection (Rohe, Becker, Kölle, Eisenreich, & Eyerer, 1999; Vojinović, Cabral, & Fonseca, 2006). It is obvious, that such devices are the most expensive ones compared to all other instruments presented in this work due to the additional fibre optic probe. Installation and (re)calibration of fibre optic NIR devices should be managed by experts and can cause additional expense.
M AN U
SC
RI PT
An increase of intrinsic scattering effects of the fiber optic cable accompanied with losses of intensity due to a small cross-section of the light collector for intake of diffuse reflected radiation are disadvantages of setups like these. All measurements done in this work applying the industrial device were carried out by hand without rotating the sample. Thus, no averaging process while measuring the samples was possible. As a consequence, reproducibility is poorer and spectra are weaker in intensity compared to the laboratory setup. However even measurements through the polymer packaging of the samples were of high quality. Resulting in successful clustering down to the lowest adulteration levels by means of both, meat and fat adulteration. In summary, the industrial device provides a high-quality standard combined with peerless flexibility in terms of use. 4.1.3. On-site setup
TE D
Mobility is of course an important advantage of the on-site setup. In addition to that, lowest cost of all devices, easy use and operation are features that can be essential for non-specialized users. Organisations in the field of consumer protection or end-product control can extend their capabilities utilising a handheld spectrometer. A typical measurement can be done with a minimum of time exposure of less than 3 seconds, which is also of advantage.
AC C
EP
With regard to technical characteristics, the handheld spectrometer cannot cope with the laboratory and industrial setup, resulting in spectra of lower spectral resolution, signal intensity, signal-to-noise ratio and wavelength range (see section 2.2.). These disadvantages limit the possibilities of multivariate analyses and result in lower reproducibility and higher limits of detection. Also perturbations of any kind taking effect on the spectral data can superimpose information of interest if there is a lack of intensity. In conformity with the latter topic, measurements through the polymer packaging for example came along with significant influence on the spectra. Thus the limit of detection of the on-site setup was higher than for all other devices. Analysing meat and fat adulteration, the limit of detection of the handheld device was found between 20 % to 10 % contamination. However for a realistic scenario the accuracy of the handheld spectrometer should be absolutely satisfactory. As we expect, real adulterations of food or feed products will be of a higher percentage than 20 % of the meat and fat part. On closer examination of the percentage values of correct classification applying SVM, the disturbing influence of the polymer packaging can be pointed out. Furthermore limitations of the handheld device are clearly indicated due to decreasing classification accuracy over the course of all adulteration levels starting with a value of 91.7 % for 50 % adulteration and ending up at 75 % for the lowest degree of contamination. 4.2. Differences between the detection of meat and fat adulteration
20
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
As the results show, the analysis of meat adulteration was successful for lower levels of contamination compared to the attempts to reveal fat adulteration. This could be a consequence of two reasons: First, the meat part of the sausages was 76.0 % of the total weight whereas the fat part was only 14.4 % - that means 10 % adulteration results in either 7.6 % or 1.4 % for meat and fat, respectively. Thus, the method for revealing pork fat as part of the sausage has to be of higher sensitivity. Second, substitution of veal with pork results always in a minor change of the total fat content of the whole sausage, due to the fact that pork (muscles only) contains 1.86 g fat (average) with a variation of 1.00 to 2.80 g per 100 g of weight, while veal (muscles only) contains 0.81 g fat (average) (Scherz & Senser, 1994). As a consequence, changes in the meat composition are accompanied by changes in the total fat content. Therefore, in case of meat adulteration PCA clustering is based on two aspects: the protein profile and the total fat content. In case of fat adulteration only one aspect is changed: the fatty acids profile. However, total fat content and protein profile remains constant.
21
ACCEPTED MANUSCRIPT 5. Conclusions This work illustrates the suitability of NIR technology for adulteration detection in sausages. However, Results show different advantages and disadvantages for the three instrumental setups applied in this work. The question, which one is the best can´t be answered in general. It is mainly depending on the field of application, required precision, requested number of measurements per day, financial effort and other reasons like expertise of the user.
M AN U
SC
RI PT
Meat and fat adulteration could be revealed down to the lowest level of contamination (10 %) applying the laboratory setup and the industrial fibre optics setup. Furthermore, it was possible to measure directly through the polymer packaging of the samples. Analyses with the on-site setup (using a handheld NIR spectrometer) also led to successful separation between the contaminated samples and the genuine products down to the lowest degree of meat adulteration (10 %). However, measurements through the polymer packaging were possible down to a contamination level of 20 %. Measuring fat adulteration, the limits of the handheld spectrometer came to light more clearly. Contamination levels less than 20 % could not be separated from the unadulterated samples in a satisfactory manner. Nevertheless 20 % adulteration means only 2.8 % of the sausage in total was modified (adulteration indication concerns the fat part only).
AC C
EP
TE D
Processing the spectroscopic data, wavelength selection turned out as a critical task to eliminate perturbations of the spectral information. SVM classification has proven to be an optimal technique to evaluate PCA results in a numeric way and to be utilized for automated discrimination between genuine and adulterated samples.
22
ACCEPTED MANUSCRIPT Acknowledgements
AC C
EP
TE D
M AN U
SC
RI PT
The authors want to thank the Interreg IV initiative of the European Union (project “OriginAlp”) for financial support and the whole team of the project “OriginAlp” for good advice and support for all reasons. Special thanks goes to AMTirol, Ing. Alexander Walser and DI. Wendelin Juen. Last but not least we want to thank the Spar Holding AG (Salzburg, Austria) for supplying us with all the ingredients used for producing the “Edelbrater” veal sausage, the original recipe and for their confidence.
23
ACCEPTED MANUSCRIPT References
AC C
EP
TE D
M AN U
SC
RI PT
Balabin, R. M., & Lomakina, E. I. (2011). Support vector machine regression (SVR/LS-SVM)—an alternative to neural networks (ANN) for analytical chemistry. Comparison of nonlinear methods on near infrared (NIR) spectroscopy data. Analyst, 136(8), 1703-1712. Blanco, M., & Villarroya, I. (2002). NIR spectroscopy: a rapid-response analytical tool. TrAC Trends in Analytical Chemistry, 21(4), 240-250. Boser, B. E., Guyon, I. M., & Vapnik, V. N. (1992). A training algorithm for optimal margin classifiers. In Proceedings of the fifth annual workshop on Computational learning theory (pp. 144-152): ACM. Chauchard, F., Cogdill, R., Roussel, S., Roger, J., & Bellon-Maurel, V. (2004). Application of LS-SVM to non-linear phenomena in NIR spectroscopy: development of a robust and portable sensor for acidity prediction in grapes. Chemometrics and Intelligent Laboratory Systems, 71(2), 141150. Chih-Chung, C., & Chih-Jen, L. (2001). LIBSVM: A Library for Support Vector Machines. Retrieved from http://www.csie.ntu.edu.tw/~cjlin (Accessed 04.01.15). Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines and other kernel-based learning methods: Cambridge university press. dos Santos, C. A. T., Lopo, M., Páscoa, R. N., & Lopes, J. A. (2013). A Review on the applications of portable near-infrared spectrometers in the agro-food industry. Applied Spectroscopy, 67(11), 1215-1233. European Commission. (2013). Horse meat - questions and answers. Retrieved from http://ec.europa.eu/food/safety/official_controls/food_fraud/horse_meat/q-ans_en.htm (Accessed 28.03.15). European Commission. (2015). Food and Feed Safety. Retrieved from http://ec.europa.eu/food/food/index_en.htm (Accessed 28.03.15). European Food Safety Authority. (2015). About EFSA. Retrieved from http://www.efsa.europa.eu/en/aboutefsa.htm (Accessed 28.03.15). European Union law. (2002). Access to European Union law EUR-Lex - 32002R0178 - EN. Retrieved from http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32002R0178 (Accessed 20.03.15). Fernández-Novales, J., López, M.-I., Sánchez, M.-T., García, J.-A., & Morales, J. (2008). A feasibility study on the use of a miniature fiber optic NIR spectrometer for the prediction of volumic mass and reducing sugars in white wine fermentations. Journal of Food Engineering, 89(3), 325-329. Godfray, H. C. J., Beddington, J. R., Crute, I. R., Haddad, L., Lawrence, D., Muir, J. F., Pretty, J., Robinson, S., Thomas, S. M., & Toulmin, C. (2010). Food security: the challenge of feeding 9 billion people. Science, 327(5967), 812-818. Hattori, Y., Tajiri, Y., Yonai, T., & Otsuka, M. (2013). Real-time monitoring of pharmaceutical powder blending using a wireless near infrared spectrophotometer. NIR news, 24(4), 9-11. Huang, H., Yu, H., Xu, H., & Ying, Y. (2008). Near infrared spectroscopy for on/in-line monitoring of quality in foods and beverages: A review. Journal of Food Engineering, 87(3), 303-313. Huck, C. W. (2014). Advances of vibrational spectroscopic methods in phytomics and bioanalysis. Journal of Pharmaceutical and Bomedical Analysis, 87, 26-35. Kowalski, B. R. (1984). Chemometrics: mathematics and statistics in chemistry: Springer. Lanza, E. (1983). Determination of moisture, protein, fat, and calories in raw pork and beef by near infrared spectroscopy. Journal of Food Science, 48(2), 471-474. Martens, H. (1989). Multivariate calibration: John Wiley & Sons. McClure, W. F. (1994). Near-infrared spectroscopy. The giant is running strong. Analytical Chemistry, 66(1), 43A-53A. Meyer, R., & Candrian, U. (1996). PCR-based DNA analysis for the identification and characterization of food components. LWT-Food Science and Technology, 29(1), 1-9. 24
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
Muehlemann, T., Haensse, D., & Wolf, M. (2008). Wireless miniaturized in-vivo near infrared imaging. Optics Express, 16(14), 10323-10330. Murugaiah, C., Noor, Z. M., Mastakim, M., Bilung, L. M., Selamat, J., & Radu, S. (2009). Meat species identification and Halal authentication analysis using mitochondrial DNA. Meat Science, 83(1), 57-61. O’Brien, N., Hulse, C. A., Pfeifer, F., & Siesler, H. W. (2013). Near infrared spectroscopic authentication of seafood. Journal of Near Infrared Spectroscopy, 21, 299-305. Ozaki, Y., McClure, W. F., & Christy, A. A. (2006). Near-infrared Spectroscopy in Food Science and Technology: John Wiley & Sons. Pierna, J. A., Baeten, V., Renier, A. M., Cogdill, R. P., & Dardenne, P. (2004). Combination of support vector machines (SVM) and near-infrared (NIR) imaging spectroscopy for the detection of meat and bone meal (MBM) in compound feeds. Journal of Chemometrics, 18, 341-349. Premanandh, J. (2013). Horse meat scandal–A wake-up call for regulatory authorities. Food Control, 34(2), 568-569. Reich, G. (2005). Near-infrared spectroscopy and imaging: basic principles and pharmaceutical applications. Advanced drug delivery reviews, 57(8), 1109-1143. Rohe, T., Becker, W., Kölle, S., Eisenreich, N., & Eyerer, P. (1999). Near infrared (NIR) spectroscopy for in-line monitoring of polymer extrusion processes. Talanta, 50(2), 283-290. Scherz, H., & Senser, F. (1994). Food Composition and Nutrition Tables: Medpharm GmbH Scientific Publishers. Schmutzler, M., & Huck, C. W. (2014). Automatic sample rotation for simultaneous determination of geographical origin and quality characteristics of apples based on near infrared spectroscopy (NIRS). Vibrational Spectroscopy, 72, 97-104. Siesler, H. W., Ozaki, Y., Kawata, S., & Heise, H. M. (2008). Near-infrared spectroscopy: principles, instruments, applications: John Wiley & Sons. Sorak, D., Herberholz, L., Iwascek, S., Altinpinar, S., Pfeifer, F., & Siesler, H. W. (2012). New developments and applications of handheld Raman, mid-infrared, and near-infrared spectrometers. Applied Spectroscopy Reviews, 47(2), 83-115. Tøgersen, G., Isaksson, T., Nilsen, B., Bakker, E., & Hildrum, K. (1999). On-line NIR analysis of fat, water and protein in industrial scale ground meat batches. Meat Science, 51(1), 97-102. Tsenkova, R., Atanassova, S., Morita, H., Ikuta, K., Toyoda, K., Iordanova, I. K., & Hakogi, E. (2006). Near infrared spectra of cow's milk for milk quality evaluation: disease diagnosis and pathogen identification. Journal of Near Infrared Spectroscopy, 14(6), 363-370. Vapnik, V. (1998). Statistical learning theory (Vol. 2): Wiley New York. Vojinović, V., Cabral, J., & Fonseca, L. (2006). Real-time bioprocess monitoring: Part I: In situ sensors. Sensors and Actuators B: Chemical, 114(2), 1083-1091. Windham, W., Lawrence, K., & Feldner, P. (2003). Prediction of fat content in poultry meat by nearinfrared transmission analysis. The Journal of Applied Poultry Research, 12(1), 69-73. Wold, J., O'Farrell, M., Høy, M., & Tschudi, J. (2011). On-line determination and control of fat content in batches of beef trimmings by NIR imaging spectroscopy. Meat Science, 89(3), 317-324. Wolfbeis, O. S. (2000). Fiber-optic chemical sensors and biosensors. Analytical Chemistry, 72(12), 8190. Wu, W., Massart, D., & De Jong, S. (1997). Kernel-PCA algorithms for wide data Part II: Fast crossvalidation and application in classification of NIR data. Chemometrics and Intelligent Laboratory Systems, 37(2), 271-280. Xiaobo, Z., Jiewen, Z., Xingyi, H., & Yanxiao, L. (2007). Use of FT-NIR spectrometry in non-invasive measurements of soluble solid contents (SSC) of ‘Fuji’apple based on different PLS models. Chemometrics and Intelligent Laboratory Systems, 87(1), 43-51. Yao, P., Guo, W., Sheng, X., Zhang, D., & Zhu, X. (2014). A portable multi-channel wireless NIRS device for muscle activity real-time monitoring. In Engineering in Medicine and Biology Society (EMBC), 36th annual international conference of the IEEE (pp. 3719-3722): IEEE. 25
ACCEPTED MANUSCRIPT
EP
TE D
M AN U
SC
RI PT
Adulteration or contamination of veal sausages with unwelcome meat was detected. We compared FT-NIR devices and methods for food authentication control. Instruments for laboratory, industrial and on-site (handheld) analyses were used Principal component analysis and support vector machines classification were utilized.
AC C
-