Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Food Research International 60 (2014) 59–65

Contents lists available at ScienceDirect

Food Research International


journal homepage: www.elsevier.com/locate/foodres

Verification of fresh grass feeding, pasture grazing and organic farming


by FTIR spectroscopy analysis of bovine milk
Edoardo Capuano a,⁎, Jan Rademaker b, Harrie van den Bijgaart b, Saskia M. van Ruth a
a
RIKILT, Institute of Food Safety, Wageningen University and Research Centre, P.O. Box 230, 6700 AE Wageningen, The Netherlands
b
Qlip N.V., Postbus 119, 7200 AC Zutphen, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: In the present study, a total of 116 tank milk samples were collected from 30 farms located in The Netherlands
Received 12 August 2013 and analysed by Fourier-transform infrared (FTIR) spectroscopy. Samples were collected in April, May and
Received in revised form 11 December 2013 June 2011 and in February 2012. The samples differed in the time spent by the cows on pasture, presence/absence
Accepted 20 December 2013
of fresh grass in the daily ration and the farming system (organic/biodynamic or conventional). Classification
Available online 28 December 2013
models based on partial least square discriminant analysis (PLS-DA) of FTIR spectra were developed for the pre-
Keywords:
diction of fresh grass feeding, pasture grazing and organic farming. The PLS-DA model discriminated between
Milk milk from cows that had fresh grass in the daily ration and milk from cows that had not fresh grass with sensi-
Authentication tivity and specificity values of 88% and 83% in external validation and all the samples from cows that had no
FTIR spectroscopy fresh grass collected in spring were correctly classified. The PLS-DA model developed for the authentication of
Grass feeding pasture grazing showed comparable accuracy when the whole sample set is considered but was less accurate
Pasture on the spring samples (75% of samples from cows indoors in spring correctly classified). Discrimination of organic
Chemometrics and conventional milk was also accomplished with acceptable accuracy with % correct classification of 80% and
PLS-DA
94% respectively in external validation. The results suggest that milk FTIR spectra contain valuable information
on cows' diet that can be used for authentication purposes.
© 2014 Elsevier Ltd. All rights reserved.

1. Introduction verification of cheese geographical origin (Boscaini, van Ruth, Biasioli,


Gasperi, & Mark, 2003; Galle et al., 2011; Karoui, Bosset, Mazerolles,
Chemometrics can be defined as the science of extracting chemical Kulmyrzaev, & Dufour, 2005; Pillonel, Ampuero, Tabacchi, & Bosset,
information from data by using mathematical and statistical tools. 2003) or ripening time (Frau, Massanet, Rossello', Simal, & Canellas,
Chemometrics has proven to be particularly suited for the verification 1997). Fluid milk has been also extensively investigated. Chemometrics
of certain food quality attributes or the detection of adulterations has been applied for the discrimination of the animal species (Andreotti,
which could not be attained by the analysis of one single food property Trivellone, Lamanna, Di Luccia, & Motta, 2000; Nicolaou, Xu, &
but requires the generation of multivariate dataset, i.e. based on the Goodacre, 2011; Smiddy, Huppertz, & van Ruth, 2012), for the determi-
analysis of more than one sample property. The application of chemo- nation of milk composition and origin (Ntakatsane, Yang, Lin, Liu, &
metric modelling aims at detecting specific patterns or fingerprints in Zhou, 2011), for the discrimination between raw and processed milk
a multivariate dataset that can be used to infer about the genuine nature (Dufour & Riaublanc, 1997) and for the detection of milk adulteration
of a food sample. (Souza et al., 2011).
A large number of scientific studies have been published over the Further examples of application of chemometric modelling to dairy
last years wherein chemometrics has been applied to data generated products other than cheese and fluid milk may be given: discrimination
from a variety of analytical methodologies (chromatography, spectros- of low-fat and full-fat yoghurt (Cruz et al., 2013), verification of purity of
copy, mass spectrometry, calorimetry, wet chemistry) for the authenti- butter (Koca, KocaoglufVurma, Harper, & Rodriguez-Saona, 2010; Lipp,
cation of dairy products. Many studies have focused on cheese. 1996) and detection of adulterations in milk powders (Borin, Ferrao,
Chemometric modelling has been applied, for instance, for the Mello, Althmann Maretto, & Poppi, 2006; Maraboli, Cattaneo, Maria, &
Giangiacomo, 2002).
Consumers are nowadays increasingly interested in information
about the origin of their foods including information on dairy cows'
⁎ Corresponding author at: RIKILT, Institute of Food Safety, Wageningen UR, P.O. Box
diet, housing and herd management system. That is because specific
230, 6700 AE Wageningen, The Netherlands. Tel.: +31 317 480260; fax: +31 317 417717. feeding regimen and herd management practices such as fresh grass
E-mail address: edoardo.capuano@wur.nl (E. Capuano). feeding or pasture grazing are linked to superior organoleptic and

0963-9969/$ – see front matter © 2014 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.foodres.2013.12.024
60 E. Capuano et al. / Food Research International 60 (2014) 59–65

nutritional quality attributes of milk and dairy products (Elgersma, 2. Material and methods
Tamminga, & Ellen, 2006). Furthermore those specific management sys-
tems are perceived by large groups of consumers as more natural, 2.1. Study design
healthy and respectful of animal welfare. Therefore such products
would gain an added value compared to conventional counterparts A total of 116 tank milk samples were collected from 30 different
and might therefore be fraudulently mislabelled. In The Netherlands, farms located in the Gelderland region (6 farms) and in the Friesland/
for instance, a quality label has been recently introduced for milk Groningen region (24 farms) in The Netherlands. The sampling was
and dairy products that are produced from cows on pasture at conducted in the framework of a larger project in which other method-
least 4 months per year, 6 h per day: “Weidemelk”, which is the ologies were tested for the authentication of fresh grass feeding, pasture
Dutch for “pasture milk” (Elgersma, 2012). The European regula- grazing and organic farming. With few exceptions, each farm provided 4
tion for organic farming also requires that the animals are given milk samples, one in each of the following month: in April, May and
an adequate access to an outdoors area and that at least 60% of June 2011 and in February 2012. Questionnaires and interviews were
the dry matter in the daily ration of herbivores should consist of used to collect information on farm management and cow diet from
roughage (Commission regulation (EC) No 889/2008). In addition the week of sampling. The grass intake with cut fresh grass fed indoor
to the paper-trail and inspection based systems of certification of was estimated from the acreage of cut land and weight of fed fresh
product traceability, reliable analytical methodologies may help grass. The grass intake via grazing was estimated by the farmer through
the various stakeholders in verifying that the claimed quality attributes their experience from the animal energy needs (based on milk produc-
are fully met. tion) minus what was fed next to the grassland. The milk samples were
In an earlier research we already found that milk fatty acid (FA) pro- collected from the tank by using 30-mL plastic screw-top containers
file can be proficiently used for the authentication of fresh grass feeding after adequate stirring of the tank. The milk samples were cooled
at a level of the amount of fresh grass as low as 30% of the daily ration. (4 °C) immediately after sampling and frozen (− 18 °C) within 6 h
PLS-DA binary classification models based on algorithm appear to be after sampling. Samples were categorized in five groups: Milk samples
highly successful, with sensitivity and specificity values close to 100% from cows at least 19 h outdoors on pasture on a daily basis at the
(Capuano et al., submitted for publication). However, the proposed an- time of sampling (group P2); Milk samples from cows 6–9 h outdoors
alytical methodology is based on gas chromatography with flame on pasture on a daily basis at the time of sampling (group P1); Milk
ionisation detection analysis which involves labour intensive and time samples from cows indoors with fresh grass in the diet (group GI);
consuming fat extraction and analysis and is therefore relatively expen- Milk samples collected in spring from cows indoors with no fresh
sive. For the routine analysis of dairy samples for authentication pur- grass in the diet (group NG); Milk samples collected in winter (group
poses, cheap, fast, non-destructive and user friendly technologies W). In groups P2, P1 and GI, the estimated amount of fresh grass in
would be more valuable. Infrared (IR) spectroscopy is a suitable candi- the cows' daily ration varied from 36% to 94%. All the milk samples of
date for that. The IR regions of the electromagnetic spectrum that are group W were from cows indoors all the time with no fresh grass in
relevant to food analysis are the mid IR (MIR, 4000–400 cm−1) region the daily ration. Two certified organic farms, three certified biodynamic
and the near IR region (NIR, 14,000–4000 cm−1). Recently, it has been farms and one farm converting to organic farming were included in the
shown that NIR spectroscopy has a potential for the discrimination be- sample set. In Table 1 an overview of the study design is provided.
tween raw milk from cows on pasture and milk from stabled cows
(Coppa et al., 2012). Discrimination (95% of correct classification in 2.2. FTIR analysis
cross-validation) between milk from cows on pasture from stabled
cows was reported based on the spectral interval from 25,000 cm− 1 The content of fat, protein, lactose, urea, total unsaturated FAs as
to 4000 cm−1. MIR spectroscopy is applied for the quantitative analysis well as the milk freezing point depression (FPD) were measured using
of major milk components since decades and is extensively employed MilkoScan FT 6000 equipment (Foss, Hillerod, Denmark) with the man-
for routine compositional quality control analysis in the dairy industry ufacturer supplied basic calibration models according to manufacturer's
and in milk production registration of individual cows. Applications of recommended procedures (Foss, Hillerod, Denmark) at Qlip N.V.
MIR spectroscopy in milk include the prediction of milk coagulation (Zutphen, The Netherlands) on 10-mL milk subsamples in conformity
properties (Cecchinato, De Marchi, Gallo, Bittante, & Cainier, 2009; De with ISO FDIS 9622|IDF 141: 2013. The applying reference methods
Marchi, dal Zotto, Cassandro, & Bittante, 2007), protein composition were ISO 1211|IDF 1 (fat), ISO 8968-1|IDF 20-1 (total protein), ISO |IDF
(Rutten, Bovenhuis, Heck, & van Arendonk, 2010; Sorensen, Lund, & 198 (lactose), ISO 14637|IDF 195 (urea), ISO 15884|IDF 182 and ISO
Jull, 2003) and fat composition (Rutten, Bovenhuis, Hettinga, van 15885|IDF 184 (total unsaturated FAs) and ISO 5764|IDF 108 (FPD).
Valenberg, & van Arendonk, 2009). A notable advancement in IR spec- FTIR absorption spectra were separately recorded for all 116 milk sam-
troscopy technology is represented by Fourier transform infrared ples. They consisted of 1060 infrared frequencies (wavenumbers) ranging
(FTIR) spectroscopy. FTIR differs from standard IR technology because from 925 to 5008 cm−1.
it employs the interferometric modulation of radiation to measure mul-
tiple frequencies simultaneously. The resulting interferogram is then 2.3. Statistical analysis
converted to the original spectrum using complex algorithms. The
major advantage of FTIR over dispersive-based instrumentations are: First, the normality of the distribution of fat, protein, lactose, urea,
improved sensitivity due to higher signal to noise ratio, improved total unsaturated FAs and FPD within the groups P2, P1, GI, NG and
speed of analysis, higher energy throughput and superior accuracy W was checked by means of a Shapiro–Wilk test for normality
(Rodriguez-Saona & Allendorf, 2011). (P b 0.05). Since the distributions appeared to be non-normal in the
Despite these advantages, to our knowledge only one study has in- groups, one Kruskal–Wallis 1-way ANOVA test for group comparisons
vestigated the potential of FTIR spectroscopy for milk authentication was performed among the groups. If the Kruskal–Wallis test result
purposes. Maamouri et al. (2008) have used FTIR spectroscopy to mon- was significant (P b 0.05), pairwise comparisons between groups
itor the quality of milk from Sicilo-Sarde ewes and to discriminate be- were performed by means of a Mann–Whitney U test (P b 0.05 was
tween milk samples from ewes that have been fed soybean meal or considered significant). SPSS v19.0 (IBM Statistics Inc., Armonk, NY)
scotch bean. The aim of the present study was therefore to investigate was used to perform all those tests. Principal Component Analysis
whether FTIR spectroscopy combined with chemometrics is an accurate (PCA) of the FTIR spectra was performed to screen the multivariate
and valid technique for the authentication of cows' milk with respect to dataset for outliers and to explore the presence of any grouping in the
grass feeding, pasture grazing and organic farming. data (Berrueta, Alonso-Salces, & Héberger, 2007). Partial least squares-
E. Capuano et al. / Food Research International 60 (2014) 59–65 61

Table 1
Summary of the study design. Characteristics and composition of the different groups.

Group Hours on pasture per day Min–max (average) % fresh grass per day Samples collected in each montha Totala

April May June February

GI 0 38–71 (54) 4 4 4 0 12
P1 6–9 36–74 (51) 12 (5) 11 (1) 8 0 31 (6)
P2 N19 52–94 (76) 4 (1) 9 (5) 12 (6) 0 25 (12)
NG 0 0 7 6 6 0 19
W 0 0 0 0 0 29 (5) 29 (5)
Totala 27 (6) 30 (6) 30 (6) 29 (5) 116 (23)
a
In parenthesis, the number of samples from organic/biodynamic farms.

discriminant analysis (PLS-DA) was performed to develop binary classi- target class before the regression is performed. PCA and PLS-DA were per-
fication models to verify the cows' feeding regime (fresh grass feeding formed in Pirouette 4.5 (Infometrix, Seattle, USA).
vs no fresh grass feeding), the housing management system (indoors
vs pasture) and the farming management system (organic vs conven- 3. Results and discussion
tional). In a PLS-DA model, the dimensionality of the dataset is reduced
by calculating new variables (latent components or factors) from the 3.1. Physical-chemical analysis
combination of the original variables in order to find the maximum cor-
relation between the new variables and the class variable and therefore The content in fat, protein, lactose, urea, unsaturated FAs, and FPD
to obtain the maximum separation among the target classes. To develop values of milk samples were directly measured by the MilkoScan
and validate the PLS-DA classification models, the sample set was divid- based on absorbance in specific MIR regions. One of the samples showed
ed into a training set, consisting of a random selection of 75% of the sam- abnormal values for fat (9.6%), protein (2.95%), unsaturated FAs (2.92 g/
ples from each of the class (class GRASS or NO GRASS, INDOORS or 100 g) and FPD (−489 m °C(H)). This sample was therefore excluded
PASTURE, ORGANIC or CONVENTIONAL, see also Section 3.3. for the from further statistical analysis. The results for groups P2, P1, GI, NG
composition of the classes from the original 5 groups) and a validation and W are reported in Table 2. From Table 2 it appears that the main dif-
set consisting of the remaining 25% of samples from each class. The ference among the groups concerns the milk fat content which is signif-
training set of samples was used to develop the models. Various data icantly higher in the winter samples compared to milk from cows on
pre-processing techniques, including mean-centering, autoscaling pasture and milk from cows stabled without fresh grass in the diet but
(scaling to unit variance), Pareto scaling (scaling to square root of the no differences were found between milk from stabled cows and cows
variance), first and second derivatives, smoothing and orthogonal signal on pasture in spring. A significantly higher fat content in stabled cows
correction (OSC) and a combination of them were tested. The models compared to grazing cows is reported, for instance, by White et al.
were internally validated by leave 10% out cross-validation and exter- (2001) and Morales-Almaraz et al. (2010) but the opposite is reported
nally validated on the samples of the validation set. For each model, by Slots et al. (2008). The effect of grass feeding and pasture grazing
the optimal number of factors was selected for each class which gave on FA profile has been repeatedly reported. The inclusion of fresh
the least standard error of cross validation (SECV). For each binary clas- grass in the ration is known to increase the level of unsaturation of
sification, the best performing model was selected which gave the milk fat (Couvreur, Hurtaud, Lopez, Delaby, & Peyraud, 2006;
highest % of correct predictions in external validation. The PLS1 algo- Dewhurst, Shingfield, Lee, & Scollan, 2006; Ellis et al., 2006). In our
rithm described in Manne (1989) was used for the development of study the unsaturated FA distribution was significantly different only
PLS-DA classification models. A tie-breaker criterion was used to classify between group NG and group P1. Lactose distribution was significantly
the samples in either of the two classes: the sample is considered to different only between groups P2 and NG whereas the protein distribu-
better fit the class with the Y prediction N 0.5. The first derivative, the tion was significantly different only between groups NG and W. The dis-
second derivatives and the smoothing transforms were based on a tribution of the FPD values was not statistically different among the 5
Savitzky–Golay polynomial filter (Savitzky & Golay, 1964) modified as groups even though a higher average value was found in milk from
suggested by Gorry (1990). A convolution is applied to independent cows on continuous grazing. The FPD of genuine milk depends on
variables in a window containing a central data point and n points on ei- many factors including cow breed, lactation stage, diet and year
ther side. A weighted second-order polynomial is fit to the data window (Elisses & Zee, 1980; Elschner, Jacobi, Buchberger, & Grun, 1997). A
and the centre point is replaced by the fitted value. Different weighting higher FPD during the grazing period has already been reported
coefficients are applied for the three transforms. OSC pre-processing (Henno, Ots, Joudu, Kaart, & Kart, 2008). All in all, with the exception
removes the variables that contribute little to the correlation with the of the fat content in winter milk and the unsaturated FAs content in

Table 2
Physical–chemical properties of milk samples analysed by FTIR in groups GI (milk from cows fed grass indoors), P1 (milk from cows on pasture 6–9 h per day), P2 (milk from cows on
pasture N19 h per day), NG (milk from cows with no fresh grass collected in spring) and W (samples collected in winter). Values are means ± SD.

Parameter GI P1 P2 NG W

Fat (%) 4.17 ± 0.16ab 4.15 ± 0.19 a 3.94 ± 0.31 a 4.10 ± 0.24a 4.66 ± 1.00b
Proteins (%) 3.45 ± 0.19ab 3.46 ± 0.11ab 3.52 ± 0.19ab 3.42 ± 0.09a 3.54 ± 0.17b
Lactose (%) 4.52 ± 0.08ab 4.51 ± 0.08ab 4.50 ± 0.08a 4.56 ± 0.07b 4.50 ± 0.12ab
Urea (mg/100 g) 20.0 ± 7.2 21.6 ± 6.6 20.3 ± 9.4 21.1 ± 4.4 20.6 ± 4.0
Unsaturated fatty acids (g/100 g) 1.29 ± 0.10ab 1.35 ± 0.11 a 1.27 ± 0.11ab 1.18 ± 0.08b 1.32 ± 0.34ab
Freezing point depression (−m °C(H)) 518 ± 5 518 ± 6 516 ± 9 518 ± 6 519 ± 7
a,b,x,y
Different superscripts in a row indicate significant differences between distributions (Kruskal–Wallis test for group comparisons, P ≤ 0.05 and Mann–Whitney U test for pairwise
comparisons, P b 0.05).
62 E. Capuano et al. / Food Research International 60 (2014) 59–65

spring milk from stabled cows few remarkable differences among the 5 very atypical profile at different wavelengths. This sample was the
groups appear from Table 2. same which showed an unusual proximate composition (see
Section 3.1.) This sample was thus excluded from further chemometric
modelling. PCA was then conducted with the MIR spectral data (900–
3.2. Exploring the FTIR data: raw spectra and PCA 3000 cm−1 with the exclusion of the region 1800–2800 cm−1 of the re-
maining 115 samples) to reveal natural clustering of the samples and to
As a first step, a PCA was carried out on the whole dataset using the 6 detect outliers. Several data pre-processing methods were tested. In
physical-chemical properties described in Section 3.1. as variables. The Fig. 2 the PCA scores plot of the pre-processed (autoscaling, smoothing
PCA scores plot did not highlight any natural clustering of the milk sam- and second derivative) FTIR spectra is depicted, whereas Fig. 3 shows
ples between the 5 groups (data not shown). Additional to the physical– the eigenvectors corresponding to the first two PCs. The PCA scores
chemical parameters, the information contained in the whole FTIR spec- plot revealed that the milk samples collected in winter are clearly dis-
tra themselves was explored. In Fig. 1, the FTIR spectra of a milk sample tinct from the milk samples collected in spring even though some over-
of group P2 is shown as an example. The absorption bands observed in lapping occurs. Milk from cows with no fresh grass in the diet collected
the MIR region are associated with the fundamental vibrations of func- in spring are clearly distinct from the winter samples and almost
tional groups of the molecules. For quantification purposes, the absorp- completely separated from milk from cows under continuous grazing
tion bands of the MIR spectrum occurring at approximately 1175 cm−1 whereas a more extensive overlapping is apparent with milk from
(corresponding to transitions in the vibrational energy of the triglycer- cows on pasture only 6–9 h. A slight difference appeared between the
ide ester linkage C\O), 1750 cm− 1 (carbonyl group of FAs), and in two groups of cows on pasture whereas the difference between the
the region 2800–3000 cm− 1 (FAs acyl chain C\H) are commonly milk from cows fed fresh grass indoors and the pasture groups seems
used to quantify fat (Dupuy, Duponchel, Huvenne, Sombret, & very limited. The PCA scores plot also revealed that one sample belong-
Legrand, 1996; Safar, Bertrand, Roberts, Devaux, & Genot, 1994; Yang ing to group P2 always laid very far from the sample cloud and exhibited
& Irudayaraj, 2000) while the bands at 1653 cm−1 (amide I) and a Mahalanobis distance far beyond the 95% probability threshold value
1567 cm−1 (amide II) are used for the quantification of proteins and calculated from the PCA models irrespective of the pre-processing tech-
to obtain information on protein secondary structure (Dufour, Robert, nique applied. This sample also exhibited a residual variance significant-
Renard, & Llamas, 1998; McQueen, Wilson, Kinnunen, & Jensen, 1995; ly larger than PCA model variance (F test, P b 0.05). This sample was
Van de Voort, Sedaman, Emo, & Ismail, 1992). Vibrations arising from therefore considered an outlier and excluded from further modelling.
C\O and C\H stretching produce bands in the region between
1100 and 1000 cm−1 which give information on sugar molecules
3.3. Binary classification models: PLS-DA
(Hashimoto & Kameoka, 2008). Water significantly absorbs in the MIR
region and can interfere with the quantification of other major milk
The aim of the present study was to investigate the possibility of
components. Major bands are present at 3920, 3490, 3280 cm− 1
using FTIR spectra for the authentication of fresh grass feeding, pasture
(O\H stretching vibrations), and 1645 cm−1 (resulting from water
grazing and organic farming. We used PLS-DA for building three binary
bending vibrations), the exact location depending on the specific inter-
classification models. Binary classification models aim at predicting the
actions of water molecules with solutes (Lin & Brown, 1992) and tem-
actual class membership of a new unknown sample. In our case model 1
perature (Libnau, Kvalheim, Christy, & Toft, 1994). In our milk
aimed at predicting whether the cows have been fed fresh grass or not,
samples, strong absorption from water molecules was apparent be-
model 2 aimed at predicting whether the cows have been grazing on
tween 1620 and 1670 cm−1 and between 3000 and 3600 cm−1. The re-
pasture or not, whereas model 3 aimed at predicting whether milk
gion between 3000 cm−1 and 3600 cm−1 was eliminated for further
was from organic/biodynamic or conventional farms.
analysis whereas the region between 1620 and 1670 cm− 1 was
For model 1 the groups P2, P1 and GI were pooled in the class GRASS,
retained because information on protein composition is embedded in
whereas the groups NG and W were pooled in the class NO GRASS. The
it. The regions ranging from 1800 cm−1 to 2800 cm−1 as well as the re-
whole sample set was constituted by 114 raw milk samples, after the
gion between 4000 cm− 1 and 5000 cm−1 was eliminated prior to
elimination of two outliers as described in Section 3.2. The most satisfac-
further modelling because it does not contain valuable information on
torily results in terms of number of misclassified samples in external
milk composition.
validation was obtained by a PLS-DA model after autoscaling, smooth-
Few evident differences among the 5 groups of Table 1 could be
ing and 2nd derivative transformation of the raw data and the applica-
highlighted by comparing the FTIR raw spectra. Chemometric treatment
tion of one OSC component (Table 3). The smoothing and the 2nd
of the MIR spectra combined is usually required to exploit the multivar-
iate information contained in these spectra and to reveal subtle differ-
ences that might exist between different types of samples (Woodcock,
Downey, & O'Donnell, 2008). In the raw spectra, it was apparent that
one sample was very different from the bulk of the data showing a

Fig. 2. PCA scores plot of the FTIR data (autoscaling, smoothing and second derivative).
Variance explained by each factor is provided in parentheses. In the circle the outlier sam-
Fig. 1. FTIR raw spectra of a milk sample from cows on pasture N19 h per day. ple that was excluded from further modelling.
E. Capuano et al. / Food Research International 60 (2014) 59–65 63

very satisfactory in internal validation with an average % of correct pre-


diction of 93% and 95% for class PASTURE and class INDOORS respective-
ly, whereas in external validation 91% of milk samples from class
PASTURE and 81% of milk samples from class INDOORS were correctly
classified in their actual class (Table 3). Out of the three samples
misclassified in class INDOORS, two were collected in spring and one
in winter which reduces to 75% the specificity of the model on spring
samples. The lower specificity of model 2 compared to model 1 was ex-
pected and related to the higher misclassification rate of milk from cows
fed grass indoors compared to milk from stabled cows without fresh
grass. Indeed the two samples of class INDOORS misclassified in internal
validation and one of the samples misclassified in external validation
belonged to group GI. The PLS-DA regression vector showed high scores
(in absolute value) for bands around 1640 cm−1 (positive score) and
around 1585 cm− 1 and 1695 cm− 1 (negative scores) (data not
shown) and smaller peaks around 1020 cm−1, 1380 cm−1 and in the
region 2800–3000 cm−1. The unambiguous assignment of the first
Fig. 3. Eigenvectors of the first two PCs for the PCA model based on the FTIR data
(autoscaling, smoothing and second derivative). The % of explained variance is provided three bands to specific vibrations is difficult but they are related to pro-
in parenthesis. tein content, protein secondary structure and interactions with water
and other solutes as well as amino acid side chains, i.e. carboxylic
group of aspartic acid. Water also strongly absorbs in the region
derivative transformation were calculated including 25 points on either 1600–1650 cm−1. The structure of the PLS-DA regression vector was
side of the centre point of the window. In general, the performance of very different from that of the PLS-DA model developed for the verifica-
the PLS-DA models was found to be dependent on the width of the win- tion of grass feeding even though the two pairs of classes (GRASS/NO
dow (number of data points) used for the computation of the smooth- GRASS and PASTURE/INDOORS) are very similar. This behaviour de-
ing and the 2nd derivative. The sensitivity (% of corrected pends on the different data pre-processing applied in the two models.
classification in class GRASS) and specificity (% of correct classification For model 3, since major differences were highlighted between win-
in class NO GRASS) of the model was very high in internal validation ter and spring samples, only the latter ones were used to develop the bi-
with approximately 88% of the samples of the class GRASS and 100% nary classification model. A classification model based on winter milk
of the samples of the class NO GRASS correctly assigned to their actual was not developed because of the few number of samples available
classes. In external validation, the sensitivity was ≈88% and the speci- (29). The milk samples from the farms converting to organic farming
ficity was ≈83%. The two samples from class NO GRASS that were incor- were included in the class ORGANIC. A total of 87 samples were includ-
rectly predicted as GRASS in external validation were collected in ed in the model with 69 samples in the class CONVENTIONAL and 18 in
February whereas the two samples from class GRASS incorrectly pre- the class ORGANIC. The most accurate model was obtained after mean-
dicted as NO GRASS were collected in spring. That means that all the centering, smoothing (25 points) and 2nd derivative transformation of
samples of the class NO GRASS collected in spring were correctly classi- the raw spectra and application of 1 OSC component and in external
fied by the model. Indeed, for authentication purposes, the ability to ac- validation gave 94% of correct classification in class CONVENTIONAL
curately predict whether the cows have been fed grass or not is by far and 80% in the class ORGANIC (Table 3). The PLS-DA regression
more important on samples collected during the grazing period than vectors gave very high scores (in absolute values) for bands around
on samples collected during the winter indoor period when fresh 1640 cm− 1 (positive score) and around 1592 cm− 1 and 1695 cm− 1
grass is not available. In Fig. 4, the regression vectors of class GRASS (negative scores) (data not shown). The scores associated to these
for the PLS-DA model described above is shown (the regression vector bands were 1–2 orders of magnitude higher than any other wave-
of class NO GRASS is not shown since it is the mirror-image of that of lengths which implies they are by far most important for the discrimina-
class GRASS). The regression vector can be regarded as the weighted tion between organic and conventional milk samples.
sum of the loadings included in the model. It gives therefore informa-
tion on the importance of each single variable for the model prediction. 3.4. FTIR vs FA profiling
Variables that have lower scores in the regression vector contribute
less in the prediction. From the regression vector it appears that In a previous part of this study, using the same set of samples as that
bands around 1175 cm − 1 , 1215 cm − 1 , 1260 cm − 1 , 1375 cm − 1 , used here, we found that it is also possible to highly satisfactorily au-
1500 cm−1, 1550 cm−1, 1750 cm−1 and the region 2800–3000 cm−1 thenticate fresh grass feeding by means of FA fingerprinting (Capuano
are important for the prediction. The last regions and the bands around et al., submitted for publication). In fact, the results achieved using FA
1175 cm−1 and 1750 cm−1 are attributed to the absorption by milk fat profiling were better, in terms of sensitivity and specificity than those
and that would confirm what emerged from the results reported in achieved by FTIR reported in the present paper (≈100% in external val-
Table 2. The region of 1500–900 cm−1 is called the fingerprint region idation). However, the analytical procedure required to obtain the FA
and refers to C\O and C\C stretching modes. The bands around profile is relatively time consuming, labour intensive, requires expen-
1199–1474 cm−1 are due to the bending modes of O\C\H, C\C\H sive instruments and trained personnel and makes use of chemicals.
and C\O\H bonds. In comparison with the regions corresponding to FTIR on the other hand is a more rapid, high-throughput technique, it
the absorption of fats, the regions corresponding to the absorption of is non-destructive, does not require specialized personnel, is environ-
proteins and of lactose seem to contribute less to the PLS-DA models. mental friendly since it does not require solvents and chemicals and
For model 2, the groups P1 and P2 were merged together in the class does not produce waste and may also be applied in situ. In addition,
PASTURE whereas the groups GI, NG and W were merged together in FTIR is a common and widespread technique in the dairy routine QA
the class INDOORS. The best performing PLS-DA model was obtained laboratories, being widely applied to dairy quality control, milk pay-
after Pareto scaling of the data, smoothing and 2nd derivative transfor- ment and herd improvement purposes. Based on those practical as-
mation of the raw data and the application of one OSC component. The pects, FTIR might be used as screening tool to verify the fresh grass
smoothing and the 2nd derivative were calculated including 25 points feeding of cows. The samples that are claimed from cows that have
on either side of the centre point of the window. The results were been fed fresh grass but predicted by the FTIR analysis otherwise can
64 E. Capuano et al. / Food Research International 60 (2014) 59–65

% correctly classified in CLASS 2c

81% (13/16)
83% (10/12)

80% (4/5)
% correctly classified in CLASS 1b

91% (12/13)
88% (15/17)

94% (16/17)

Fig. 4. Regression vector for the class GRASS in the PLS-DA model (autoscaled data,
smoothing, second derivative and one OSC component).

then be subsequently confirmed by FA profiling. From this standpoint it


RMSECV

is remarkable that the FTIR model developed herein exhibits an espe-


0.206
0.266
0.221

cially good specificity for the spring samples, which means that the per-
centage of misclassifications is very low for the samples from cows that
have not been fed fresh grass during spring. This would reduce in prac-
0.825
0.718
0.697
r2a

tice the rate of mislabelling detection failure. On the other hand, the di-
rect authentication of pasture grazing is not achievable with acceptable
Explained variance %

accuracy by FTIR combined with chemometrics on samples collected in


spring (the same was found with FA profiling), but the verification of
fresh grass feeding might be the starting point for the authentication
of pasture grazing provided that additional biomarkers are considered.
86
29

90

4. Conclusions
No. of factors

In this feasibility study we showed that FTIR spectra of raw, herd bulk
milk contain valuable information on cows' diet that can be used for the
6
3

authentication of cow feeding regime and herd management system.


Mean-centering, smoothing (25 points), 2nd derivative (25 points) and 1 OSC

PLS-DA classification models developed to predict whether cows have


Pareto scaling, smoothing (25 points), 2nd derivative (25 points) and 1 OSC

been fed fresh grass exhibited satisfactory accuracy. An analogous binary


Autoscaling, smoothing (25 points), 2nd derivative (25 points) and 1 OSC

model developed to predict whether cows have been on pasture or not


Class 2 is NO GRASS for model 1, INDOORS for model 2 and ORGANIC/BIODYNAMIC for model 3.

showed comparable accuracy even though the model specificity reduced


when applied on spring samples only. The discrimination between
Description of the optimal PLS-DA models and their performance during external validation.

organic and conventional milk collected in spring seems also feasible by


Class 1 is GRASS for model 1, PASTURE for model 2 and CONVENTIONAL for model 3.

FTIR but the relatively few samples in the class ORGANIC would call
for more cautious conclusions and validation with a larger and more
balanced dataset. The results reported in this study suggest that FTIR
spectroscopy might be used as a screening technique for the authentica-
tion of fresh grass feeding before more accurate but time-consuming
techniques are applied (e.g. FA profiling) for confirmatory purposes.
However, the validation of the proposed model with a truly independent
validation set is mandatory to support these conclusions.
RMSECV = Root Mean Square Error of Cross Validation.
Data pre-processing

Acknowledgements

The authors gratefully acknowledge the Dutch Ministry of Economic


Affairs and Innovation for financial support, Grishja van der Veer for the
OSC = orthogonal signal correction.

fruitful discussion on the modelling and Nisha Shetty for reading the
Obtained in cross validation.

manuscript. They also acknowledge Anjo Elgerma and the farmers for
collecting and providing the milk samples.
Model 3: Organic farming
Model 2: Pasture feeding
Model 1: Fresh grass

References
PLS-DA model

Andreotti, G., Trivellone, E., Lamanna, R., Di Luccia, A., & Motta, A. (2000). Milk identifica-
tion of different species: 13C NMR spectroscopy of triacylglycerols from cows and
Table 3

buffaloes' milks. Journal of Dairy Science, 83, 2432–2437.


Berrueta, L. A., Alonso-Salces, R. M., & Héberger, K. (2007). Supervised pattern recognition
b
a

in food analysis. Journal of Chromatography A, 1158(1–2), 196–214.


E. Capuano et al. / Food Research International 60 (2014) 59–65 65

Borin, A., Ferrao, M. F., Mello, C., Althmann Maretto, D., & Poppi, R. J. (2006). Least-squares Libnau, F. O., Kvalheim, O. M., Christy, A. A., & Toft, J. (1994). Spectra of water in the near-
support vector machines and near infrared spectroscopy for quantification of com- and mid-infrared region. Vibrational Spectroscopy, 7, 24–2543.
mon adulterants in powdered milk. Analytica Chimica Acta, 579, 25–32. Lin, J., & Brown, C. W. (1992). Near-IR spectroscopic determination of NaCl in aqueous so-
Boscaini, E., van Ruth, S., Biasioli, F., Gasperi, F., & Mark, T. D. (2003). Gas chromatogra- lution. Applied Spectroscopy, 46, 1809–1815.
phy–olfactometry (GC–O) and proton transfer reaction–mass spectrometry (PTR– Lipp, M. (1996). Determination of the adulteration of butter fat by its triglyceride compo-
MS) analysis of the flavour profile of grana padano, parmigiano reggiano, and grana sition obtained by GC. A comparison of the suitability of PLS and neural networks.
trentino cheeses. Journal of Agricultural and Food Chemistry, 51, 1782–1790. Food Chemistry, 55, 389–395.
Capuano, E., van der Veer, G., Boerrigter-Eenling, R., Elgersma, A., Rademaker, J., Sterian, Maamouri, O., Rouissi, H., Dridi, S., Kammoun, M., De Baerdemaeker, J., & Karoui, R.
A., et al. (2013). Verification of Fresh Grass Feeding, Pasture Grazing and Organic Farm- (2008). Mid infrared attenuated total reflection spectroscopy as a rapid tool to assess
ing by Cows Farm Milk Fatty Acid Profile. (submitted for publication). the quality of Sicilo-Sarde ewe's milk during the lactation period after replacing soy-
Cecchinato, A., De Marchi, M., Gallo, L., Bittante, G., & Cainier, P. (2009). Mid-infrared bean meal with scotch bean in the feed ration. Food Chemistry, 106, 361–368.
spectroscopy predictions as indicator traits in breeding programs for enhanced coag- Manne, R. (1989). Analysis of two partial-least-squares algorithms for multivariate cali-
ulation properties of milk. Food Chemistry, 92(10), 5304–5313. bration. Chemometrics and Intelligent Laboratory Systems, 2, 187–197.
Coppa, M., Martin, B., Agabriel, C., Chassaing, C., Sibra, C., Constant, I., et al. (2012). Au- Maraboli, A., Cattaneo, P., Maria, T., & Giangiacomo, R. (2002). Detection of vegetable pro-
thentication of cow feeding and geographic origin on milk using visible and teins from soy, pea, and wheat isolates in milk powder by near infrared spectroscopy.
near-infrared spectroscopy. Journal of Dairy Science, 95, 5544–5551. Journal of Near Infrared Spectroscopy, 10(1), 63–69.
Couvreur, S., Hurtaud, C., Lopez, C., Delaby, L., & Peyraud, J. L. (2006). The linear relation- McQueen, D. H., Wilson, R., Kinnunen, A., & Jensen, E. P. (1995). Comparison of two infra-
ship between the proportion of fresh grass in the cow diet, milk fatty acid composi- red spectroscopic methods for cheese analysis. Talanta, 42, 2007–2015.
tion, and butter properties. Journal of Dairy Science, 89, 1956–1969. Morales-Almaraz, E., Soldado, A., Gonzales, A., Martinez-Fernandez, A., Dominguez-Vara,
Cruz, A. G., Cadena, R. S., Alvaro, M. B. V. B., Sant'Ana, A. S., Oliveira, C. A. F., Faria, J. A. F., I., de la Rosa-Delgado, B., et al. (2010). Improving the fatty acid profile of dairy cow
et al. (2013). Assessing the use of different chemometric techniques to discriminate milk by combining grazing with feeding of total mixed ration. Journal of Dairy
low-fat and full-fat yogurts. LWT — Food Science and Technology, 50, 210–214. Research, 77, 225–230.
De Marchi, M., dalZotto, R., Cassandro, M., & Bittante, G. (2007). Milk coagulation ability of Nicolaou, N., Xu, Y., & Goodacre, R. (2011). MALDI-MS and multivariate analysis for the
five dairy cattle breeds. Journal of Dairy Science, 90(8), 3986–3992. detection and quantification of different milk species. Analytical and Bioanalytical
Dewhurst, R. J., Shingfield, K. J., Lee, M. R. F., & Scollan, N. D. (2006). Increasing the concen- Chemistry, 399, 3491–3502.
trations of beneficial polyunsaturated fatty acids in milk produced by dairy cows in Ntakatsane, M. P., Yang, X. Q., Lin, M., Liu, X. M., & Zhou, P. (2011). Short communication:
high-forage systems. Animal Feed Science and Technology, 131, 168–206. Suitability of fluorescence spectroscopy for characterization of commercial milk of
Dufour, E., & Riaublanc, A. (1997). Potentiality of spectroscopic methods for the character- different composition and origin. Journal of Dairy Science, 94, 5375–5380.
isation of dairy products, I — Front-face fluorescence study of raw, heated and Pillonel, L., Ampuero, S., Tabacchi, R., & Bosset, J. O. (2003). Analytical methods for the de-
homogenised milks. Le Lait, 77, 657–670. termination of the geographic origin of Emmental cheese: Volatile compounds by
Dufour, E., Robert, P., Renard, D., & Llamas, L. (1998). Investigation of β-lactoglobulin ge- GC/MS-FID and electronic nose. European Food Research and Technology, 216, 179–183.
lation in water/ethanol solutions. International Dairy Journal, 8, 87–93. Rodriguez-Saona, L. E., & Allendorf, M. E. (2011). Use of FTIR for rapid authentication and
Dupuy, N., Duponchel, L., Huvenne, J. P., Sombret, B., & Legrand, P. (1996). Classification of detection of adulteration of food. Annual Review in Food Science and Technology, 2,
edible fats and oils by principal component analysis of Fourier transform infrared 467–483.
spectra. Food Chemistry, 57, 245–251. Rutten, M. J. M., Bovenhuis, H., Heck, J. M. L., & van Arendonk, J. A.M. (2010). Predicting
Elgersma, A. (2012). New developments in The Netherlands: Dairies reward grazing be- bovine milk protein composition based on Fourier transform infrared spectra.
cause of public perception. Grassland Science in Europe, 17, 420–422. Journal of Dairy Science, 94, 5683–5690.
Elgersma, A., Tamminga, S., & Ellen, G. (2006). Modifying milk composition through for- Rutten, M. J. M., Bovenhuis, H., Hettinga, K. A., van Valenberg, H. J. F., & van Arendonk, J.
age — A review. Animal Feed Science and Technology, 131, 207–225. A.M. (2009). Predicting bovine milk fat composition using infrared spectroscopy
Elisses, J., & Zee, B. (1980). The freezing point of authentic cows' milk and farm tank milk based on milk samples collected in winter and summer. Journal of Dairy Science, 92,
in The Netherlands. Netherlands Milk and Dairy Journal, 34, 162–180. 6202–6209.
Ellis, K. A., Innocent, G., Grove-White, D., Cripps, P., McLean, W. G., Howard, C. V., et al. Safar, M., Bertrand, D., Roberts, P., Devaux, M. F., & Genot, C. (1994). Characterization of
(2006). Comparing the fatty acid composition of organic and conventional milk. edible oils, butters and margarines by Fourier transform infrared spectroscopy with
Journal of Dairy Science, 89, 1938–1950. attenuated total reflectance. Journal of the American Oil Chemists' Society, 71(4), 371.
Elschner, M., Jacobi, U., Buchberger, J., & Grun, E. (1997). Untersuchungen zum Savitzky, A., & Golay, M. J. E. (1964). Smoothing and differentiation of data by simplified
Gefrierpunkt von Kuhmilch am Beispiel eines grossen Milcherzeugerbetriebes. Teil least squares procedures. Analytical Chemistry, 36, 1627–1639.
2. Lebensmittelindustrie und Milchwirtschaft, 5, 188–192. Slots, T., Butler, G., Leifert, C., Kristensen, T., Skibsted, L. H., & Nielsen, J. H. (2008). Poten-
Frau, M., Massanet, J., Rossello', C., Simal, S., & Canellas, J. (1997). Evolution of free amino tials to differentiate milk composition by different feeding strategies. Journal of Dairy
acid content during ripening of Mahon cheese. Food Chemistry, 60, 651–657. Science, 92, 2057–2066.
Galle, S. A., Koot, A., Soukoulis, C., Cappellin, L., Biasioli, F., Alewijn, M., et al. (2011). Typ- Smiddy, M.A., Huppertz, T., & van Ruth, S. (2012). Triacylglycerol and melting profiles of
icality and geographical origin markers of protected origin cheese from The milk fat from several species. International Dairy Journal, 24(2), 64–69.
Netherlands revealed by PTR–MS. Journal of Agricultural and Food Chemistry, 59(6), Sorensen, L. K., Lund, M., & Jull, J. (2003). Accuracy of Fourier transform infrared spectrometry
2554–2563. in determination of casein in dairy cows' milk. Journal of Dairy Research, 70, 445–452.
Gorry, P. A. (1990). General least-squares smoothing and differentiation by the convolu- Souza, S. S., Cruz, A. G., Walter, E. H. M., Faria, J. A. F., Celeghini, R. M. S., Ferreira, M. M. C.,
tion (Savitzky–Golay) method. Analytical Chemistry, 62, 570–573. et al. (2011). Monitoring the authenticity of Brazilian UHT milk: A chemometric ap-
Hashimoto, A., & Kameoka, T. (2008). Applications of infrared spectroscopy to biochemi- proach. Food Chemistry, 124, 692–695.
cal, food, and agricultural processes. Applied Spectroscopy Reviews, 43, 416–451. Van de Voort, F. R., Sedaman, J., Emo, G., & Ismail, A. A. (1992). Assessment of Fourier
Henno, M., Ots, M., Joudu, I., Kaart, T., & Kart, O. (2008). Factors affecting the freezing point transform infrared analysis of milk. Journal of AOAC International, 75, 780–785.
stability of milk from individual cows. International Dairy Journal, 18, 210–215. White, S. L., Bertrand, A., Wade, M. R., Washburn, S. P., Green, J. T., Jr., & Jenkins, T. C.
Karoui, R., Bosset, J. -O., Mazerolles, G., Kulmyrzaev, A., & Dufour, E. (2005). Monitoring (2001). Comparison of fatty acid content of Milk from Jersey and Holstein cows con-
the geographic origin of both experimental French Jura hard cheeses and Swiss suming pasture or a total mixed ration. Journal of Dairy Science, 84, 2295–2301.
Gruye're and L'Etivaz PDO cheeses using mid-infrared and fluorescence spectros- Woodcock, T., Downey, G., & O'Donnell, C. P. (2008). Confirmation of declared provenance
copies: A preliminary investigation. International Dairy Journal, 15, 275–286. of European extra virgin olive oil samples by NIR spectroscopy. Journal of Agricultural
Koca, N., Kocaoglu-Vurma, N. A., Harper, W. J., & Rodriguez-Saona, L. E. (2010). Applica- and Food Chemistry, 56, 11520–11525.
tion of temperature controlled attenuated total reflectance-mid-infrared (ATR-MIR) Yang, H., & Irudayaraj, J. (2000). Characterization of semisolid fats and edible oils by
spectroscopy for rapid estimation of butter adulteration. Food Chemistry, 121, Fourier transform infrared photoacoustic spectroscopy. Journal of the American Oil
778–782. Chemists' Society, 77, 291–295.

You might also like