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

Food Chemistry 181 (2015) 31–37

Contents lists available at ScienceDirect

Food Chemistry
journal homepage: www.elsevier.com/locate/foodchem

Analytical Methods

Development and analytical validation of a screening method


for simultaneous detection of five adulterants in raw milk using
mid-infrared spectroscopy and PLS-DA
Bruno G. Botelho a, Nádia Reis b, Leandro S. Oliveira b, Marcelo M. Sena a,c,⇑
a
Departamento de Química, ICEx, Universidade Federal de Minas Gerais, 31270-901 Belo Horizonte, MG, Brazil
b
Faculdade de Farmácia, Universidade Federal de Minas Gerais, 31270-901 Belo Horizonte, MG, Brazil
c
Instituto Nacional de Ciência e Tecnologia em Bioanalítica (INCT Bio), 13083-970, Campinas, SP, Brazil

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

Article history: This paper proposed a new screening method for the simultaneous detection of five common adulterants
Received 30 August 2014 in raw cow milk by using attenuated total reflectance (ATR) mid infrared spectroscopy and multivariate
Received in revised form 11 January 2015 supervised classification (partial least squares discrimination analysis – PLSDA). The method was able to
Accepted 14 February 2015
detect the presence of the adulterants water, starch, sodium citrate, formaldehyde and sucrose in milk
Available online 20 February 2015
samples containing from one up to five of these analytes, in the range of 0.5–10% w/v. A multivariate
qualitative validation was performed, estimating specific figures of merit, such as false positive and false
Keywords:
negative rates, selectivity, specificity and efficiency rates, accordance and concordance. The proposed
Adulteration
Milk
method does not need any sample pretreatment, requires a small amount of sample (30 lL), is fast
Qualitative validation and simple, being suitable for the control of raw milk in a dairy industry or for the quality inspection
Partial least squares discriminant analysis of commercialized milk.
FT-IR Ó 2015 Elsevier Ltd. All rights reserved.
Screening analysis

1. Introduction citrate, sodium hydroxide, sodium chloride, sucrose, phosphates,


carbonates, bicarbonates and hydrogen peroxide to correct milk
Milk can be defined as the fluid secreted by the female of all defects, such as high acidity and microbial growth, and to increase
mammalian species, primarily to meet the complete nutritional its volume (Hoorfar, 2012). In 2013, the ‘‘Operação Leite Com-
requirement of the neonate, and it is the major component in pen$ado’’ (Compensated Milk Operation) revealed the utilization
human diets in many parts of the world. In 2011, more than 735 of fertilizers containing urea and formaldehyde to mask water
million tons of milk were produced around the world (including addition in milk (UOL, 2014). Souza et al. (2011) have analyzed
cow, buffalo, sheep, goat and camel milk). USA was the largest pro- 100 samples of Brazilian UHT milk and detected high rates of
ducer, with 87.5 millions of tons, and Brazil was the fifth major non-conformities, such as 55% for urine, 44% for formaldehyde,
producer (31.7 millions of tons). However, as fast as the production 30% for hydrogen peroxide and 12% for chlorine.
and demand for milk have grown, a larger frequency of sophisticat- Currently, the detection of adulterants in milk is performed by
ed milk adulterations has been reported in different countries using physicochemical methods based on specific reactions or dif-
(EMBRAPA, 2014). ferences in the freezing point and specific gravity (IAL, 2005;
The practice of unscrupulous producers that add adulterating MAPA, 2006). These methods are time consuming, labor-intensive
substances in milk in order to raise their profits is not a new and may not yield accurate results. In the last years, the use of
subject, but recently, it has attracted more attention worldwide. chemometric methods of supervised classification and/or multi-
In Brazil, two major milk adulteration scandals were recently variate calibration combined with near or mid infrared spec-
reported by the Federal Police. In 2007, the so-called ‘‘Operação troscopy has allowed the development of rapid and non-
Ouro Branco’’ (White Gold Operation) reported the use of sodium destructive methods to detect and/or quantify adulterants in milk,
such as urea, synthetic urine, synthetic milk, whey, hydrogen per-
oxide, water, melamine and ammonium nitrate (Kasemsumran,
⇑ Corresponding author at: Departamento de Química, ICEx, Universidade Federal Thanapase, & Kiatsoonthon, 2007; Nicolaou, Xu, & Goodcare,
de Minas Gerais, 31270-901 Belo Horizonte, MG, Brazil. Tel.: +55 31 34096389; fax:
2010; Santos, Pereira-Filho, & Rodriguez-Saona, 2013a, 2013b;
+55 31 34095700.
E-mail address: marcsen@ufmg.br (M.M. Sena).
Zhang et al., 2014). An alternative method based on image analysis

http://dx.doi.org/10.1016/j.foodchem.2015.02.077
0308-8146/Ó 2015 Elsevier Ltd. All rights reserved.
32 B.G. Botelho et al. / Food Chemistry 181 (2015) 31–37

has also been applied for detecting and quantifying water and 2. Materials and methods
sodium hydroxide as adulterants in milk (Santos, Wentzell, &
Pereira-Filho, 2012). 2.1. Instruments and software
Once the addition of any kind of foreign substance in raw milk is
prohibited in Brazil, at any level, only qualitative methods are Mid infrared spectra were obtained using a Shimadzu IR Affi-
described in Brazilian official guidelines for detection of adulter- nity1 spectrophotometer (Shimadzu Corp., Kyoto, Japan) in
ation in milk (IAL, 2005; MAPA, 2006). It is more important to iden- attenuated total reflectance (ATR) mode, equipped with a ZnSe cell.
tify than to quantify possible adulterants. Qualitative analytical Data were handled using MATLAB software, version 7.13 (The
methods provide a dichotomous variable as answer (1/0, yes/no, MathWorks, Natick, MA, USA), the PLS-DA routine came from the
presence/absence, pass/fail, accepted/not accepted). These meth- PLS Toolbox, version 6.5 (Eigenvector Technologies, Manson, WA,
ods have been considered a declining branch of analytical chem- USA), and Minitab 16 Demo Version (Minitab, Inc, State College,
istry in recent decades, due to their association with the PA, USA) was used for the mixture design.
identification of ionic species using chemical reactions. However,
with the advances in instrumental analysis and the increasing
demands for total indices and yes/no binary responses, qualitative 2.2. Samples and reagents
methods are now identified as promising tools to develop and
implement reliable vanguard (screening)-rearguard (confirmation, Fifteen blank samples (300 mL) of whole cow’s milk were pro-
quality control and information expansion) analytical strategies vided by the model experimental farm from the Veterinary School
(Valcárcel & Cárdenas, 2005). Rapid sample screening methods of the Universidade Federal de Minas Gerais (UFMG, Belo Horizon-
can confirm the presence of the analytes of interest, simplifying te) and used as replicates in this investigation. Since its production
the analytical process, saving time and costs by reducing the is controlled, it is guaranteed that no substances are added to this
number of samples that demand the use of a more laborious milk. Thus, these blank samples did not contain any amount of
quantitative analysis (Magalhães et al., 2012; Pikeemaat, 2009). formaldehyde, sucrose or starch. Water is naturally present in milk,
Similarly to quantitative methods, qualitative methods present being its major component (85–90% w/w), and sodium citrate is
in the official guidelines must be submitted to an analytical valida- commonly present in milk, but in lower concentrations, bellow
tion procedure. However, this subject has received less attention in 0.1% w/w (Fox & McSweeney, 1998). The milk was stored at
the literature. As the response provided by qualitative methods is 18 °C in 500 mL glass bottles, and heated up using a water bath
discrete, statistical tests and procedures used differ from the ones until it reached ambient temperature, daily. Sodium citrate,
used in quantitative validation. In 2002, European Community formaldehyde, starch, sucrose (all of analytical grade) and distilled
released the first official guidelines for qualitative analytical water were used as adulterants. For the preparation of the samples,
validation (EC, 2002). Since then, the interest on this subject has all the adulterants were weighed in an analytical balance
continuously grown (Gondim, Junqueira, & Souza, 2011; NATA, (±0.0001 g). For the samples containing only one adulterant, water,
2013; Trullols, Ruisanchez, & Rius, 2004). With the increasing sucrose, formaldehyde, starch and sodium citrate were added in a
number of developed chemometric methods of supervised classifi- concentration range from 0.5% to 10.0% w/v, with 0.5% increments,
cation, the need to discuss multivariate qualitative validation has resulting in 20 adulterated samples for each adulterant. For the
increased, as it has been the case of multivariate quantitative samples with more than one adulterant, a simplex centroid mix-
validation, which has gained more attention in the last years ture design with five components, with the concentrations of each
(Botelho, Mendes, & Sena, 2013). Nevertheless, this subject is still adulterant ranging from 0.5% to 10.0%, and a maximum of 10.0% of
almost completely absent in the literature. total adulteration. This mixture design resulted in 71 samples, con-
The main objective of this paper was to develop a rapid and taining from two to five adulterants simultaneously (30 binary, 30
non-destructive qualitative method for the simultaneous detection tertiary, 5 quaternary and 6 pentenary mixtures) and their compo-
of five adulterants in raw cow milk, using attenuated total reflec- sitions are presented in Table 1 of Supplemental materials.
tion (ATR) mid infrared spectroscopy and partial least squares dis-
criminant analysis (PLS-DA). There were chosen five adulterants 2.3. Procedure
recently identified in Brazilian milk: water, starch, sodium citrate,
formaldehyde and sucrose. Water is the most common adulterant All the samples were analyzed directly, without any pre-treat-
used for increasing the milk volume by dilution, resulting in the ment, placing approximately 30 lL in the ATR accessory. The spec-
decrease of nutrition substances, such as protein and solid content. tra were collected as an average of 20 scans, from 600 to
Starch is used as thickener, increasing the solid content of adulter- 4000 cm1. Before each measurement, a background correction
ated milk. Citrate is used as stabilizer and preservative, aiming to was performed to avoid atmospheric interference and reduce
avoid precipitation of nutrients. Formaldehyde is used for preserv- instrumental noise. After each spectrum acquisition, the accessory
ing adulterated milk from microbial contamination, since the was cleaned with acetone P.A.
reduction of nutritional value due to water addition increases this
risk, and sucrose is used to restore the normal analytical values of
adulterated milk in physicochemical tests and to improve its sen- 2.4. Qualitative analytical validation
sory properties. Considering the milk adulterations usually report-
ed, the simultaneous presence of different adulterants is common, The first step in qualitative multivariate validation is to deter-
but most of the chemometric methods found in the literature have mine the limit whether the sample is classified as 0 (negative/
determined only one adulterant at a time or at most binary mix- not accepted/absence) or 1 (positive/accepted/presence). Modern
tures (Kasemsumran et al., 2007; Santos et al., 2013b; Zhang qualitative methods usually rely on instrumental reading for clas-
et al., 2014). In this paper, mixtures of higher order (up five adul- sifying a sample as positive or negative, and this limit value is
terants) were also analyzed for specific detection of each adulter- called threshold limit (TL), estimated using the Bayes Theory
ant. The developed method was also validated, by the estimate of (Gondim et al., 2011; Pulido, Ruisanchez, Boqué, & Rius, 2003).
specific figures of merit (FOM) used in qualitative analytical valida- The TL estimate has been extended from univariate to multivariate
tion, such as sensitivity, specificity, efficiency rate, accordance and methods, specifically for PLS-DA (Wise et al., 2006). Assuming that
concordance. all the samples in the training test belong to one of the proposed
B.G. Botelho et al. / Food Chemistry 181 (2015) 31–37 33

classes, i.e., class A or class B (here, we assumed only two classes estimation, the probability that two samples provide the same
for simplification), it can be stated: result is calculated, and then this probability is averaged over all
batches. The ACO can be estimated by the following equation:
PðAjyÞ þ PðBjyÞ ¼ 1 ð1Þ
fkðk  1Þ þ ðn  kÞðn  k  1Þg
Using the Bayes’ Theory, it is possible to estimate the probabil- ACO ¼ ð10Þ
nðn  1Þ
ity that a sample is from class A given a particular value of y, P(A|y),
from the following equation: where n is the total number of samples in each batch and k is the
number of positive results in each batch. CON is the qualitative
PðyjAÞ  PðAÞ
PðAjyÞ ¼ ð2Þ FOM equivalent to the quantitative estimation of intermediate pre-
½PðyjAÞ  PðAÞ þ PðyjBÞ  PðBÞ
cision or reproducibility. It estimates the probability of two testing
where P(A) and P(B) are the probabilities for the observation of class samples providing the same results in different laboratories or
A or B in the future. If we assume that the probability of observing A batches in intermediary precision conditions. The CON can be esti-
or B is similar to how many samples of A and B were in the original mated using the following equation:
training set, we can reduce this to: 2½kðk  nbÞ þ nbðnb  1Þ  ACO½nbðn  1Þ
CON ¼ ð11Þ
PðyjAÞ n2 bðb  1Þ
PðAjyÞ ¼ ð3Þ
½PðyjAÞ þ PðyjBÞ where b is the number of laboratories or analytical batches (Gondim
et al., 2011; Langton, Chevennement, Nagelkerke, & Lombard,
and
2002).
PðyjBÞ
PðBjyÞ ¼ ð4Þ
½PðyjAÞ þ PðyjBÞ 3. Results and discussion
These two distributions typically intersect in a unique value,
3.1. Mid infrared spectra
which leads to a single point where both P(B|y) and P(A|y) are
0.5. This point is selected as the threshold limit.
The spectra of all the samples are shown in Fig. 1. By observing
After the TL has been established, the next step is to determine
these spectra, it is possible to attribute the most intense absorption
the numbers of false positive (FP) and false negative (FN) samples.
peak around 3600–3000 cm1, which can be associated with
FP is the number of samples that do not belong to a class classified
strong OAH stretching vibrations. The two small peaks near
as belonging to it, i.e., a sample that is not adulterated been classi-
2800 cm1 are related to milk’s fatty acid CH2 stretching. Another
fied as it is, and FN is the opposite parameter. Using FP and FN, the
peak, near 1700 cm1, is characteristic from amide I and II, which
following FOM are estimated, False Negative Rate (FNR), False Posi-
can be related to milk protein content (Nicolaou et al., 2010;
tive Rate (FPR), Selectivity Rate (STR), Specificity Rate (SPR) and
Santos et al., 2013a). Initially, the region between 2430 and
Efficiency Rate (EFR). FPR is defined as the ratio between FP and
2230 cm1 was removed due to the interference of atmospheric
the sum of FP and the total number of known negative samples
CO2. Even using the instrumental atmospheric correction, some
(TN).
samples presented a characteristic peak in this spectral region, so
FP it was removed from all samples.
FPR ¼ ð5Þ
FP þ TN Fig. 2(a) shows a spectrum of a blank milk sample, without the
addition of any adulterant. As stated before, two distinguishable
Analogously, the FNR is defined as the ratio between FN and the
peaks can be noticed, related to water and protein content.
sum of FN and the total number of known positive samples (TP).
Fig. 2(b) shows the spectrum of a milk sample adulterated with
FN 10.0% v/v of water, and no significant changes can be visually not-
FNR ¼ ð6Þ ed, compared with Fig. 2(a). Fig. 2(c) shows a spectrum for milk
FN þ TP
with 10.0% w/v of starch. In comparison with Fig. 2(a), it is possible
The STR is defined as the ratio between TP and the sum of TP to note slight changes in the fingerprint region, between 1200 and
and FN, and the SPR in the ratio between TN and the sum of TN 900 cm1. The absorbance in this region can be associated with
and FP. In addition, EFR can be defined as the difference between skeletal mode vibrations of the a-1,4 glycosidic linkage, CAH bend-
the total of results (100%) and the sum of FPR and FNR. ing, CAOAH bending and CAO and CAC stretching (Kizil,
TP Irudayaraj, & Seethraman, 2002). Fig. 2(d) shows the spectrum of
STR ¼  100 ð7Þ a sample adulterated with sodium citrate 10.0% w/v. Two charac-
TP þ FN
teristics peaks around 1400 and 1600 cm1 can be related to sym-
TN metric and asymmetric stretching of CAO bonds from carboxylate
SPR ¼  100 ð8Þ groups. Fig. 2(e) shows the spectrum of the adulteration of a blank
TN þ FP
milk with 10.0% v/v of formaldehyde. A distinguishable peak
around 1000 cm1 can be noticed. This absorption peak can be
EFR ¼ 100  ðFPR þ FNRÞ ð9Þ
related to formaldehyde CH2 rocking and wagging bending vibra-
Although some FOM used in qualitative and quantitative valida- tions (Pavia, Lampman, Kriz, & Vyvyan, 2008). Fig. 2(f) shows a
tions have the same name, the concepts attached to them and their spectrum of a sample adulterated with 10.0% w/v of sucrose. The
evaluation are different. While FPR, FNR and EFR express the true- presence of sucrose can be associated with the appearance of sev-
ness in qualitative analysis, STR and SPR are considered FOM relat- eral peaks in the fingerprint region (near 1200–1000 cm1), associ-
ed to the selectivity of the method (Gondim et al., 2011; Trullols ated with CAH and CAO stretching, ketone C@O stretching and
et al., 2004). Another two FOM, accordance (ACO) and concordance bending, and CAO stretching and bending (Tewari & Malik,
(CON), express the precision of qualitative methods. ACO is the 2007). The region between 2000–900 cm1 presented the greatest
probability that two identical samples provide the same result variations and is considered to be the most discriminative, contain-
(both found positive or negative) when analyzed by the same ing characteristic bands for almost all the adulterants evaluated.
laboratory or batch, under repeatability conditions. For this Thus, local models were tested and presented smaller classification
34 B.G. Botelho et al. / Food Chemistry 181 (2015) 31–37

Fig. 1. Mid infrared spectra of all the analyzed milk samples.

Fig. 2. Mid infrared spectra of (a) a blank milk sample; (b) a sample adulterated with water (10% w/v); (c) a sample adulterated with starch (10% w/v); (d) a sample
adulterated with sodium citrate (10% w/v); (e) a sample adulterated with formaldehyde (10% w/v); (f) a sample adulterated with sucrose (10% w/v). Spectral region used in
the models is marked by the dashed squares.

errors when compared to the full spectra models. All the models 25% of the samples (38) of each model was selected for the test set,
developed and validated in this paper used the above cited region. chosen in equally spaced intervals of composition according to the
experimental design. To select the best number of latent variables
3.2. PLS-DA models (LV), venetian blind cross validation was used, and the LV number
that presented the smallest cross validation classification error
Five PLS1-DA models were built, one for each adulterant. Each (CVCE) was chosen (Table 1). Table 1 also presents the percentages
model was built using 155 samples (15 blank samples, 100 sam- of variance accounted for each model, which were 99.5% in the X
ples containing only one of the five adulterants, and 40 samples block for all the models, and between 69.3% and 77.0% in the Y
containing the adulterant of choice for that model in mixtures with block. The classification plots, with the y predicted values of all
others adulterants). Samples were split in training and test sets and samples, for all the models are shown in Fig. 3. In these plots, the
B.G. Botelho et al. / Food Chemistry 181 (2015) 31–37 35

Table 1 coefficients. Nevertheless, as pointed out by Brown and Green


Number of latent variables (LV), variance accounted for in X and Y blocks, and CVCE (2009), this interpretation should not rely on the regression vec-
for the developed PLS-DA models.
tors, since they are dependent on the samples in the calibra-
Adulterant Number of Explained Explained Cross validation tion/training set, on the implicit covariance of the components,
latent variance – variance – classification and on the signal to noise ratio of the data. A better tool for the
variables X (%) Y (%) error (CVCE)
spectral interpretation of PLS models is the variable importance
Water 9 99.5 76.2 0.07 in projection (VIP) scores, which measure the importance of each
Starch 8 99.5 72.5 0.07
Sodium 9 99.5 77.0 0.02
variable in the projection used by a particular PLS model via their
citrate coefficients in every component, together with the significance of
Formaldehyde 9 99.5 69.3 0.04 each component in regression (Chong & Jun, 2005). Fig. 4 shows
Sucrose 9 99.5 73.1 0.07 the VIP scores for each model. There is a close agreement between
the spectral discussion presented in Section 3.1 and the variables of
greater importance observed through the highest VIP scores in
data were ordered always starting with the samples not containing Fig. 4.
the adulterant of the model. The threshold for each class/model As can be seen in Fig. 4, all the VIP scores presented distinguish-
was estimated using the Bayes’ theorem (Wise et al., 2006) in order able peaks in regions specifically characteristic of each adulterant,
to minimize classification errors. The Bayesian threshold estima- such as the two peaks between 1080–1040 and 1015–980 cm1 of
tion assumes that the predicted Y variance follows a distribution citrate (Fig. 4(c)), the broad band with two peaks between around
similar to what will be observed for future samples. Plots of the 1100 and 1000 cm1 of formaldehyde (Fig. 4(d)), and the two
PLS-DA predictions as a function of different concentrations are major peaks at 1080 and 980 cm1 for sucrose (Fig. 4(e)).
also shown in Fig. 1 of Supplemental materials for two adulterants,
sodium citrate and water. 3.3. Multivariate qualitative validation
For all the models, the data were preprocessed with first deriva-
tive and Savitzky–Golay smoothing, followed by mean centering. As stated previously, multivariate validation of analytical meth-
Other preprocessing were tested, such as orthogonal signal correc- ods is still an evolving subject, not well established, and the esti-
tion (OSC), multiplicative scatter correction (MSC) and standard mation of FOM for qualitative multivariate methods has received
normal variate (SNV), but all of them provided higher CVCE. even less attention. In this paper, some basic FOM were estimated.
The spectral interpretation of PLS or PLS-DA models is As can be seen in Table 2, seven different FOM were estimated for
sometimes carried out through the observation of the regression the training and test sets.

Fig. 3. Classification plots for (a) water; (b) starch; (c) sodium citrate; (d) formaldehyde; (e) sucrose. The following symbols apply for the samples of each class: 5 – blank
samples; / – water adulteration; } – starch adulteration; s – sodium citrate adulteration; 4 – formaldehyde adulteration; h – sucrose adulteration. The full and empty
symbols represent the samples allocated in the training and test sets, respectively. The solid lines indicate the threshold values for each model.
36 B.G. Botelho et al. / Food Chemistry 181 (2015) 31–37

Fig. 4. Variable importance in projection (VIP) scores for the models detecting adulterations with: (a) water; (b) starch; (c) sodium citrate; (d) formaldehyde; (e) sucrose.

Table 2 and citrate being responsible for this high FPR. Formaldehyde
Estimated FOM for evaluating the developed methods for simultaneous detection of and sucrose models presented similar FPR (0.0% and 1.0%, respec-
adulterants in milk.
tively) and FNR (3.2% and 4.8%, respectively) in the training set;
FPR (%) FNR (%) STR (%) SPR (%) EFR (%) and identical values for the test set, with a FPR of 4.2% and a FNR
Training set of 0% for both analytes. The citrate model presented the best
Water 4.0 3.2 96.8 96.0 92.8 results, with FPR and FNR of 0% for training and test sets. According
Starch 3.0 4.8 95.2 97.0 92.2 to the Directive 657/2002 of the European Community (EC, 2002),
Sodium citrate 0.0 0.0 100.0 100.0 100.0
the FNR should be smaller than 5%. False negative results should be
Formaldehyde 0.0 3.2 96.8 100 96.8
Sucrose 1.0 4.8 96.8 99.0 95.7 more rigorously controlled for consumer’s safety, avoiding that an
adulterated sample be classified as non-adulterated. False positive
Test set
Water 11.5 0.0 93.8 88.5 82.8 results represent a lower concern, because all positive samples in a
Starch 8.0 16.7 83.3 92.0 75.3 screening test may be confirmed by a reference method. Except for
Sodium citrate 0.0 0.0 100.0 100.0 100.0 the starch model, all of the proposed methods are in accordance
Formaldehyde 4.2 0.0 93.8 95.8 89.6
with the above cited validation guide.
Sucrose 4.2 0.0 93.8 95.8 89.6
The more representative FOM of trueness, EFR, was also shown
FPR – False Positive Rate; FNR – False Negative Rate; STR – Sensitivity Rate; SPR – in Table 2. EFR for the training set of each model ranged from 92.2%
Specificity Rate; EFR – Efficiency Rate.
to 100%, and for test sets ranged from 75.3% to 100%, with the two
worst results observed for the starch adulteration model.
The CON was estimated using three replicates of three different
Water adulteration model presented the highest FPR (4.0%) and levels (0.5%, 5.0% and 10% w/v) analyzed in repeatability conditions
the second highest FNR (3.2%) in the training set, and also the high- for each adulterant. The ACO was estimated using the same proto-
est FPR in the test set (11.5%). The greatest number of misclassifi- col employed for CON estimation, repeated in three different days
cations in this model may be due to the close similarity between by three different analysts. The water model presented the small-
the water–adulterated samples and the blank samples or samples est ACO and CON values (78% and 92%, respectively), with all other
of other adulterants in lower concentrations (Fig. 4(a)). Starch four models presenting 100% of both ACO and CON. These results
adulteration presented a low number of FPR and FNR (3.0% and suggest that the detection of adulteration with water was not so
4.8%, respectively) in the training set. However, this model pre- precise, maybe because of the large amount of water already pre-
sented relative high FPR and FNR for the test set (8.0% and 16.7%, sent in non-adulterated milk and the natural variation of its con-
respectively), with misclassification of samples containing water tent between 85% and 90% w/w.
B.G. Botelho et al. / Food Chemistry 181 (2015) 31–37 37

4. Conclusion cnpgl.embrapa.br/nova/informacoes/estatisticas/producao/producao.php>.
Accessed in June 2014.
Fox, J. P., & McSweeney, P. L. H. (1998). Dairy Chemistry and Biochemistry. Berlin:
A simple multivariate screening method for simultaneous Springer Science & Business Media (pp. 478).
detection of five common adulterants in milk, based on ATR mid Gondim, C. S., Junqueira, R. G., & Souza, S. V. C. (2011). Trends in implementing the
validation of qualitative methods of analysis. Revista do Instituto Adolfo Lutz, 70,
infrared spectroscopy, was developed and validated. The method
433–447.
was able to detect the specific presence of water, starch, sodium Hoorfar, J. (2012). Case Studies in Food Safety and Authenticity: Lessons From Real-Life
citrate, formaldehyde and sucrose in raw cow milk samples in con- Situations (1st ed.). Cambridge: Woodhead Publishing Limited, pp. 408.
IAL (2005). Métodos Físicos-Químicos para Análise de Alimentos (4th ed.). São Paulo:
centrations as low as 0.5% w/v. The analytical range chosen for all
Instituto Adolfo Lutz (IAL) (pp. 1020).
the adulterants, between 0.5% and 10.0% w/v, was consistent with Kasemsumran, S., Thanapase, W., & Kiatsoonthon, A. (2007). Feasibility of near-
the forensic interest, considering the reported cases of milk adul- infrared spectroscopy to detect and to quantify adulterants in cow milk.
teration. All the models, except for the starch, provided results in Analytical Sciences, 23, 907–910.
Kizil, R., Irudayaraj, J., & Seethraman, K. (2002). Characterization of irradiated
accordance with international validation guidelines (EC, 2002; starches by using FT-Raman and FTIR spectroscopy. Journal of Agricultural and
NATA, 2013). The proposed method requires a small amount of Food Chemistry, 50, 3912–3918.
sample (about 30 lL), is very rapid (up to 1 min for spectral acqui- Langton, S. D., Chevennement, R., Nagelkerke, N., & Lombard, B. (2002). Analysing
collaborative trials for qualitative microbiological methods: Accordance and
sition) and non-destructive, not demanding any sample prepara- concordance. International Journal of Food Microbiology, 79, 175–181.
tion nor using reagents, and all the adulterants can be evaluated Magalhães, C. G., de Paiva, C. R., Botelho, B. G., de Oliveira, A. M. G., de Souza, L. F.,
simultaneously using the same spectra. The analytical validation Nonaka, C. V., et al. (2012). In-house validation of PremiÒTest, a microbiological
screening test with solvent extraction, for the detection of antimicrobial
was carried out searching to extend the estimate of specific figures residues in poultry muscles. Food Additives & Contaminants: Part A, 29, 535–540.
of merit from univariate to multivariate qualitative methods. This MAPA (2006). Instrução Normativa no 22 de 19 de abril de 2006. Secretaria de
harmonization is important once it aims the acceptance of multi- Defesa Agropecuária. Ministério da Agricultura, Pecuária e Abastecimento
(MAPA). Available. In: <www.agricultura.gov.br>. Accessed in June 2014.
variate screening methods by the official guidelines. The simplicity
NATA (2013). Technical Note 17. Guidelines for the validation and verification of
of the proposed method makes it suitable for quality control in dif- quantitative and qualitative test methods. Canberra: National Association of
ferent steps of the chain of production, such as the receipt of raw Testing Authorities (pp. 32).
Nicolaou, N., Xu, Y., & Goodcare, R. (2010). Fourier transform infrared spectroscopy
milk in a dairy product factory or the detection of adulterants in
and multivariate analysis for the detection and quantification of different milk
commercialized milk. species. Journal of Dairy Science, 93, 5651–5660.
Pavia, D., Lampman, G., Kriz, G., & Vyvyan, J. (2008). Introduction to Spectroscopy (4th
ed.). Stanford: Cengage Learning (pp. 752).
Acknowledgements Pikeemaat, M. (2009). Microbial screening methods for detection of antibiotic
residues in slaughter animals. Analytical and Bioanalytical Chemistry, 395,
B. G. B. thanks CAPES and CNPq for fellowships and Denise 893–905.
Pulido, A., Ruisanchez, I., Boqué, R., & Rius, X. F. (2003). Uncertainty of results in
Aquino from LANAGRO (Pedro Leopoldo, MG) for providing the routine qualitative analysis. Trends in Analytical Chemistry, 22, 647–654.
milk samples. Santos, P. M., Pereira-Filho, E. R., & Rodriguez-Saona, L. E. (2013a). Application of
hand-held and portable infrared spectrometers in bovine milk analysis. Journal
of Agricultural and Food Chemistry, 61, 1205–1211.
Appendix A. Supplementary data Santos, P. M., Pereira-Filho, E. R., & Rodriguez-Saona, L. E. (2013b). Rapid detection
and quantification of milk adulteration using infrared microspectroscopy and
chemometrics analysis. Food Chemistry, 138, 19–24.
Supplementary data associated with this article can be found, in Santos, P. M., Wentzell, P. D., & Pereira-Filho, E. R. (2012). Scanner digital images
the online version, at http://dx.doi.org/10.1016/j.foodchem.2015. combined with color parameters: A case study to detect adulterations in liquid
02.077. cow’s milk. Food Analytical Methods, 5, 89–95.
Souza, S. S., Cruz, A. G., Walter, E. H. M., Faria, J. A. F., Celeghini, R. M. S., Ferreira, M.
M. C., et al. (2011). Monitoring the authenticity of Brazilian UHT milk: A
References chemometric approach. Food Chemistry, 124, 692–695.
Tewari, J. C., & Malik, K. (2007). In situ laboratory analysis of sucrose in sugarcane
bagasse using attenuated total reflectance spectroscopy and chemometrics.
Botelho, B. G., Mendes, B. A. P., & Sena, M. S. (2013). Development and analytical International Journal of Food Science and Technology, 42, 200–207.
validation of robust near-infrared multivariate calibration models for the Trullols, E., Ruisanchez, I., & Rius, X. F. (2004). Validation of qualitative analytical
quality inspection control of mozzarella cheese. Food Analytical Methods, 6, methods. Trends in Analytical Chemistry, 23, 137–145.
881–891. UOL (2014). Operação desmonta esquema de empresas que adulteravam leite no
Brown, C. D., & Green, R. L. (2009). Critical factors limiting the interpretation of Sul. Available in <http://noticias.uol.com.br/cotidiano/ultimas-noticias/2013/
regression vectors in multivariate calibration. Trends in Analytical Chemistry, 28, 05/08/operacao-desmonta-esquema-de-empresas-que-fraudavam-leite-no-sul.
506–514. htm>. Accessed in June 2014.
Chong, I.-G., & Jun, C.-H. (2005). Performance of some variable selection methods Valcárcel, M., & Cárdenas, S. (2005). Modern qualitative analysis. Trends in Analytical
when multicollinearity is present. Chemometrics and Intelligent Laboratory Chemistry, 24, 467.
Systems, 78, 103–112. Wise, B. W., Gallagher, N. B., Bro, R., Shaver, J. M., Windig, W., & Koch, R. S. (2006).
EC (2002). European Community Directive 657. Comission Decision of 12 August PLS-Toolbox 4.0 for use with MatlabTM (Manual). Manson: EigenVector Research
2002 implementing Council Directive 96/23/EC concerning the performance of Inc.
analytical methods and the interpretation of results, Official Journal, L221 (pp. Zhang, L.-G., Zhang, X., Ni, L.-J., Xue, Z.-B., Gu, X., & Huang, S.-X. (2014). Rapid
8–36), Brussels. identification of adulterated cow milk by non-linear pattern recognition
EMBRAPA (2014). Estatísticas de produção de leite (Statistics of milk production). methods based on near infrared spectroscopy. Food Chemistry, 145, 342–348.
Brasília: Brazilian Agriculture Research Corporation. Available in <http://www.

You might also like