od

of L

d, Ch

k inme8 nanddinnalremSNV can improve classication accuracy and reduce the complexity of PLSDA.

nd aducomposcessedre, ver

medium rich in social meaning because of its association with cultural methods (Aida, Che Man, Raha, & Son, 2007; Aida, Che Man, Wong,

Meat Science 92 (2012) 506510

Contents lists available at SciVerse ScienceDirect

Meat Sc

l shabits and rituals, both religious and secular. Religious food prescrip-tions are far easier to adopt than to discard because once a ban isadopted it tends to be reinforced by strong feelings of disgust forexample the strong aversion of Muslims for pork in general. Meatspecies identication and Halal authentication are a major concernin Asia, France, Russia, Sweden, Germany, Switzerland, Greece,Spain, Italy, United Kingdom, South and North America and mostother countries (Murugaiah et al., 2009).

Rapid and reliable methods for detection of Halal meat adultera-tion are indispensable for implementation of food labeling regula-

Raha, & Son, 2005).Ham sausage accounts for nearly one third of the total meat prod-

ucts in China and the yearly output exceeds 10,000,000 tons. Chineseham sausage is a complex mixture consisting mainly of meat andstarch and with low concentrations of water, vegetable oil, salt,monosodium glutamate and other food additives. The main meat con-tents of Halal ham sausage are beef, chicken or sh. Unfortunately,some food manufacturers choose to use pork as a substitute ingredi-ent for Halal meats because it is cheaper and easily available, whichwould trigger serious dispute on national relationships. Mincedtions and product quality control. Methodsto be specic, sensitive, rapid, economic andof different morphological characteristics (M

Corresponding authors.E-mail addresses: zhye@cjlu.edu.cn (Z.H. Ye), yxp@c

1 These authors contributed equally to the work.

0309-1740/$ see front matter 2012 Elsevier Ltd. Alldoi:10.1016/j.meatsci.2012.05.019ication that the compo-ble to special consumers. Meat in particular is a

(Marikkar, Ghazali, Che Man, Peiris, & Lai, 2005; Rashood, Shaaban,Moety, & Rauf, 1995; Saeed, Ali, Rahman, & Sawaya, 1989), electronicnose (Che Man, Gan, NorAini, Nazimah, & Tan, 2005), and DNA-basednents are authentic and from sources acceptamay be required (Lockley & Bardsley, 2000)1. Introduction

The concern of food authenticity aincreased awareness regarding theThe identity of the ingredients in prois not always readily apparent. Therefoperformance. For the best models, the sensitivity and specicity was 0.913 and 0.929 for PLSDA with SNVspectra, 0.957 and 0.929 for LS-SVM with second derivative spectra, respectively.

2012 Elsevier Ltd. All rights reserved.

lteration has resulted inition of food products.or composite mixtures

Velzquez, & Osorio-Revilla, 2010). Various techniques have beenproposed for the analysis of pork or lard, including differential scan-ning calorimetry (Coni, Pasquale, Cappolelli, & Bocca, 1994;Kowalski, 1989), gas chromatography (Farag, Abo-raya, Ahmed,Hewedi, & Khalifa, 1983), high pressure liquid chromatographyLS-SVMPLSDAPossibly due to the loss of detailed high-frequency spectral information, smoothing degrades the modelHalalFTIR spectroscopy

tigated. The results indicateination. Taking derivatives,Rapid discrimination of pork in Halal and ntransform infrared (FTIR) spectroscopy an

L. Xu a,1, C.B. Cai b, H.F. Cui a,1, Z.H. Ye a,, X.P. Yu a,a Zhejiang Provincial Key Laboratory of Biometrology and Inspection & Quarantine, CollegeHangzhou 310018, Chinab Department of Chemistry and Life Science, Chuxiong Normal University, Luchengnan Roa

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

Article history:Received 19 November 2011Received in revised form 26 February 2012Accepted 18 May 2012

Keywords:Chinese Ham sausage

Rapid discrimination of porform infrared (FTIR) spectro4000 cm1 of 73 Halal and 7nely grinding of samplespreprocessing methods incluleast squares discriminant a

j ourna l homepage: www.efor these purposes needable to analyze sampleseza-Mrquez, Gallardo-

jlu.edu.cn (X.P. Yu).

rights reserved.n-Halal Chinese ham sausages by Fourierchemometrics

ife Sciences, China Jiliang University, Xueyuan Street, Xiasha Higher Education District,

uxiong 675000, China

Halal and non-Halal Chinese ham sausages was developed by Fourier trans-try combined with chemometrics. Transmittance spectra ranging from 400 toon-Halal Chinese ham sausages were measured. Sample preparation involvedformation of KBr disks (under 10 MPa for 5 min). The inuence of datag smoothing, taking derivatives and standard normal variate (SNV) on partialysis (PLSDA) and least squares support vector machine (LS-SVM) was inves-oval of spectral background and baseline plays an important role in discrim-

ience

ev ie r .com/ locate /meatsc imeat production removes the morphological characteristics of mus-cle, making it difcult to identify one type of muscle from another.For this reason, meat substitution with unspecied species, usuallyof lower quality, is the most common form of economic adulterationin the minced meat industry (Hargin, 1996). For routine analysis ofham sausage, some of the above methods are too laborious, time-consuming and expensive, therefore, rapid and economical yet reli-able analysis methods are highly demanded.

As a promising alternative approach to the traditional methods ofchemical and sensory analysis, the combination of spectrometry andchemometricmethods has been successfully used inHalal food analysis.Previous research has shown the potential of FTIR spectroscopy foranalysis of lard in cake formulation (Syahariza, Che Man, Selamat, &Bakar, 2005) and chocolate products (Che Man, Syahariza, Mirghani,Jinap, & Bakar, 2005). FTIR spectroscopy was also used to characterizelard and other edible oils (Guillen & Cabo, 1997). Other successful appli-cations include detection of lard in the mixture with other animal fats(Che Man & Mirghani, 2001; Jaswir, Mirghani, Hassan, & Mohd Said,2003; Rohman & Che Man, 2010) and analysis of pork adulteration inbeef meatball (Rohman, Sismindari, Erwanto, & Che Man, 2011). Theobjective of this study is to develop an accurate and reliable methodto discriminateHalal and non-Halal Chinese ham sausages by FTIR spec-trometry combined with multivariate discriminant analysis. Given theperformances of different classication models and preprocessingmethods are similar or have no signicant differences, models with

507L. Xu et al. / Meat Science 92 (2012) 506510least complexity and preprocessing were sought to ensure the general-ization of models.

2. Materials and methods

2.1. Collection of samples

Representative Halal and non-Halal ham sausage samples by mainproducers from China were collected. 73 Halal and 78 non-Halal hamsausage samples were analyzed. The ham sausage samples wereobtained from the domestic markets and the identities of sampleswere ascertained by the quality branch of manufacturers. All of thesamples retained integral packaging and the original labels indicatingdetailed sample information. For all the samples, according to nation-al and professional standards, the content of starch was less than 10%.The contents of protein and fat were no less than 12% and no morethan 10%, respectively. The detailed information concerning samplesis shown in Table 1. All of the samples were stored at about 0 Cwith integral packaging before spectrometry analysis.

2.2. Sample preparation and FTIR spectroscopy analysis

Sample preparation involved nely grinding of ham sausage sam-ples followed by preparation of KBr pellets. Sampling was performedon different parts of each ham sausage to consider the potential het-erogeneity of materials. Then the granular samples were manuallyground into ne particles with KBr using an agate pestle and mortar.25 mg (1:40 w/w) of each powder sample was mixed with 975 mg(39:40 w/w) of KBr (Xi'an Shiji, Xi'an, China). KBr pellets were pre-pared by exerting a pressure of 10 MPa for approximately 5 min in apellet press (Tuopu Instrument., Tianjin, China). To examine whether

Table 1Detailed information of the ham sausage samples analyzed.

Brands Batch size Meat content Typesa

Jinluo 12 Pork NJinluo 11 Chicken HJinluo 10 Beef HJinluo 11 Chicken and Pork NShineway 10 Pork NShineway 12 Chicken and Pork NShineway 15 Beef HShineway 14 Chicken HYurun 12 Beef HYurun 11 Chicken HYurun 8 Pork NTRS 11 Pork NMeihao 7 Pork NMeihao 7 Chicken and Pork N

a N=non-Halal and H=Halal.the variation in pellet thickness cause signicant interference in themeasured spectra, different pellets were prepared from the samesample and their FTIR spectra were compared (Garip, Gozen, &Severcan, 2009). The correlation coefcient between each spectrumwith the average spectrum was higher than 0.99, indicating the mea-sured FTIR spectra were nearly identical to their average spectrumused for analysis.

FTIR spectra were collected using an Avatar-360 FTIR spectrome-ter (Thermo Scientic, Waltham, MA) working in the wavelengthrange of 4004000 cm1. For each pellet, 128 scans were performedwith a resolution of 4 cm1 at room temperature using OMNIC soft-ware. An increase in scanning time did not signicantly improve thesignal. The average of the 128 scans was used as a raw spectrumfor further data analysis. The scanning interval was 1.929 cm1.Therefore, each spectrum contained 1868 individual points forchemometric analysis.

2.3. Preprocessing and data splitting

All the preprocessing and further data analysis were performed onMatlab 7.0.1 (Mathworks, Sherborn, MA). Considering the lack of suf-cient prior information concerning the measured spectra, differentoptions were investigated to optimize data preprocessing. Smoothingcan remove part of the random noise present in the signal and en-hance the signal-to-noise ratio (SNR). The algorithm of polynomialtting (Savitzky & Golay, 1964) was adopted for this purpose becauseof its popularity and simplicity. Taking derivatives can enhance spec-tral differences and remove baseline and background, so rst and sec-ond derivatives were applied. Because direct differencing tends todecrease the SNR by enhancing noise, the derivative spectra werealso computed by polynomial tting algorithms. Standard normalvariate (SNV) (Barnes, Dhanoa, & Lister, 1989) transforms each mea-sured spectrum into a signal with zero mean and unit variance. It wasoriginally proposed to reduce scattering effects in the spectra but wasalso proved to be effective in correcting the interference caused byvariations in pellet thickness or optical path. Although the inuenceof the thickness of pellets was found to be insignicant in this work,SNV was performed to reduce the possible variations caused by scat-tering effects or uneven mixing of KBr and sample powders.

The DUPLEX algorithm (Snee, 1977) was used to split the mea-sured spectra data into a representative training set and test set.DUPLEX selects the two samples with largest distance and putsthem in the training set, then selects the two samples with largestdistance among the left samples and puts them in the training set,and so on. By alternatively selecting the spectral data for the calibra-tion set and test set, DUPLEX gives data in the test set with a distribu-tion almost equal to that of the training set. Because the distributionsof Halal and non-Halal samples were different, DUPLEX method wasperformed separately for the Halal and non-Halal ham sausages.

2.4. Multivariate statistical analysis and method validation

Partial least-squares discriminant analysis (PLSDA) (Barker &Rayens, 2003) is a classication method based on partial least squares(PLS) regression. As the cornerstone of chemometrics, PLS has beensuccessfully used to solve various regression problems. For two-classproblems, suppose an np matrix X including p wavelength variablesfor n training objects, the response vector y (n1) is constructed withthe category variable of each object in X, for instance, +1 and 1were used to denote Halal and non-Halal samples, respectively. There-fore, the cut-off value of predicted response values was set to be 0,e.g., an object with a predicted response value above/under 0 wouldbe classied into Halal/non-Halal class.

Support vector machine (SVM) considers the trade-off betweenthe capacity and generalization performance of a learned model by

regression coefcients regularization. Based on the structural risk

minimization principle, SVM has been proved to be an effective androbust method for both classication and regression. Least squaresSVM (LS-SVM) (Suykens & Vandewalle, 1999) was suggested as asimplied version of SVM. With equality constraints in the formula-tion, LS-SVM obtains the solution by solving a set of linear equations,instead of quadratic programming for traditional SVM algorithms. LS-SVM can model nonlinear relationship by using a kernel matrix trans-formation of the original X.

In this paper, linear PLSDA and nonlinear LS-SVMwere applied to thediscrimination of non-Halal and Halal ham sausages based on the rawand preprocessed spectra. For LS-SVM, the mostly frequently usedGaussian radical basis functionwas applied for nonlinear transformation.To reduce the risk of overtting, Monte Carlo cross validation (MCCV)(Xu& Liang, 2001)was used to evaluate the number of PLSDA latent var-iables and optimize the parameters of LS-SVM. The mean err rate ofMCCV (MERMCCV) was used as the validity criteria for the classication.

Peaks in 30002800 cm1 also involve the contributions of C-Hstretching vibrations. Other obvious bands include 1720 cm1 (C Ostretching) 1530 cm1 (asymmetric stretching vibration of COOH inamino acids) and 1070 cm1 (stretching vibration of COC). Accurateassignments of most peaks were difcult due to low resolution and sig-nicant baselines, therefore, pattern recognition methods are necessaryto extract the useful information from spectral data for discrimination.

Fig. 2 demonstrates the PCA plot of raw FTIR spectra of Halal andnon-Halal ham sausages. The rst two principal components (PCs)explain 83.2% of the total variances. It can be seen that the projectionsof the samples in both classes onto the 2-PC subspace are very dis-perse. This can be attributed to the fact that both Halal and non-Halal groups were known to contain samples composed of differentmeats. Obviously, the data have a complex structure and two PCsare insufcient to discriminate them from each other.

Fig. 2. PCA plot of raw FTIR spectra of Halal and non-Halal ham sausages.

Fig. 3. Average spectra of Halal and non-Halal ham sausages preprocessed by smoothing

508 L. Xu et al. / Meat Science 92 (2012) 5065102.5. Evaluation of model performance

Sensitivity and specicity (Forina, Armanino, Leardi, & Drava, 1991)were used to evaluate the performance of different classicationmodelsand data preprocessing methods. Denote Halal as positive and non-Halal as negative, sensitivity (Sens) and specicity (Spec) were com-puted as:

Sens TPTP FN

Spec TNTN FP

2

where TP, FN, TN, and FP denote the numbers of true positives, falsenegatives, true negatives, and false positives, respectively.

Moreover, the total accuracy of classication was also used:

Accu TN TPTN TP FN FP 3

3. Results and discussions

3.1. FTIR spectral analysis

Some of the raw FTIR spectra of Halal and non-Halal ham sausageswere shown in Fig. 1. Seen from Fig. 1, the spectra of Halal and non-Halal samples have very similar absorbance bands in the range of4004000 cm1 (Ripoche & Guillard, 2001). The wide bands in36001700 cm1 can be attributed to the overlapping of thestretching of various OH groups (36002000 cm1) and NHgroups (34001700 cm1), where the peak resolution is very low.Fig. 1. Raw FTIR spectra of Halal (solid line) and non-Halal (dotted line) ham sausages. (1), SNV (2), rst-order derivative (3) and second derivative (4).

Fig. 3 demonstrates the average preprocessed spectra of Halal andnon-Halal ham sausage samples. By comparison of the raw spectrawith smoothed spectra, although smoothed spectra can slightly im-prove the SNR, it might lose some useful high-frequency informationin the raw data. Compared with rst-order derivative spectra, thesecond-order derivative spectra can remove most of the baselinesand enhance some detailed information and peak resolution. SNVspectra can remove some spectral variations while enhancing others.The effects of data preprocessing should be evaluated by modelperformance.

that preprocessing generally improved the classication performancein terms of sensitivity and specicity. However, PLSDA and LS-SVMbased on smoothed spectra had inferior performance, which mightbe attributed to the possible loss of detailed frequency information(Kokalj, Rihtari, & Kreft, 2011). Second derivative and SNV signi-cantly sharpened the classication models by reducing the baselineand backgrounds. The model complexity of PLSDA based on rst de-rivative, second derivative and SNV was reduced compared with themodel based on smoothed and raw spectra. The results by rst deriv-ative spectra were not satisfying and this might be partially attributedto the baseline remained in rst derivative spectra as seen in Fig. 3.For the best models, the sensitivity and specicity was 0.913 and0.929 for PLSDA with SNV spectra and 0.957 and 0.929 for LS-SVMwith second derivative spectra, respectively. The best prediction re-sults were also demonstrated in Fig. 4. The comparison of differentpreprocessing methods demonstrated that the spectral variationscaused by scattering effects and baseline shifts played a more impor-tant role than SNR. Since LS-SVM involves nonlinear transformationof the raw variables, PLSDA with SNV preprocessing should be

Fig. 4. The predicted response values by the best linear PLSDA model and nonlinear LS-SVM models. Samples 123 are Halal (positive) samples and 2451 are non-Halal

Table 3Results of LS-SVM models with different preprocessing methods.

Preprocessing Sensitivity Specicity 2, MERMCCV Accuracy

Raw data 0.826 (19/23)a 0.893 (25/28)b 0.80, 8 0.170 0.824Smoothing 0.783 (18/23) 0.893 (25/28) 0.65, 11 0.191 0.7841st derivative 0.870 (20/23) 0.786 (22/28) 0.85, 14 0.211 0.7842nd derivative 0.957 (22/23) 0.929 (26/28) 0.45, 7 0.091 0.922SNV 0.913 (21/23) 0.929 (26/28) 0.40, 11 0.094 0.902

a True positive/Total positive.b True negative/Total negative.

509L. Xu et al. / Meat Science 92 (2012) 5065103.2. Optimization of model parameters

The DUPLEX method was performed on the Halal and non-Halalsamples separately. Each data set was divided into a training set of50 samples and a test set containing the remainder samples. Thetwo training sets were then combined for developing classicationmodels. Therefore, the nal training set had 100 objects (50 Halalplus 50 non-Halal) and 51 test samples (23 Halal plus 28 non-Halal).

LS-SVM and PLSDA models were developed based on raw andpreprocessed spectra. For LS-SVM, two parameters, and need tobe optimized. The kernel width parameter, , is related to the con-dence in the data; the magnitude of also inuences the non-linearnature of the regression. As decreases, the kernel becomesnarrower, forcing the model toward a more complex (nonlinear) so-lution. The regularization parameter controls the tradeoff betweenmaximizing the margin and minimizing the training error, a toosmall value of will lead to an under-tted model; if is too large,the model tends to overt the training data and the model willhave poor prediction performance. Therefore, should be optimizedtogether with the kernel width parameter . To optimize the twoparameters in LS-SVM and the model complexity (number of latentvariables) of PLSDA, MERMCCV was computed with different combi-nations of parameters. Compared with the traditional leave-one-outcross validation (LOOCV), MCCV can effectively reduce the risk ofovertting by multiple resampling of the training set and a higherrate of leave-out samples. In this paper the resampling time is set tobe 200 and the number of left-out samples was 10 (5 Halal and 5non-Halal) for each resampling. The MERMCCV was computed as:

ERMCCV XN

i1

ENiN L 4

where N is the number of resampling time, L is the number of leave-out samples, ENi is the misclassied samples for the ith resampling.The number of PLSDA latent variables and LS-SVM parameter combi-nation ( and ) were optimized to obtain the lowest MERMCCVvalues.

3.3. Comparison of model performances

With different preprocessing methods, the prediction results andoptimized parameters were listed in Tables 2 and 3. It can be seen

Table 2Results of PLSDA models with different preprocessing methods.

Preprocessing Sensitivity Specicity Lva MERMCCV Accuracy

Raw data 0.783 (18/23)b 0.857 (24/28)c 8 0.170 0.824Smoothing 0.739 (17/23) 0.821 (23/28) 7 0.191 0.7841st derivative 0.826 (19/23) 0.750 (21/28) 6 0.211 0.7842nd derivative 0.913 (21/23) 0.893 (25/28) 6 0.102 0.922SNV 0.913 (21/23) 0.929 (26/28) 7 0.091 0.902

a The number of PLS latent variables.b True positive/Total positive.c True negative/Total negative. (negative) samples.

recommended because it is linear and simpler and expected to have amore reliable generalization performance.

4. Conclusion

Rapid discrimination of Halal/non-Halal Chinese ham sausageswas developed by FTIR spectroscopy and chemometric data analysis.PLSDA with SNV spectra (sensitivity 0.913 and specicity 0.929)and LS-SVM with second derivative spectra (sensitivity 0.957 andspecicity 0.929) achieved best classication performance in terms

CheMan, Y. B., Syahariza, Z. A., Mirghani, M. E. S., Jinap, S., & Bakar, J. (2005). Analysis ofpotential lard adulteration in chocolate and chocolate products using Fouriertransform infrared spectroscopy. Food Chemistry, 90, 815819.

Coni, E., Pasquale, M. D., Cappolelli, P., & Bocca, A. (1994). Detection of animal fats inbutter by SC: A pilot study. Journal of the American Oil Chemists' Society, 71,807810.

Farag, R. S., Abo-raya, S. H., Ahmed, F. A., Hewedi, F. M., & Khalifa, H. H. (1983). Frac-tional crystallization and gas chromatographic analysis of fatty acids as a meansof detecting butterfat adulteration. Journal of the American Oil Chemists' Society,60, 16651669.

Forina, M., Armanino, C., Leardi, R., & Drava, G. (1991). A class modelling techniquebased on potential functions. Journal of Chemometrics, 5, 435453.

Garip, S., Gozen, A. C., & Severcan, F. (2009). Use of Fourier transform infrared spectros-copy for rapid comparative analysis of Bacillus and Micrococcus isolates. FoodChemistry, 113, 13011307.

Guillen, M. D., & Cabo, N. (1997). Characterization of edible oils and lard by Fourier

510 L. Xu et al. / Meat Science 92 (2012) 506510removal of spectral background and baseline plays a more importantrole than a higher signal-to-noise ratio (SNR). Taking derivatives, SNVcan not only improve classication accuracy but also reduce the com-plexity of PLSDA. Possibly due to the loss of detailed high-frequencyspectral information, spectra smoothing degrades the model perfor-mance. Although we can hardly perform an exhaustive sampling ofall types of ham sausages, this study built a reliable model for Halal/non-Halal discrimination of some mainstream and representativesamples in China. This paper demonstrates FTIR combined withchemometrics provides a useful tool for Halal authentication of simi-lar ham sausages. However, if one wants to analyze ham sausageswith very different compositions and production procedure, it is rec-ommended specic models be built with the samples of interest, be-cause a model with very diverse training samples would have muchmore complexity and degraded generalization performance.

Acknowledgements

This work was nancially supported by the National PublicWelfare Industry Projects of China (no. 201210010 and 201210092)and Hangzhou Programs for Agricultural Science and TechnologyDevelopment (no. 20101032B28). Chen-Bo Cai is grateful to thenancial aid of the Applied and Basic Research Project of YunnanProvincial Science and Technology Department (no. 2010CD087).

References

Aida, A. A., Che Man, Y. B., Raha, A. R., & Son, R. (2007). Detection of pig derivatives infood products for Halal authentication by polymerase chain reactionrestrictionfragment length polymorphism. Journal of the Science of Food and Agriculture, 87,569572.

Aida, A. A., Che Man, Y. B., Wong, C. M. V. L., Raha, A. R., & Son, R. (2005). Analysis of rawmeats and fats of pigs using polymerase chain reaction for Halal authentication.Meat Science, 69, 4752.

Barker, M., & Rayens, W. (2003). Partial least squares for discrimination. Journal ofChemometrics, 17, 166173.

Barnes, R. J., Dhanoa, M. S., & Lister, S. J. (1989). Standard normal variate transforma-tion and detrending of near infrared diffuse reectance. Applied Spectroscopy, 43,772777.

Che Man, Y. B., Gan, H. L., NorAini, I., Nazimah, S. A. H., & Tan, C. P. (2005). Detection oflard adulteration in RBD palm olein using an electronic nose. Food Chemistry, 90,829835.

Che Man, Y. B., & Mirghani, M. E. S. (2001). Detection of lard mixed with body fats ofchicken, lamb, and cow by Fourier transform infrared spectroscopy. Journal of theAmerican Oil Chemists' Society, 78, 753761.transform infrared spectroscopy. Relationships between composition and frequen-cy of concrete bands in the ngerprint region. Journal of the American Oil Chemists'Society, 74, 12811286.

Hargin, K. D. (1996). Authenticity issues in meat and meat products. Meat Science, 43,277289.

Jaswir, I., Mirghani, M. E. S., Hassan, T. H., & Mohd Said, M. Z. (2003). Determination oflard in mixtures of body fats of mutton and cow by Fourier transform-Infra red(FTIR) spectroscopy. Journal of Oleo Science, 52, 633638.

Kokalj, M., Rihtari, M., & Kreft, S. (2011). Commonly applied smoothing of IR spectrashowed unappropriate for the identication of plant leaf samples. Chemometricsand Intelligent Laboratory Systems, 108, 154161.

Kowalski, B. (1989). Sub-ambient differential scanning calorimetry of lard and lardcontaminated by tallow. International Journal of Food Science and Technology, 24,415420.

Lockley, A. K., & Bardsley, R. G. (2000). DNA-based methods for food authentication.Trends in Food Science and Technology, 11, 6777.

Marikkar, J. M. N., Ghazali, H. M., Che Man, Y. B., Peiris, T. S. G., & Lai, O. M. (2005). Dis-tinguishing lard from other animal fats in admixtures of some vegetable oils usingliquid chromatographic data coupled with multivariate data analysis. Food Chemis-try, 91, 514.

Meza-Mrquez, O. G., Gallardo-Velzquez, T., & Osorio-Revilla, G. (2010). Applicationof mid-infrared spectroscopy with multivariate analysis and soft independentmodeling of class analogies (SIMCA) for the detection of adulterants in mincedbeef. Meat Science, 86, 511519.

Murugaiah, C., Noor, Z. M., Mastakim, M., Bilung, L. M., Selamat, J., & Radu, S. (2009).Meat species identication and Halal authentication analysis using mitochondrialDNA. Meat Science, 83, 5761.

Rashood, K. A., Shaaban, R. R. A., Moety, E. M. A., & Rauf, A. (1995). Triacylglycerolsprolingby high performance liquid chromatography: A tool for detection of pork fat inprocessed foods. Journal of Liquid Chromatography, 18, 2631.

Ripoche, A., & Guillard, A. S. (2001). Determination of fatty acid composition of pork fatby Fourier transform infrared spectroscopy. Meat Science, 58, 299304.

Rohman, A., & Che Man, Y. B. (2010). FTIR spectroscopy combined with chemometricsfor analysis of lard in the mixtures with body fats of lamb, cow, and chicken. Inter-national Food Research Journal, 17, 519527.

Rohman, A., Sismindari, Erwanto, Y., & Che Man, Y. B. (2011). Analysis of pork adulter-ation in beef meatball using Fourier transform infrared (FTIR) spectroscopy. MeatScience, 88, 9195.

Saeed, T., Ali, S. G., Rahman, H. A., & Sawaya, W. N. (1989). Detection of pork and lard asadulterants in processed meat: Liquid chromatographic analysis of derivatized tri-glycerides. Journal of the Association of Ofcial Analytical Chemists, 72, 921925.

Savitzky, A., & Golay, M. J. E. (1964). Smoothing and differentiation of data by simpli-ed least-squares procedures. Analytical Chemistry, 36, 16271639.

Snee, R. D. (1977). Validation of regression models: Methods and examples.Technometrics, 19, 415428.

Suykens, J. A. K., & Vandewalle, J. (1999). Least squares support vector machine classi-ers. Neural Processing Letters, 9, 293300.

Syahariza, Z. A., Che Man, Y. B., Selamat, J., & Bakar, J. (2005). Detection of lard adulter-ation in cake formulation by Fourier transform infrared (FTIR) spectroscopy. FoodChemistry, 92, 365371.

Xu, Q. -S., & Liang, Y. -Z. (2001). Monte Carlo cross validation. Chemometrics and IntelligentLaboratory Systems, 56, 111.of prediction sensitivity and specicity. The analysis results indicate

Rapid discrimination of pork in Halal and non-Halal Chinese ham sausages by Fourier transform infrared (FTIR) spectroscopy and chemometrics1. Introduction2. Materials and methods2.1. Collection of samples2.2. Sample preparation and FTIR spectroscopy analysis2.3. Preprocessing and data splitting2.4. Multivariate statistical analysis and method validation2.5. Evaluation of model performance

3. Results and discussions3.1. FTIR spectral analysis3.2. Optimization of model parameters3.3. Comparison of model performances

4. ConclusionAcknowledgementsReferences