Analytica Chimica Acta 538 (2005) 181–193

Determination of the energetic value of fruit and milk-based beveragesthrough partial-least-squares attenuated total reflectance-Fourier

transform infrared spectrometry

Javier Morosa, Fernando A. Inonb, Salvador Garriguesa, Miguel de la Guardiaa, ∗a Departamento de Qu´ımica Anal´ıtica, Universidad de Valencia, Edificio Jer`onim Munoz C/ Dr. Moliner, 50, 46100 Burjassot (Valencia), Spain

b Laboratorio de An´alisis de Trazas, Departamento de Qu´ımica Inorganica, Anal´ıtica y Quımica Fısica, Universidad de Buenos Aires,Pabellon 2, Ciudad Universitaria, 1428 Buenos Aires, Argentina

Received 8 November 2004; received in revised form 25 January 2005; accepted 1 February 2005Available online 2 March 2005

Abstract

The estimation of important nutritional parameters, such as carbohydrates content and energetic value (calories) in commercially availablef IR) using ap et, coveringf uilding theP

samples, thea mL,r 5 g/100 mL.I ondedt kJ/100 mLa

d providedfi©

K

1

pac

tbo

, itshigh-

onlyutyper-rterialever-ntrol

withdges int it.

0d

ruit juice and flavour milk shakes has been made by attenuated total reflectance-Fourier transform infrared spectroscopy (ATR-FTartial-least-square (PLS) calibration approach. A highly heterogeneous population of 65 samples obtained from the Spanish mark

ruit juices, flavour milk shakes and milk-added fruit juices was used. The spectral range and the size of the calibration set for bLS model have been evaluated.Considering a calibration set comprised of 27 samples, selected via hierarchical cluster analysis, and a validation data set of 38

bsolute mean difference (dx–y) and standard deviation of mean differences (sx–y) of the total carbohydrate content were 0.06 and 0.66 g/100espectively. The reproducibility of this determination established as the mean standard deviation of each triplicate analysis was 0.0n the case of energetic value, thedx–y andsx–y were 2.8 and 18 kJ/100 mL, respectively. The reproducibility of this determination correspo a standard deviation of 2.4 kJ/100 mL, for three replicate analyses. The root-mean-square error of prediction (RMSEP) was 18.4nd 0.72 g/100 mL for energetic value and total carbohydrates, respectively.The developed methodology favourably compares with that reported in previous works in much restricted sample composition an

gures of merit which agree with the US-FDA statuary tolerance values.2005 Elsevier B.V. All rights reserved.

eywords:Carbohydrates determination; Fruit juice; Milk shake; Milk; Hierarchical cluster analysis; Partial-least-squares; Attenuated total reflectance; Infrared

. Introduction

Carbohydrates offer energy sources for vital metabolicrocesses and are constituents of cellular substances suchs nucleic acids, being also enzyme cofactors and structuralomponents of cell walls and cell membranes[1].

The main intake of carbohydrates in human beings ishrough food stuff and liquid beverages. As an example, theasic sources of carbohydrates, in a usual diet, are fruit juicer milk shakes that people generally have for breakfast.

∗ Corresponding author. Tel.: +34 3544900; fax: +34 3544900.E-mail address:miguel.delaguardia@uv.es (M. de la Guardia).

Apart form the beneficial properties of carbohydratesexcess also cause generalized vascular disease. Thecarbohydrate diet, which is now so popular, can result notin chronichyperglycemias, also called “high blood sugar” balso in obesity. Moreover, excess insulin also causes htension and helps initiate the sequence of events in the awall which leads to atherosclerosis and heart disease. Ntheless, it is not healthy to avoid carbohydrates; just cothe amount that someone eats[1].

Scientists in the food and beverage industry are facedmany different quality control tasks. Nowadays, knowleabout all foodstuffs is an important task for manufacturethe food industry in order to inform the consumers abou

003-2670/$ – see front matter © 2005 Elsevier B.V. All rights reserved.oi:10.1016/j.aca.2005.02.004

182 J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193

In this sense, there are many reasons for analysing juices andin general fruit-based beverages such as identity and authen-ticity control of raw materials, purity control, control of thehygienic status, determination and control of the parametersfor nutritional information, control of the conformity of therecipe, official regulations and quality control.

Traditional methods of food analysis, usually regulatedby organizations such as the US-food and drugs administra-tion (US-FDA, for food safety), the United State Departmentof Agriculture (USDA), or the AOAC (Association of Of-ficial Analytical Chemists), normally feature wet-chemicalmethods or some form of chromatographic separation[2–4].However, the need of a fast analytical control of too manysamples at day demands for alternative procedures based onspectrometric data of untreated samples.

A bibliographic search has been carried out in orderto investigate previous works related with the analysis ofcarbohydrates in juices, juices containing milk and milk by-products such as milk shakes by means of Fourier trans-form infrared (FTIR) spectroscopy and chemometrics (seeTable 1). The mid-infrared region at 4000–400 cm−1 is themost widely used range for various applications, generallyusing transmission, and more recently attenuated total re-flectance (ATR), measurements due to the high content ofwater in fruit-based beverages in the wavenumber range be-tween 1400 and 900 cm−1.

ina-td ym cov-es

h ast forr entf

the-a ges,T iousw cal-i ffer-eb PLS-F ean-s antw

ble,m ofc n-t threem ost oft ermi-n fc sa sion,b ssion

(PCR). Moreover, in some cases, a previous data treatmentsuch as First or Second Derivative was employed.

Taking into account all these considerations, there is anyprecedent about the determination of carbohydrates and en-ergetic value in complex datasets, comprising different kindsof samples. Thus, we decided to evaluate a fast and accuratemethod which allows us to estimate these important qual-ity parameters in several fruit and milk-based liquid foodsobtained from the Spanish market using ATR-FTIR spec-troscopy in order to provide an appropriate tool for the qual-ity control of these products, widely consumed around theworld.

As fruit juice and milk shakes are basically water solu-tions, suspension and/or emulsions, we decided to evaluatethe use of a ZnSe ATR cell for FTIR measurements, insteadof transmission flow cells in order to avoid problems foundin the transmission measurements for high content of watersamples.

Due to the heterogeneity of the proposed sample set, theselection of calibration dataset is a critical point, which wasaddressed using a previously developed methodology basedin hierarchical cluster analysis[18].

Calibration models were selected by means of the (i)cross-validation (RMSECV), (ii) the root-mean-square-errorof prediction (RMSEP) and (iii) the relative standard devia-tion of predicted values of an independent validation set ofs

ew t mul-t ucedf t min-i

nalc se-l eticv

2

theO ntC mpI omGt tion.R mon-i of± ndw

2

eo rtantr kes

The most significant works are related with the determion of carbohydrates in apple juice[5–7], mango juice[8],iverse mixtures of sugars[9], milk [10], synthetic ternarixtures of glucose, fructose and sucrose applied to rery studies in soft-drinks and fruit juice[11,12], fruit-basedoft-drinks[13] and sugar cane juices[14,15].

Some other application of FTIR can be mentioned suche detection of adulteration in fruit concentrates by ATRaspberry purees[16] and using transmission measuremor apple juices[17].

In order to provide a complete picture of the state-of-rt of PLS-FTIR analysis of fruit and milk-based beveraable 1 summarizes the main characteristics of prevorks like the number of samples and composition of the

bration and validation set, the typical absolute mean dince (dx–y) and standard deviation of mean differences (sx–y)etween predicted versus actual values obtained byTIR methods and chemical methods and the root-mquared error of prediction (RMSEP) of most significorks.As it can be concluded from the aforementioned ta

ost of the works reviewed involved the quantificationarbohydrates in juice of only one kind of fruit or for syhetic samples prepared, using aqueous mixtures of theain carbohydrates (glucose, fructose and sucrose). M

hese works were focused only on the main sugars detation, but few estimates (i.e.,[11,12]) the total content oarbohydrates. As it can be seen inTable 1calibration modelre mainly built using partial-least-squares (PLS) regresut less recent works uses principal component regre

amples.As it is well known, the elimination of non-informativ

avenumbers yields, in many cases, the best and robusivariate models. Therefore, direct efforts have been condor selecting the best spectral range for each analyte thamizes both prediction and calibration errors.

Analytical figures of merit, based on net analyte sigalculations[19], were obtained for the sensitivity and

ectivity of the determination of carbohydrates and energalue in fruit and milk-based beverages.

. Experimental

A Nicolet 750 FTIR spectrometer controlled bymnic for Windows software from Nicolet Instrumeorp. (Madison, WI, USA) and equipped with Specacla

N-Compartment Contact Sampler horizontal ATR frraseby Specac (Orpington, UK) with a 45◦ crystal ZnSe

hrough top-plate, was employed for spectra acquisioom and sample compartment temperatures were

tored using a mercury thermometer with a precision0.5◦C. Both temperatures did not differ significantly aere 26± 1◦C during spectra acquisition of all samples.

.1. Samples

Sixty five commercial bottle or tetrabrick® samples werbtained from the Spanish market, covering an impoange of available types of fruit juices, flavour milk sha

J.Mo

ros

eta

l./An

alytica

Ch

imica

Acta

53

8(2

00

5)

18

Table 1Summary of the content of previously published PLS-FTIR procedures for the determination of carbohydrates in juices and related samples

Reference Calibration set Validation set Selectionmode

Factors number dx–y sx–y RMSEP Technique

Type Standardsamount

Type Samplesamount

[6] Synthetic apple juice(sucrose, glucose,fructose, sorbitol, citricacid, and malic acid)

39 Commercial juicevarieties and juiceextracted from differentapple varieties

11 – PCR 9, PLS 8 PCR−0.012, PLS−0.009

PCR 0.030, PLS0.021

PCR 2.5, PLS 2.2 PLS-ATR-FTIR

[8] Standard mixtures ofglucose, fructose,sucrose, and citric acid

22 Mango juice samples 8 – Sucrose 4, fructose4, glucose 4

0.18 3 Sucrose 5.1%,fructose 6.7%,glucose 8.0%

First derivative-PLS-ATR-FTIR

[9] Standard mixtures ofglucose, fructose andsucrose

– Standard mixtures ofglucose, fructose andsucrose

– – – – – 0.34% w/v (SEC),1.3% w/v (RMSEP)

Transmission PLS-I(first derivative)

[10] Commercially availablemilks

33 Commercially availablemilks

48 Clusteranalysis

Fat 6, protein 7,carbohydrates 12,calories 10, Ca 12

0.06% w/v fat,0.03% w/v protein,−0.15% w/v CH,0.06 kcal,−4.5 mgCa/100 mL

0.38% w/v fat,0.18% w/v protein,0.41% w/v CH,3.8 kcal, 9 mgCa/100 mL

0.38% w/v fat,0.18% w/v protein,0.41% w/v CH,3.8 kcal, 9.5 mgCa/100 mL

PLS-ATR-FTIR

[7] Pure apple juice dilutedwith Milli-Q water

First set(2–20%)[A] 134,second set(25–100%)[B] 86

Pure apple juice dilutedwith Milli-Q water

A 39, B 44 – PLS A 2, B 2;SIMCA (differentfor each standard)

– – – PLS, potentialfunctions, SIMCAand ANNs modelsto classify 23 applejuice-basedcommercialbeverages

[11] Ternary mixtures ofglucose, fructose,sucrose with Milli-Qwater

8 Synthetic ternarymixtures

14 – Sucrose 4, fructose4, glucose 4

– – 0.7% PLS-ATR-FTIR(absorbance andfirst-derivativemode)

[12] Ternary mixtures ofglucose, fructose,sucrose with Milli-Qwater

8 Synthetic ternarymixtures

[13] Sucrose, glucose,fructose and�-gluconolactone aqueouscalibration standards

8 Commercially availablefruit juice drinks

[14] Raw sugar cane juices 107 Samples that cover thentire sugar harvest anddifferent geographicalregions

[15] Raw sugar cane juices 20 Spectra of raw, sugarcane juices

1–

19

3183

14 – Sucrose 4, fructose4, glucose 4

– – 0.7% PLS-ATR-FTIRand PLS-Circle cell-FTIR

3 – 2 (mean centeringpre-processingdata)

−8.3 8.9 7.2% FTIR

e1267 – 5 0.17 1.4 – PCA-PCR-ATR-FTIR

19 – – 0.09 0.17 – ATR-FTIR

184 J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193

(cocoa, vanilla, etc.) and mixtures of fruit juices and milkproducts. Thirty nine samples were juices from different fruitmixtures (19 of them also contained milk) and the others26 samples were composed by: 10 juices obtained from asingle fruit, 13 flavour milk shakes and 3 milks (whole, semi-skimmed and skimmed).Table 2shows a brief description ofeach sample.

The sell-by-date of the samples varied between end De-cember 2002 and beginning of November 2003, while theproduction dates range from beginning of June to middle ofAugust 2002. Therefore it is considered that samples werenearly of the same age, while the differences in the sell-by-date values are ascribed to the differences in the preservationstrategy and the characteristics of the product. ATR-FTIRspectra were obtained in October 2002. Concentration ref-erence data of energetic value (EV) and total carbohydrates(CH) values of the samples were provided by the producers.

2.2. FTIR analysis

Original samples were placed in the same temperaturecontrolled room where the spectrometer was located before tocarry out the analysis. The sample compartment temperature

Table 2General description of samples employed in this study

Sn p

Sano

33

3

444

11 41 41 41 41 411112 5222 52222223333

was monitored and it remained stable at 26± 1◦C during theacquisition of IR spectra of all analyzed samples.

Samples were only shaken gently before filling the ATRcell. After shaking, three sub-samples were poured into theATR cell and the FTIR spectra were taken as described below.As some samples did contain some pulp, the effect of theacquisition time and the possible deposition of substances onthe ATR cell were assessed (see Section3).

In order to select the instrumental conditions for carry-ing out the ATR measurements, several parameters includingthe number of accumulated scans per spectra, nominal reso-lution and mirror velocity were tested. After this evaluationthe following conditions were chosen: sample spectra werescanned between 4000 and 600 cm−1, by averaging 75 scansper spectrum with a nominal resolution of 8 cm−1 (data spac-ing of 3.86 cm−1) and a mirror velocity of 0.6329 cm s−1. Theacquisition of each averaged spectrum requires 57 s.

The background and blank spectra were collected fillingthe ATR plate cell with Millipore Q-purified water (Bedford,MA, USA) and using the same instrumental conditions thanthose employed in the case of samples. Background spectrumwas scanned at a seven samples interval, while blank spec-trum was collected after the measurement of each sample, for

ampleo.

Brand Milkproduct

Juice Multifruits Fiberor pul

1 Pascual Milk Yes Yes No2 Choleck Milk Yes Yes Yes3 Juver Milk Yes Yes Yes4 Central Asturiana Milk Yes Yes No5 Hero No Yes Yes No6 Juver No Yes Yes Yes7 Feiraco Milk No No No8 Feiraco Milk No No No9 Carrefour Milk No No No0 Pascual Yogurt Yes Yes No1 Hero No Yes Yes No2 Juver Milk Yes Yes Yes3 Juver Milk Yes Yes Yes4 Celta Milk Yes No Yes5 Hero No Yes Yes No6 Juver No Yes Yes Yes7 Carrefour No Yes No No8 Carrefour No Yes No No9 Carrefour No Yes Yes No0 Son Fil Milk Yes No No1 Kas Fruit No Yes Yes No2 Kas Fruit No Yes Yes No3 Capel Big Vit No Yes Yes No4 Compal No Yes Yes Yes5 Compal No Yes Yes Yes

6 Compal Milk Yes No Yes 57 Compal Milk Yes Yes No 68 Compal Milk Yes Yes Yes 69 Solan cabras No Yes Yes No0 Solan cabras No Yes Yes Yes1 Solan cabras No Yes Yes No2 Puleva Milk Yes Yes Yes3 Puleva Milk Yes No No

mple.

Brand Milkproduct

Juice Multifruits Fiberor pulp

34 Puleva Milk Yes Yes Yes5 Nutrexpa Milk No No No6 Central Asturiana Milk Yes Yes No

37 Central Asturiana Milk No No No8 RAM Milk Yes No No

39 Fuente Liviana No Yes Yes No0 Kasfruit No Yes No No1 Juver No Yes Yes Yes2 Hero No Yes No No43 Emig No Yes Yes No4 Juver No Yes No Yes5 Carrefour Milk No No No6 Carrefour Milk No No No7 Compal No Yes Yes Yes8 Carrefour No Yes Yes No

49 Puleva Milk No No No50 Puleva Milk No No Yes51 Puleva Milk Yes No No52 Capel Big Vit Milk Yes No No3 Fuensanta No Yes No Yes54 Fuensanta No Yes Yes No55 Puleva Milk No No No6 Hacendado Milk Yes Yes Yes

57 Hacendado Milk Yes Yes Yes58 Hacendado Milk Yes Yes Yes

9 Hacendado Milk Yes Yes Yes0 Pascual Milk Yes Yes No1 Pascual Milk Yes Yes Yes62 Carrefour Milk Yes Yes Yes63 Auchan Milk No No No64 Auchan Milk No No No

65 Auchan Milk No No No

J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193 185

assessing that cross-contamination of IR spectra was mini-mized after cleaning the ATR crystal.

Both, sample and blank spectra were collected in the ab-sorbance mode. The regions between 4000 and 3200 cm−1

and between 840 and 600 cm−1 were eliminated prior thecalculations as it was observed that variations in these re-gions cannot be ascribed to variations in sample compo-sition. Furthermore, the absorption band around 2400 and2260 cm−1, which are due to atmospheric CO2, were alsocut-off from raw spectra. These bands could be also min-imized by purging the ATR cell with N2. Nevertheless, asthe IR region in which they are located does not provide anyrelevant information about samples, N2 purging was con-sidered unnecessary and this region was cut-off from rawspectra.

The spectra of the three sub-samples of each sample weretaken by refilling the ATR cell.

2.3. Data analysis

Data obtained from Ommic were exported in text formatand analyzed using Matlab® (The Mathworks Inc., SouthNatick, MA, USA). First of all internal and home-made Mat-lab functions were used for hierarchical cluster analysis inorder to evaluate the similarity of samples in terms of theirATR-FTIR spectra and to assess the number of characteristics Sim-i ionh sm

re-d f thec of them

raticst ) isa , andi

R

w r (ino sam-p bero libra-tn

M-S delf f thed tion,et s inw one-

out cross-validation d.f. is equal to the number of calibrationsamples).

As the ability of the model to fit the calibration data isnot a direct measurement of its prediction capabilities, it ismandatory to compare the values predicted for new samplesnot used to build the model. This can be performed by cal-culating the root-mean-square error of prediction (RMSEP)when the model is applied to new data for which the refer-ence values are known. RMSEP is calculated exactly as inEq. (1) except that the estimates forCi are based on a previ-ously developed model, in which the sample concentrationsof the validation set are excluded in the model-building stepand the degree of freedom is the number of samples in thisset.

In order to build and select PLS models, the following it-erative procedure was carried out for. For building the bestcalibration model, a selection of the optimum number of fac-tors, which minimize the root-mean-square error of cross-validation, was made based on the criterion of Haaland andThomas[21]. To improve the prediction performance of theregression method, a search for suitable sensors was consid-ered. In this sense, one subroutine from MVC1 toolbox wasused to find the minimum PRESS, as a function of the num-ber of factors, based on a moving spectral window strategy[20]. Several spectral windows were tested in order to eval-uate their prediction capabilities for the validation set. Onlym

2

eens latedu ) ofn t dif-f amplei nd ac Thisd in dif-f link-a ords,b com-p atedw pre-s lisedo max-i plesi t tol om-p m isu um-bt ini-m n-c sent ut-offv

ubsets in which the available samples could be divided.lar criteria to that already published for milk classificatave been used[10]. Multivariate calibration calculation waade with the MVC1 toolbox[20].The following figures of model’s fit to the data and p

ictive power have been use throughout the text. In all oases, the scope was to estimate the average deviationodel from the data.PRESS is the sum of squares prediction error (quad

um term in Eq.(1)) for the model, which includesA fac-ors. The root-mean-square error of calibration (RMSECmeasure of how well the model fits the calibration data

s defined as:

MSEC=[(

n∑i=1

(Ci − Ci)2

)(d.f.)−1

]0.5

(1)

hereCi means the values of the predicted parameteur case energetic value and carbohydrates) when allles are included in the model building and d.f. is the numf degrees of freedom calculated as the number of ca

ion samples with known concentration (Ci) minusA+ 1, theumber of factors kept in the model plus one.

The root-mean-square error of cross-validation (RECV) is a measure of predictive ability of the mo

ormed on part of a dataset to predict the remainder oata. The RMSECV is defined as the previous equaxcept thatCi are predictions for samplesnot included inhe model formulation, and d.f. is the number of timehich the cross-validation is repeated (i.e., in the leave-

ost significant results will show here.

.4. Cluster analysis

In hierarchical cluster analysis, the similarity betwamples is calculated using the distance concept, calcusing a mathematical relationship (i.e., Euclidian normumerical properties of the samples (i.e., absorbance a

erent wavelengths). In a successive procedure, each ss linked to the closest sample or group of samples aharacteristic distance is used to describe this union.istance between groups of samples can be evaluated

erent ways and is the main difference among commonge methods (Ward, complete, average, etc.). In other wy this procedure, each sample is replaced by a grouprised of the sample and their neighbour samples locithin the given similarity distance. The results are reented in a dendrogram, which shows at which normar rescaled distance (i.e., each distance rationed to the

mum distance, multiplied by a factor) a group of sams differentiated from others, when it is read from righeft. At the far-left end each replicate of each sample crises a group of one member, that is, each spectrunique. Thus, for a given rescale distance, different ner of groups are kept. At this stage we have proposed[10]

o use the similarity distance between triplicates as mum cut-off criterion. Actually, taking into account the co

epts of limit of detection and quantification, we have choen times the average distance between replicates as calue.

186 J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193

3. Results and discussion

3.1. Instrumental and experimental conditions for FTIRdata acquisition

Before the acquisition of IR sample spectra, several instru-mental and experimental conditions were evaluated in orderto improve the spectra quality.

The number of accumulated scans per spectrum and thenominal resolution were tested in terms of signal to noiseratio and the time required for spectra collection.

The assays were carried out by varying the nominal res-olution between 2 and 16 cm−1 and the number of scans be-tween 25 and 100 for MilliQ-water, juice and a flavour milkshake as samples. The noise in the spectrum of these totallydifferent samples was compared, using the RMS (root-mean-square) estimator, in the region where no bands were observed1900–2200 cm−1. As in this spectral range no significant dif-ferences in the noise level has been noticed for these samplesin all tested conditions, it was considered that the use of wa-ter was suitable for selecting the instrumental conditions in aspectral range with more analytical significance. This regionwas selected seeking the greater number of bands observedand related to the sample composition (see below). Therefore,the RMS was evaluated using water as sample in the spectralregion compressed between 1800 and 1000 cm−1 (seeFig. 1).

isew sa e of5 vablew tet s).

al-u s

F ratioo TR-F 1800a s in-d icatest signalt ls seet

with a nominal resolution of 8 cm−1. It also was evaluatedin terms of signal to noise ratio and the time required forspectrum collection (67, 57 and 43 s respectively). The besttime/benefits compromise was found for a mirror velocity of0.6329 cm s−1.

On the other hand, in order to minimize the cross-contamination and to establish an appropriate strategy forcleaning the ATR cell between samples, different proce-dures were tested using background and blank controls. Sat-isfactory results were achieved when samples were removedfrom the ATR device with a 5% w/v commercial deter-gent/water solution, which was removed afterward by a 10%v/v ethanol/water solution. Methanol was finally used for dry-ing quickly the surface of the ATR cell. The whole cleaningprocedure consumes less than 60 s.

To evaluate possible sample sedimentation on the ATR cellduring the spectra acquisition for samples containing fiber orpulp continuous a series of spectra were collected with a1 min interval for the ATR cell filled with a multifruit juice.No changes on the spectra were observed after 15 min.

3.2. Juice and flavour milk shake FTIR spectra

Fig. 2shows the spectra of a flavour milk shake sample, amultifruit juice sample, a multivitamin fruit-based softdrinkand MilliQ-water blank. Regions compressed between 4000a dh o-n werei

ar-b akea einsa ofs om2 tionb fatm lyc-e ) thea ups( rbe a-sC nlyir o thea

ea and1 ofp and1 willb witht -p ratesc on of

Using this criterion, the best time/benefits compromas found for a nominal resolution of 8 cm−1 and 75 scanccumulated per spectrum for which a measurement tim7 s was required. The decrease in the RMS value achiehen acquiring 100 scans (seeFig. 1) does not compensa

he increment on the acquisition time (approximately 20Mirror velocity was also tested at three different v

es (0.4747, 0.6329 and 0.9494 cm s−1) averaging 75 scan

ig. 1. Effect of the instrumental parameters on the signal to noisef ATR-FTIR spectra of flavour milk shake and fruit juice samples. ATIR analysis evaluated in the spectral region compressed betweennd 1000 cm−1 using RMS (root-mean-square) estimator; black pointicate actual condition measured. Dashed line in the floor plane ind

he instrumental condition chosen to ensure a compromise betweeno noise ratio and time required for spectra collection (for more detaihe text).

nd 3200 cm−1 and between 840 and 600 cm−1 presenteigh absorption of the incident light principally by compents of the system (water and the ZnSe crystal) and

gnored for this study.In Fig. 2 the main wavelengths belong to protein, c

ohydrates and fat present in fruit juice, flavour milk shnd milk are shown. The amide groups (CONH) of protbsorb at 1546 cm−1, whereas the hydroxyl groups (OH)ugars absorb at 1052 cm−1. The increase in absorbance fr600 cm−1 upward is ascribed to overtones and combinaands of saccharides[22]. The chemical structure of theolecules can be divided into two portions: (A) the grol connected by an ester linkage (carbonyl) and (Blkyl chains of the fatty acid. The carbon–hydrogen groC–H) and the carbonyl groups (CO) of samples absonergy at 2929 and 1653 cm−1, respectively. The Fat A meurement counts the ester linkage, while Fat B uses CH2 andH3 absorption centres[22], being these bands presents o

n products containing milk. The band around 928 cm−1 cor-esponds to the symmetric COC skeletal stretch due tliphatic ethers[23].

As it can be seen inFig. 2, spectra from multifruit juicnd milk shake differs mainly in the region 3000–2500700–1500 cm−1. These differences are due to the lackroteins and fats in juices. The region between 1500200 cm−1 correspond to combination bands and, as ite shown, it can be correlated for both type of samples

he energetic value. In the 1200–1000 cm−1 region both samles shows absorption and it is related to the carbohydontent. It can be also seen that the number and positi

J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193 187

Fig. 2. ATR-FTIR spectra of a flavour milk shake (a), tropical multifruit juice (b) and a multivitamin soft drink containing multifruit juice (c), in comparisonwith a MilliQ-water blank (d). Experimental conditions: nominal resolution of 8 cm−1, 75 scans per spectra accumulated and a mirror velocity of 0.6329 cm s−1.Sample composition: (a) 0.9% fat, 2.5% protein and 12.8% carbohydrates; (b) 0.1% fat, 0.31% protein and 9.6% carbohydrates; (c) 0.04% fat, 0.1% proteinand 1.7% carbohydrates.

bands in this region changes depending on the type of sample,because the total and relative amount of the different saccha-rides are sample dependent. As an example, the occurrenceof lactose band will be only seen in milk-based beverages.For samples containing fruit juices and milk derivatives,bands related with lactose, fructose, glucose and sucrosewill be observed. This variation on the relative and totalamount of saccharides really imposes a challenge when seek-ing the determination of total carbohydrates by multivariatemethods. Therefore a suitable correction of ATR-FTIR spec-trum and selection is mandatory for achieving representative-ness.

3.3. Correction of ATR-FTIR data

When comparing all ATR-FTIR collected spectra for sam-ples considered and in spite of the careful control of the cellcleaning a significant shift of the baseline, sometimes evenwithin replicates of the same sample, was observed. Thesevariations make difficult the subtraction of blanks to the cor-responding sample spectra.

Different procedures to deal with this problem in milksamples has been evaluated in the literatures[10,24]. Ourprevious experience shows that good results, evaluated asminimum misclassification of samples (see below) and bet-ter prediction capabilities were achieved when the averagea spec-t

In this case, the average of absorbance values between2000 and 2200 cm−1 was used for performing the spectralcorrection. In consequence, these corrected spectra were uti-lized for further calculations.

As the optics involved in ATR are quite different fromthose used in the transmission experiments, the infrared spec-trum of a sample obtained by ATR exhibits some significantdifferences when compared to its transmission counterpart.Some of these differences are desirable and have been used toimprove the possibility of measuring aqueous samples. Lessdesirable characteristics of ATR are its distortion of the rel-ative intensities of bands and the introduction of a shift tolower frequencies.

The distortion of relative peak intensities in an ATR spec-trum is well known. In the transmission experiment, thepathlength is defined by the thickness of the sample andis therefore constant across the spectrum. However, in theATR experiment, the depth to which the sample is penetrated(dp) by the infrared beam is a function of wavelength (λ),incident angle (θ), ATR crystal refractive index (nc), andsample refractive index (ns) (see Eq.(2)). The most sim-ple and common-applied correction is based on the hypoth-esis that the effect of the incident angle, ATR crystal refrac-tive index, and sample refractive index on the absorbanceis fairly constant between samples. Therefore, as the pen-etration depth is inversely proportional to the wavenum-b ectedb ber

bsorbance in a fairly flat region was subtracted to eachrum for correcting additive artifacts.

er, the absorbance at a given wavenumber is corry multiplying the absorbance value by the wavenum

188 J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193

value and rationing the result by the maximum wavenumberin the spectrum. The resulting spectrum can be optionallyrescaled to the highest absorbance acquired for visualisationpurposes.

dp = λ(2π)−1(n2c sin2θ − n2

s)−0.5

(2)

This correction has dissimilar application rate in the analyti-cal literature. In general it is always done when spectra willbe compared to a reference spectrum taken in transmissionmode, but is more rarely applied for quantitative analysis.A priori, it is possible to expect that this correction will beinsignificant when quantitative analysis is done at a singlewavelength or small wavelength range. Nevertheless, whenthe whole spectra are used with multivariate regression anal-ysis, such a non-linear correction may improve or worsenthe prediction capabilities. Moreover, when a cluster analy-sis strategy is used, as in this work, to assess how samples canbe grouped from their IR spectra, the increase in absorbancevalues at high wavenumbers, produced after correction, mayaffect the cluster results.

We have made an iterative study applying or not the afore-mentioned correction, grouping the samples based on clus-ter analysis, selecting the calibration and validation set fromthese groups and generating and assessing the multivariatem ge,e uchi wnh

3 sf

con-s datat icha set,t lus-t f se-l vourm ore-o mayb eringt

nal-y wec or se-lF valu-a thesev .

-i dis-t ach ot which

were 47, 16, 44 and 24, respectively, with different numbersof samples in each one.

When Euclidean, Euclidean square and Minkowsky fourdistances were compared between them we found that clus-ters were formed nearly by the same samples. Moreover, thethree dendrograms can be reduced to the same dendrogramby choosing in each case an appropriate cut-off rescaled dis-tance. On the other hand, the dendrogram obtained using anyof these three distances could be easily explained in termsof sample composition. Therefore, the use of any of thesethree measuring distances can be considered as equivalent.However it was not the case when it was used the Pearsoncorrelation distance. For this measuring distance there is atrend to form few cluster comprised of a high number ofsamples, but “chemically” different between them. In fact,the spectrum of a dilution of a given sample is highly cor-related (Pearson coefficient almost equals to 1) with that ofthe undiluted sample. In our opinion, this distance groups interms of relative similar composition (i.e., similar ratios offats to carbohydrates), which is not useful for selecting a cal-ibration set that spans a large composition range and it wasmore or less the same on using Euclidean square criterion.

Cluster analysis could be done also with the scores of themost significant principal components selected after a prin-cipal components analysis (PCA), centring or not data. Forthis particular case, only Euclidean distance was tested. Thed ainedi upsw s oft

raws mplee Thisf smalld viousP upsf plesi ucest , thec

den-d witha us-i andp

tedm kagem andt efore,f ge.

of anA wase sis.

tiono CAf ATR

odel from the number of factor, optimum spectral rantc., for each parameter. As this process implies too m

nformation, only the most significant results will be shoere.

.4. Clustering of juices and flavour milk shake samplerom their FTIR spectra

In order to evaluate possible classes among samplesidered, a clustering method was carried out before PLSreatment. In our opinion, it is a very important task whllows us to select properly a representative calibration

hus improving the prediction of unknown samples. This cering method should be done prior the determination oected target parameters of commercially juices and fla

ilk shakes of different composition and trademark. Mver, the classification of samples based on their spectrae interpreted through the obtained clusters also consid

he differences in sample composition.In a previous work we applied a hierarchical cluster a

sis for selecting a calibration set for milk analysis andonfirmed that this procedure has some advantages fecting the calibration set when compared to others[10,18].urther investigation has been carried out in this way, eting other distance and linkage methods. The effect ofariables on the obtained dendrogram will be discussed

In this study, several distance types[25] such as Euclidan distance, Euclidian square distance, Minkowsky fourance and Pearson correlation distance, were tested. Ehese distances generates a different number of groups

f

endrograms obtained from PCA data can be also expln terms of sample nutritional values although some groere modified due to variations in the fingerprint region

he spectra.The consideration of the Euclidean distance using

pectral data provided 36 clusters including a single saach one and incorporates one misclassified triplicate.

act may suggest that this strategy is more dependent onifferences in sample spectra than that based on a preCA, which reduces a little bit the number of sample gro

rom 47 to 43 but especially increases the number of samncluded in each group. Moreover, as PCA drastically redhe number of variables to be considered for clusteringomputer memory requirements are also reduced.

Based on the obtained results we have selected therogram classification found using Euclidean distanceverage within-groups linkage for the scores of the PC

ng not centred data to be compared with other linkagere-treatment method.

When applying the Ward linkage method, it originaore compact clusters than complete or average linethod. The number of cluster was reduced from 43 to 20

he number of samples in each cluster increased. Therollowing comparisons were performed using Ward linka

The effect on the classification of samples of the useTR correction algorithm for a previous data treatmentvaluated using the Euclidean distance after PCA analy

Fig. 3A and B shows the dendrographic classificaf samples obtained with five factors extracted via P

rom spectral data corrected and non-corrected via the

J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193 189

Fig. 3. Dendrographic classification of samples using the Euclidean distanceafter PCA analysis of ATR-FTIR spectra and applying the Ward linkagemethod. (A) Corrected with an ATR algorithm and (B) uncorrected. Cut-offdistance was set to 1.4 a.u. Typical dissimilarity distance between replicates0.5 a.u.

correction algorithm, respectively, using Ward linkage. As itcan be seen, in most cases the application of the correctiondoes not affect the group of samples in each small cluster(most samples are grouped together with the same samplesin both cases), but changes the distribution of the big groups,for example the cluster 1B inFig. 3A is a sub-cluster of the 1Ain Fig. 3B. These changes may be explained by the fact thatwhen applying the ATR correction, the intensity of the bandsin the spectral region around 2800–3000 cm−1 increase.

It is important to notice that the fact that for most samplesthe near neighbourhood of each sample does not change when

applying the ATR correction indicates that the selection of thecalibration samples will not be affected by this correction.In fact only two samples have different neighbour samplesin Fig. 3A when compared to those ofFig. 3B. Moreoverin both cases, the number of cluster formed is similar (19and 20, respectively). Therefore the application of the ATRcorrection is considered not important at the sample selectionstep.

The main group of clusters formed (from right to left upto rescaled distance equals to 5) are directly correlated withthe mean intensity of the FTIR spectra of the samples inthese groups and, thus, samples with high absorbance levelare grouped together. As the absorbance intensity is mainlyrelated with the total amount of carbohydrates, fats and pro-teins the groups are basically related with the similar contentof these analytes between the samples.

Table 3shows the mean and the standard deviation ofthe energetic value and the carbohydrate content of themain five clusters obtained using the Euclidean distanceafter PCA analysis of corrected ATR spectra (Fig. 3A).Basically, clustering criteria seems to be based on carbo-hydrates content, which separated samples into two maingroups. In each one, samples were separated using other pa-rameters such as protein and fat (which depends strong onthe milk presence). So, it can conclude that the energeticvalue of fruit and milk-based beverages is the most impor-t ew-p

thes re is as mainp ices( untsl tiono eenc

3

plest lti-v tion,t Int lustera

Table 3Characteristics of the samples grouped in the main clusters

Cluster branch Number ofsamples

EV (kJ),mean± S.D.

CH (g),mean± S.D.

Sa

1B 11 293± 39 13.8± 1.3 1,1A top 14 248± 36 12.3± 1.4 2,1A button 22 201± 22 11.2± 1.3 3, 52, 532A 14 82± 17 4.2± 1.0 112B 4 190± 45 4.5± 0.1 49

Note: button and top means the upper and lower branch of group 1A inFig. 3A. Da ean).

ant characteristic from both, alimentary and FTIR vioints.

Fig. 4 shows a plot of the carbohydrate content ofamples vs. their energetic values. As can be seen, thetrong correlation between these two parameters for theart of samples being those corresponding to pure fruit juseeTable 2). For samples containing appreciable fat amoike juices with milk samples or milk shakes, the contribuf fat to the energetic value avoids a close correlation betwarbohydrates and energetic value.

.5. Selection of the calibration set

The determination of the number and the nature of samo be used for calibration is always a critical factor in muariate analysis. As it was pointed out in the previous seche choice ofFig. 3A or Fig. 3B is of no great concern.he present work it was done based on the hierarchical cnalysis results, using dendrogram ofFig. 3A for selecting

mple number

4, 8, 9, 10, 21, 35, 36, 37, 46, 507, 33, 38, 45, 51, 55, 56, 57, 58, 59, 60, 61, 625, 6, 12, 13, 14, 17, 19, 20, 22, 24, 25, 32, 34, 40, 41, 42, 44, 47, 48,, 15, 16, 18, 23, 26, 27, 28, 29, 30, 31, 39, 43, 54, 63, 64, 65

ta are referred to 100 mL of sample. S.D.: standard deviation (of the m

190 J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193

Fig. 4. Correlation between energetic value and total carbohydrate contentin fruit and milk-based beverages. The dashed line represents a regressionline of samples in which the energetic value is highly correlated with thecarbohydrates content (CH (g/100 mL) = 0.060EV (kJ/100 mL)− 0.528). Toidentify the each samples seeTable 2.

the calibration and validation datasets. The sample selectioncriterion was based on the following principles. At least onesample of each cluster was selected for calibration. If thecluster is comprised of more than one sample, the number ofsamples selected for calibration was approximately the rootsquare of the total number of samples included in the cluster,while the remaining samples were integrated in the validationdata set. So, the number of samples assigned to the validationwas equal or higher than the number of those employed forcalibration. The samples within a given cluster were selectedrandomly.

As summary, 27 samples, with an average energeticvalue of 204± 69 kJ/100 mL and a carbohydrate content of11.0± 3.3 g/100 mL were selected for calibration (±valuescorrespond to the standard deviation) and 38 samples with

an average energetic value of 224± 84 kJ/100 mL and a car-bohydrate content of 11.4± 3.7 g/100 mL were used for pre-diction.

In order to evaluate the representativeness of the afore-mentioned calibration and validation sets, an extended cal-ibration model was also used. This extended calibration setincludes the samples of all individual clusters but now theroot square of samples from clusters with more than onewas reserved for validation and the rest included in the cal-ibration set. This calibration set was composed by 40 sam-ples, corresponding to 203± 81 kJ/100 mL and 10.7± 3.8 gof carbohydrates per 100 mL and the remaining 25 sampleswere using for validating the model. For this validation set,the energetic value and carbohydrate average content were203± 70 kJ/100 mL and 10.7± 2.9 g/100 mL, respectively.

3.6. Determination of the energetic value

Different models were built and compared in terms of RM-SECV and RMSEP values for both, calibration and extendedcalibration sets. In all the cases, slightly better RMSEP val-ues (around 2%) were obtained using spectra corrected bythe ATR algorithm (data not shown).

It is interesting to notice that there are two spectral rangesin which a good prediction of the energetic value of samplescan be carried out. These ranges are the 1300–1500 cm−1 andto eterf

nto learr ctralr , a re-d iblee rentc tures

Table 4Prediction capabilities of PLS-ATR-FTIR for energetic value determination o

Set Factors Spectral range RMSECV (kJ/100 m

1 5 2769–2923 13.2 92 5 2769–2923 16.2 941 5 1461–1500 16.1 842 4 1300–1493 16.7 66

Set 1 (2769–2923 cm−1)

Overall Milk shakes ces

ssd 4sSSL

N , where RMSEP, is theR multip of the normo e the st r mored

trip (kJ/100 mL) 2.4 2.8

reg (kJ/100 mL) 3.5 5.9

x–y (kJ/100 mL) 2.8 −10.5

x–y (kJ/100 mL) 18electivity (%) 5.1ensitivity 4.1E−05D (kJ/100 mL) 9.5

ote: Set 1 corresponds to a calibration set composed of 27 samplesMSEP divided by the mean value of EV in the validation dataset andf net analyte signal of 10 blanks divided by the sensitivity.strip andsreg aretails see the text.

he 2760–2923 cm−1 regions.Table 4andFig. 5 show theptimised prediction capabilities obtained for this param

or both sets and spectral ranges.As it can be seen, the 1300–1500 cm−1 range is depende

n the composition of the calibration set. There is not a ceason that justifies the fact that when the optimum speange of the extended set is used for the reduced oneuction on the prediction capabilities is obtained. A possxplanation is that this region is formed by many diffeombination bands, and thus, a greater variability of fea

f fruit and milk-based beverages

L) RMSEP (kJ/100 mL) RRMSEP (%) R2

18.4 8.2 0.9313.8 6.8 0.9516.9 7.5 0.9412.6 6.2 0.96

Set 2 (2769–2923 cm−1)

Juices Overall Milk shake Jui

2.3 2.2 2.0 2.33.5 3.5 3.7 3.4.1 2.8 4.7 2.0

185.3

4.1E−059.9

as Set 2 is an extended calibration set composed of 40 Samples. Rlied by 100. LD was calculated as three times the standard deviationandard deviation and the standard error of prediction respectively. Fo

J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193 191

Fig. 5. Prediction of energetic value in samples of fruit and milk-based beverages. Predicted vs. actual values and their elliptic confidence test forthe slopeand the intercept for (A)–(C) using a reduced calibration set of 27 samples (set 1) and (B)–(D) for an extended calibration set of 40 samples (set 2). Spectralranges: (A) 1461–1500 cm−1, (B) 1300–1493 cm−1, (C) 2769–2923 cm−1, (D) 2769–2923 cm−1.

should be present in the calibration set for obtaining a morerobust model.

From the consistency of the spectral range, we select the2769–2923 cm−1 region for evaluating the analytical perfor-mance of the methodology.Table 4includes the figures ofmerit obtained under the aforementioned conditions for bothcalibration sets discriminating the prediction capabilities asa function of the type of samples.

The reproducibility of the determination, established fromthe mean standard deviation of each triplicate and the stan-dard error of prediction (that includes the uncertainty fromthe model[26,27]) were 2.4 and 3.5 kJ/100 mL. Additionallyit can be seen fromTable 4that there is no significant differ-ence in the obtained figures of merit, neither between typeof samples nor on the basis of the number of samples in thecalibration set.

For comparative purposes it must be noticed that the tol-erance levels of variation accepted by the US-FDA are 21and 42 kJ/100 mL for samples with an energetic value lowerthan 210 and higher than 210 kJ/100 mL, respectively[2].These reference values indicate that ansx–y of 18 kJ/100 mLobtained by the developed procedure is clearly acceptable.

Fig. 5 shows the good regression between predicted val-ues and actual ones of the energetic value of fruit juice and

milk shake samples for both calibration sets employed andtwo spectral ranges evaluated. This figure also evidences theabsence of systematic error as the theoretical point interceptequal to 0 and slope equal to 1 were located inside the corre-sponding confidence ellipses.

3.7. Determination of total carbohydrates

A similar procedure to that followed for energetic valuedetermination in fruit and milk-based liquid foods was madefor building a calibration-prediction model for total carbohy-drates concentration. In this case, there is only one spectralregion correlated with this parameter, which is that corre-sponding to the absorption of the hydroxyl groups of sugars(seeFig. 2). Table 5andFig. 6show the optimised predictioncapabilities obtained for this parameter using both calibrationsets assayed.

As it can be seen, there is neither significant differencein the main prediction indicator values nor in the optimumspectral range (1020–1175 cm−1) selected for models con-structed with 27 or with 40 samples. Moreover, the optimumnumber of extracted factors is only 3 in both cases. This is animportant fact as the robustness of multivariate models de-creases as the number of factors increases. On the other hand,

192 J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193

Table 5Prediction capabilities of PLS-ATR-FTIR for total carbohydrates determination in fruit and milk-based beverages

Set Factors Spectral range RMSECV (g/100 mL) RMSEP (g/100 mL) RRMSEP (%) R2

1 3 1020–1175 0.89 0.72 6.3 0.97002 3 1020–1175 0.91 0.65 6.1 0.9708

Set 1 Set 2

Overall Milk shake Juices Overall Milk shake Juices

strip (g/100 mL) 0.05 0.07 0.05 0.04 0.06 0.04sreg (g/100 mL) 0.19 0.20 0.18 0.26 0.23 0.27dx–y (g/100 mL) 0.06 0.13 0.04 0.08 −0.20 0.18sx–y (g/100 mL) 0.66 0.69Selectivity (%) 31.60 33.20Sensitivity 2.6E−02 2.7E−02LD (g/100 mL) 0.2 0.3

Note: Set 1 and Set 2 include 27 and 40 Calibration samples, respectively. RRMSEP, is the RMSEP divided the mean value of total carbohydrate content in thevalidation dataset. LD was calculated as three times the standard deviation of the norm of net analyte signal of 10 blanks divided by the sensitivity.strip andsreg

are the standard deviation and the standard error of prediction respectively. For more details see the text.

Fig. 6. Prediction of total carbohydrate concentration in samples of fruit and milk-based beverages. Predicted vs. actual values and their ellipticconfidence testfor the slope and the intercept for a reduced calibration set of 27 samples (set 1) (A) and an extended calibration set of 40 samples (set 2) (B). Spectralranges:(A) 1020–1175 cm−1, (B) 1020–1175 cm−1.

three factors can be perfectly justified taking into account theheterogeneity of the sample population, comprising not onlyfruit juices of different fruits, but also different compositionof milk and milk shakes.

Table 5also shows the figures of merit and predictioncapabilities of total carbohydrates for both calibration setsdiscriminated by the type of sample. For the reduced setof 27 samples, thedx–y and thesx–y values were 0.06 and0.66 g/100 mL, respectively. The reproducibility of the de-termination established from the mean standard deviation ofeach triplicate and from the standard error of prediction (keepin mind that this latter parameter includes the uncertainty ofthe model) were 0.05 and 0.19 g/100 mL, respectively. As itcan be seen fromTable 5, there is no significant difference inthe obtained figures of merit, neither between type of samplesnor for the number of samples employed in the calibration set.

As Fig. 6 indicates, the regression between predictedand actual values of the total carbohydrates in the samplesassayed, being observed that the confidence ellipses of theintercept and slope of these regression lines include thezero intercept and the one slope values, thus evidencing

the absence of systematic errors in the PLS-ATR-FTIRmethodology.

The increment rounding for the US-FDA is 1 g/100 mL forsamples containing more 1 g/100 mL, whereas the insignifi-cance amount is 0.5 g/100 mL[2]. From obtained results, itis clear that the proposed method is suitable for complyingwith these statuary values.

The selectivity and sensitivity for the PLS-ATR-FTIR de-termination of total carbohydrates in fruit and milk beveragesare higher than those obtained for the energetic value, becauseof the specificity of the optimised spectral range for sugars. Asa final remark and regarding the high heterogeneity of the cal-ibration set, the global performance obtained here favourablycompares with that obtained in previous works (Table 1) inmuch restricted sample composition.

4. Conclusions

In this work different aspects for the ATR-FTIR estimationof energetic value and total carbohydrates in commercially

J. Moros et al. / Analytica Chimica Acta 538 (2005) 181–193 193

available fruit juice and flavour milk shakes samples werediscussed. The use of ATR measurements in front of the useof transmission ones can be justified by the good signal tonoise ratio obtained in a wide range of wavenumbers, be-ing less affected by the presence of water than transmissionmeasurements.

Hierarchical cluster analysis has been used for selectingsamples for calibration dataset, using triplicate measure-ments distance as internal distance to compare how differ-ent the clusters are. This is an interesting attribute. Due tothe high heterogeneity of the sample population, the lowestroot-mean-square error of prediction was obtained when thenumber of samples extracted from each cluster for the cali-bration set equals the root square of the number of samplesin the cluster.

The cluster classification of sample population, fromthe mid-IR spectra is mainly related to the total amountof carbohydrates and fats, and therefore to the energeticvalue.

With regards to the application of the ATR correction al-gorithm, it has been shown that although it affect the dendro-gram, it does not affect the sample selection stage using theinternal cut-off distance provided by the replicate analysis ofsamples. Moreover, this correction has only a slight effect onthe prediction performance.

The performance of the developed methodologyf rksi osedm DAs

A

-nG -118a ofB

References

[1] B.G. Davis, A.J. Fairbanks, Carbohydrate Chemistry, Oxford Uni-versity Press, Oxford, UK, 2002.

[2] FDA (United States Food and Drug Administration) Nutrition La-beling Manual, 1998. http://www.fda.gov.

[3] P. Cunniff, Official Methods of Analysis of AOAC International,AOAC International, Washington, 1995.

[4] D.M. Sullivan, D.E. Carpenter, Methods of Analysis for NutritionLabeling, AOAC International, Washington, 1993.

[5] J. Tewari, M. Joshi, A. Gupta, R. Mehrotra, S. Chandra, J. Sci. Ind.Res. 58 (1999) 19.

[6] J. Irudayaraj, J. Tewari, Appl. Spectrosc. 57 (2003) 1599.[7] M. Gestal, M.P. Gomez-Carracedo, J.M. Andrade, J. Dorado, E. Fer-

nandez, D. Prada, A. Pazos, Anal. Chim. Acta 524 (2004) 225.[8] I.F. Duarte, A. Barros, I. Delgadillo, C. Almeida, A.M. Gil, J. Agric.

Food Chem. 50 (2002) 3104.[9] S. Sivakesava, J. Irudayaraj, Appl. Eng. Agric. 16 (2000) 543.

[10] F.A. Inon, S. Garrigues, M. de la Guardia, Anal. Chim. Acta 513(2004) 401.

[11] F.J. Rambla, S. Garrigues, N. Ferrer, M. de la Guardia, Analyst 123(1998) 277.

[12] S. Garrigues, F.J. Rambla, M. de la Guardia, Fresen. J. Anal. Chem.362 (1998) 137.

[13] R. Kellner, B. Lendl, I. Wells, P.J. Worsfold, Appl. Spectrosc. 51(1997) 227.

[14] F. Cadet, B. Offmann, J. Agric. Food Chem. 45 (1997) 166.[15] F. Cadet, Talanta 48 (1999) 867.[16] E.K. Kemsley, J.K. Holland, M. Defernez, R.H. Wilson, J. Agric.

Food Chem. 44 (1996) 3864.[ . 17

[ rdia,

[ 20.[ ab.

[[ ng,

[ nd-ganic

[[ la

[[

avourably compares with that obtained in previous won much restricted sample composition and the prop

ethods are suitable for complying with the US-Ftatuary values.

cknowledgments

The financial support of the Oficina de Ciencia I Tecologia de la Conselleria d’ Innovacio i Competivitat de laeneralitat Valenciana (Project GV 01-249, Grupos 03nd invited professor F.A. Inon grant) and the Universityuenos Aires (X013) is acknowledge.

17] S. Sivakesava, J.M.K. Irudayaraj, R.L. Korach, Appl. Eng. Agric(2001) 815.

18] F.A. Inon, J.M. Garrigues, S. Garrigues, A. Molina, M. de la GuaAnal. Chim. Acta 489 (2003) 59.

19] A. Lorber, K. Faber, B.R. Kowalski, Anal. Chem. 69 (1997) 1620] A.C. Olivieri, H.C. Goicoechea, F.A. Inon, Chemometr. Intell. L

Syst. 73 (2004) 189.21] D.M. Haaland, E.V. Thomas, Anal. Chem. 60 (1988) 1193.22] H.J. Luinge, E. Hop, E.T.G. Lutz, J.A. van Hemert, E.A.M. de Jo

Anal. Chim. Acta 284 (1993) 419.23] D. Lin-Vien, N.B. Colthup, W.B. Fateley, J.G. Graselli, The Ha

book of Infrared and Raman Characteristic Frequencies of OrMolecules, Academic Press, Boston, 1991.

24] S. Sasic, Y. Ozaki, Appl. Spectrosc. 55 (2001) 163.25] N. Blanco, V. Cerda, Quimiometrıa, Bellaterra: Publicacions de

Universitat Autonoma de Barcelona, Barcelona, 1988.26] K. Faber, B.R. Kowalski, J. Chemometr. 11 (1997) 181.27] A.C. Olivieri, J. Chemometr. 16 (2002) 207.