Qualitative and quantitative analysis of peanut adulteration in almond powder samples using multi-elemental fingerprinting combined with multivariate data analysis methods

Accepted Manuscript Qualitative and quantitative analysis of peanut adulteration in almond powder samples using multi-elemental fingerprinting combined with multivariate data analysis methods

Mahnaz Esteki, Yvan Vander Heyden, Bahman Farajmand, Yadollah Kolahderazi PII:

S0956-7135(17)30312-2

DOI:

10.1016/j.foodcont.2017.06.014

Reference:

JFCO 5667

To appear in:

Food Control

Received Date:

07 January 2017

Revised Date:

08 June 2017

Accepted Date:

11 June 2017

Please cite this article as: Mahnaz Esteki, Yvan Vander Heyden, Bahman Farajmand, Yadollah Kolahderazi, Qualitative and quantitative analysis of peanut adulteration in almond powder samples using multi-elemental fingerprinting combined with multivariate data analysis methods, Food Control (2017), doi: 10.1016/j.foodcont.2017.06.014

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

1

Qualitative and quantitative analysis of peanut adulteration in almond powder samples

2

using multi-elemental fingerprinting combined with multivariate data analysis methods

3 4

Mahnaz Esteki1,*, Yvan Vander Heyden2, Bahman Farajmand1, Yadollah Kolahderazi1

5

1Department

6

2Department

7

Pharmaceutical Research (CePhaR), Vrije Universiteit Brussel (VUB), Laarbeeklaan 103, B-

8

1090 Brussels, Belgium

of Chemistry, University of Zanjan, Zanjan 45195-313, Iran of Analytical Chemistry and Pharmaceutical Technology (FABI), Center for

9 10

Mahnaz Esteki (Corresponding author)

11

Department of Chemistry, University of Zanjan, Zanjan 45195-313, Iran

12

Phone: +982415152586

13

Fax: +982415152477

14

Email: [email protected]

15 16 17

Yvan Vander Heyden

18

Phone: +32 2 477 47 34

19

Fax: +32 2 477 47 35

20

Email: [email protected]

21 22 23

Bahman Farajmand

24

Department of Chemistry, University of Zanjan, Zanjan 45195-313, Iran 1

ACCEPTED MANUSCRIPT

25

Phone: +982415152586

26

Fax: +982415152477

27

Email: [email protected]

28 29 30

Yadollah Kolahderazi

31

Department of Chemistry, University of Zanjan, Zanjan 45195-313, Iran

32

Phone: +982433052586

33

Fax: +982433052477

34

Email: [email protected]

35 36 37 38 39 40 41 42 43 44 45

2

ACCEPTED MANUSCRIPT

46

Qualitative and quantitative analysis of peanut adulteration in almond powder samples

47

using multi-elemental fingerprinting combined with multivariate data analysis methods

48

Abstract

49

In this study, adulteration of almond powder samples with peanut was analysed using multi-

50

elemental fingerprinting based on inductively coupled plasma optical emission measurements

51

(ICP-OES) combined with chemometric methods. The ability of multivariate data analysis

52

approaches, such as principal component analysis (PCA) and principal component analysis-

53

linear discriminant analysis (PCA-LDA), to achieve differentiation of samples and as partial

54

least squares (PLS) and least squares support vector machine (LS-SVM), to quantify the

55

adulteration based on the elemental contents has been investigated. Ten variables i.e. the contents

56

of B, Na, Mg, K, Ca, Fe, Cu, Cu, Zn and Sr at µg g-1 level, determined by ICP-OES were used.

57

Different almond and peanut samples were then mixed at various ratios to obtain mixtures

58

ranging from 95/5 to 5/95 w/w and PCA-LDA was applied to classify the almonds, peanuts and

59

adulterated samples. This method was able to differentiate peanut and almond samples from the

60

adulterated samples. PLS and LS-SVM models were developed to quantify the adulteration

61

ratios of almond using a training set and the constructed models were evaluated using a

62

validation set. The root mean squared error of prediction (RMSEP) and the coefficient of

63

determination (R2) of the validation set for PLS and LS-SVM were 3.81, 0.986 and 1.66, 0.997,

64

respectively, which demonstrates the superiority of the LS-SVM model. The results show that

65

the combination of multi-elemental fingerprinting with multivariate data analysis methods can be

66

applied as an effective and feasible method for testing almond adulteration.

67 68

Keywords: adulteration, almond, peanut, multi-elemental fingerprinting, PCA-LDA, LS-SVM

3

ACCEPTED MANUSCRIPT

69 70

Introduction

71

The authentication of food products has been recognized as a worldwide topic of interest

72

covering many different aspects, from adulteration to mislabeling and misleading origin.

73

Economically motivated adulteration for economic gain of the seller is a process by which the

74

quality of a substance is reduced through the addition, substitution or removal of food

75

ingredients without the consumer's knowledge (Moore, Spink, & Lipp, 2012). Most food

76

products susceptible for fraud are high commercial cost products, often produced worldwide on a

77

large scale (Cordella, Moussa, Martel, Sbirrazzouli, & Lizzani-Cuvelier, 2002). Therefore, the

78

ability of the industry, governments, and standards-setting organizations to authenticate, to

79

control food constituents and to check for food fraud is increasingly important ( Ellis, Brewster,

80

Dunn, Allwood, Golovanov & Goodacre, 2012; Gupta & Panchal, 2009; Zhang, Zhang, Dediu,

81

& Victor, 2011).

82

Several methods have been proposed for the detection of the adulteration in food products, such

83

as PCR assay for meat adulteration ( Ali, Rahman, Hamid, Mustafa, Bhassu & Hashim, 2014;

84

Karabasanavar, Singh, Kumar, & Shebannavar, 2014), visible and near infrared hyperspectral

85

imaging for meat adulteration (Kamruzzaman, Makino, Oshita, & Liu, 2015), high performance

86

liquid chromatography with mass spectrometric detection for lemon juice adulteration (Wang &

87

Jablonski, 2016), solid-phase micro extraction, gas chromatography with mass spectrometry

88

detection (SPME-GC-MS) for cognac and brandy adulteration (Mozhayeva, Zhakupbekova,

89

Kenessov & Akmoldayeva, 2014), electrospray ionization mass spectrometry for meat

90

adulteration (Ruiz Orduna, Husby, Yang, Ghosh, & Beaudry, 2015), two-dimensional gas

91

chromatography for medicinal herb adulteration ( Welke, Damasceno, Nicolli, Mentz, Caramao,

92

Pulgatid, & Zini, 2015), an electronic nose for saffron adulteration (Heidarbeigi, Mohtasebi, 4

ACCEPTED MANUSCRIPT

93

Foroughirad, Ghasemi-Varnamkhasti, Rafiee & Rezaei, 2015),

94

attenuated total reflectance (Jiménez-Sotelo, Hernández-Martínez, Osorio-Revilla, Meza-

95

Márquez, García-Ochoa & Gallardo-Velázquez, 2016) and Raman spectroscopy for milk powder

96

characterization (Karunathilaka, Farris, Mossoba, Moore, & Yakes, 2016). Most methods are

97

time consuming, use large volumes of solvents and are not readily adaptable for rapid monitoring

98

through portable instrumentation. Meanwhile, spectroscopic techniques combined with

99

multivariate data analysis methods form a promising strategies to overcome the drawbacks and

100

can effectively be used to detect adulteration of different food products (López, Trullols, Callao,

101

& Ruisánchez, 2014).

102

Almond (Prunus amygdalus) with several unique features is one of the most popular nuts

103

worldwide. It is highly nutritious and classified as a drupe in which the edible seed is a

104

commercial product (Alonso, Kodad, & Gradziel, 2012). Almond powder is generally used in a

105

variety of processed foods, particularly in bakery and confectionery products (Dourado, Barros,

106

Mota, Coimbra, & Gama, 2004). Because of its high price, almond powder is a target of illegal

107

practices, such as mixing with cheaper nuts. One of the most common adulterations consists of

108

using peanut as an adulterant with very similar chemical composition and much lower prices. To

109

the best of our knowledge, adulteration studies for this kind of nut are not extended in the

110

literature. Thus, it is important to develop analytical procedures to verify the quality of almond

111

powder, identifying peanut fraud, motivated primarily by economic gain.

112

Food products are also usually tested for metal contents for a variety of reasons. Determination

113

of the mineral composition is an important approach reflecting the nutritional value and its

114

relationship with food quality. In fact, some elements are essential for numerous bodily functions

115

and have a metabolic role (Co, Cu, Fe, Se, Zn), while others have potentially toxic characteristics

5

Fourier transform infrared

ACCEPTED MANUSCRIPT

116

(As, Cd, Pb). Therefore, trace metal profiling can be used to authenticate food Chen, Fan, Chang,

117

Pang, Hu, Lu & Wang, 2014; D’Archivio, Giannitto, Incani, & Nisi, 2014; Drivelos, Higgins,

118

Kalivas, Haroutounian, & Georgiou, 2014; Laursen, Schjoerring, Kelly, & Husted, 2014) and it

119

is also expected that food adulteration could change the elemental profile of a particular sample,

120

such as almond.

121

Fingerprinting based on chemical composition and multivariate data analysis have become one

122

of the most powerful systematic approaches to determine authenticity (Coetzee, Van Jaarsveld,

123

& Vanhaecke, 2014). A fingerprint is a characteristic profile of a sample, which can be

124

established through common techniques such as chromatography and spectroscopy to

125

characterize food products based on their quality or origin (Gan, Yang, Li, Wen, Zhu, Jiang &

126

Ni, 2015). The most common procedures used for pattern recognition purposes include: (i)

127

principal component analysis (PCA) as an unsupervised technique, which provides uncorrelated

128

objective latent variables (principal components) capable of extracting valuable information from

129

the experimental data, exploring the relationship between objects in addition to the relationship

130

between variables, and between objects and variables, (ii) linear discriminant analysis (LDA) as

131

a supervised pattern recognition technique which is based on the determination of linear

132

discriminant functions through maximizing the variance between classes and minimizing the

133

variance within the classes.

134

In this paper, multi-element fingerprints of almond and peanut samples were determined by

135

inductively coupled plasma optical emission spectrometry (ICP-OES) and subsequently two

136

methodologies were proposed for multivariate classification. First, PCA was used to investigate

137

the ability of elemental fingerprinting to discriminate the samples and then LDA was applied to

138

detection of pure and adulterated almond samples. The second main objective of this study was

6

ACCEPTED MANUSCRIPT

139

to build a reliable model using elemental fingerprinting coupled with multivariate calibration

140

methods for the quantification of peanut adulteration in almond powder samples. In this context,

141

we have focused on the development and validation of calibration models using PLS and LS-

142

SVM for the quantitative determination of adulteration in binary mixtures of peanut/almond

143

powder samples in the concentration range of 5–95% (w/w).

144 145

Materials and method

146

Reagents and samples

147

Analytical reagent-grade materials were used for all experiments. All solutions were prepared

148

using high-purity deionized water (>18MΩ) (TKASmart2pure Water Purification Systems,

149

Niederelbert, Germany). Nitric acid 65% w.v−1 and reference solutions of Al, B, Ba, Be, Bi, Ca,

150

Cd, Co, Cr, Cu, Fe, Ga, K, Li, Mg, Mn, Na, Ni, Pb, Se, Te and Zn at 1000 mg L−1 were obtained

151

from Merck (Darmstadt, Germany).

152

Due to the large variation in available almonds and peanuts as a result of flavoring and

153

processing, only raw, unflavored samples were used to minimize the sample variability. Almond

154

samples were collected from Shahr-e-kord, which is the main producer of almonds with the

155

highest quality in Iran. Peanut sampling was performed from Astaneh-ye-ashrafiyeh orchards.

156

This region covers more than 80% of the peanut production in Iran. All samples were collected

157

during the harvesting periods of 2013–2014. The adulterated samples were prepared as different

158

ratios of almond-peanut mixtures. Almonds and peanuts were oven dried at 30°C for at least 5

159

days. Afterwards, the samples were stored in a refrigerator at 4 °C till the preparation process.

160 161

Sample preparation and digestion 7

ACCEPTED MANUSCRIPT

162

The raw unshelled samples were powdered using a mortar and pestle and were used for the

163

preparation of the pure and the adulterated samples (ranging from 95/5 to 5/95 w/w). A total of

164

150 samples was measured, including 25 pure almond samples, 25 pure peanut samples and 100

165

adulterated almond samples. Almond powder was adulterated with peanut powder at a variety of

166

levels from 5 to 95%, w/w. The finely ground kernels (1.0 g) (pure or adulterated with specific

167

weight fraction of peanut) were introduced into a 50-mL PTFE (polytetrafluoroethylene) closed

168

vessel. A volume of 20.0 mL concentrated nitric acid and 5 mL of concentrated hydrogen

169

peroxide were added to the vessel. The bomb was sealed and put in a furnace set at 180 ± 10 ◦C

170

and remained at this temperature during 2h. The addition of H2O2 to the nitric acid was necessary

171

to increase the oxidation efficiency. After cooling down to room temperature, solutions were

172

quantitatively transferred to glass volumetric flasks and volumes were made up to 50.0 mL with

173

deionized water.

174 175

Analysis by ICP-OES

176

An ICP-OES (Perkin Elmer, Optima 7300 DV, Shelton, USA) was used for the Al, B, Ba, Be,

177

Bi, Ca, Cd, Co, Cr, Cu, Fe, Ga, K, Li, Mg, Mn, Na, Ni, Pb, Se, Te and Zn determination. The

178

sample introduction system was composed of a Scott spray chamber and a Gem-cone nebulizer.

179

The operational parameters are tabulated in Table 1. A multi-element calibration curve was

180

prepared by diluting standard solutions at 100 mg L−1 of each metal (CertiPur, Merck). The

181

instrument is operated with computer software WinLab32. All measurements were performed in

182

triplicate.

183

Multivariate data analysis

8

ACCEPTED MANUSCRIPT

184

Different multivariate analysis techniques were applied on the ICP-OES data. Pure almond

185

samples (25 samples) and pure peanut samples (25 samples) were used for PCA and PCA-LDA

186

analysis in order to investigate the ability of the elemental composition to differentiate both. Half

187

of the adulterated samples (50 samples) were also used for classification purposes using PCA-

188

LDA, while the remaining adulterated samples (50 samples) were used to evaluate the

189

classification accuracy of PCA-LDA model.

190

In order to perform the quantification, the adulterated almond samples data (100 samples) were

191

randomly divided into calibration (50 samples) and prediction (evaluation) set (50 samples). The

192

calibration samples were used for model construction (applying PLS and LS-SVM), while the

193

prediction samples were utilized to evaluate the model performance.

194

The LDA and PLS models were in-house programmed (Mousavi, Esteki, Mostafazadeh-Fard,

195

Dehghani, & Khorvash, 2012) using MATLAB (the MathWorks, Version 6.0, Natick, MA) and

196

the SVM software package ChemSVM, including SVR, was programmed by Suykens ( 2001).

197

This software has been validated in previous works (Esteki, Rezayat, Ghaziaskar, & Khayamian,

198

2010).

199 200

Linear Discriminant analysis (LDA)

201

Linear Discriminant analysis (LDA) is a supervised pattern recognition technique which is based

202

on maximizing the variance between the groups and minimizing the variance within them, by

203

means of a data projection from a high dimensional space to a low dimensional space. The

204

transformation tends to pivot the space in the way which maximizes the differences between the

205

groups when the observations are projected on to the new space (Varmuza & Filzmoser, 2010).

9

ACCEPTED MANUSCRIPT

206

order In to achieve a binary classification (two classes), suppose two classes with n-dimensional

207

samples x1, …, xn , where n1 and n2 represent the number of samples of the first and the second

208

class, respectively. The Fisher discriminant can be defined as the linear function 𝑽𝒕𝒙 that

209

maximizes the principle function: (𝝁𝟏 ‒ 𝝁𝟐)𝟐

210

𝑱(𝑽) =

(4) where 𝜇1 and 𝜇2

211

indicate the means of the projected points of classes 1 and 2, and 𝑠1 and 𝑠2 the scatter in classes 1

212

and 2, respectively. Now, the within-class scatter matrix is defined as:

213

𝑺𝒘 = 𝑺𝟏 + 𝑺𝟐

214

In this case, the separate class scatter matrices, 𝑺𝟏 and 𝑺𝟐, for classes 1 and 2 can be described

215

as:

𝒔𝟏𝟐 + 𝒔𝟐𝟐

∑

𝑺𝟏 =

(5)

(𝒙𝒊 ‒ 𝝁𝟏)(𝒙𝒊 ‒ 𝝁𝟏)𝒕

(6)

(𝒙𝒊 ‒ 𝝁𝟐)(𝒙𝒊 ‒ 𝝁𝟐)𝒕

(7)

𝒙𝒊𝝐𝑪𝒍𝒂𝒔𝒔 𝟏

216

∑

𝑺𝟐 =

𝒙𝒊𝝐𝑪𝒍𝒂𝒔𝒔 𝟐

217

The scatter of the projection y can then be defined as a function of the scatter matrix in feature

218

space x:

219

𝒔𝟏𝟐 = 𝑽𝒕𝑺𝟏𝑽

220

Considering the fact that 𝒔𝟐𝟐 = 𝑽𝒕𝑺𝟐𝑽, and using eq. 5, the sum can be written as

221

𝒔𝟐𝟏 + 𝒔𝟐𝟐 = 𝑽𝒕𝑺𝑾𝑽

222

The between class scatter matrix can be defined as:

223

𝑺𝑩 = (𝝁𝟏 ‒ 𝝁𝟐)(𝝁𝟏 ‒ 𝝁𝟐)𝒕

(8)

(9)

(10)

10

ACCEPTED MANUSCRIPT

224

Similarly, the difference between the projected means (in y-space) can be represented in terms of

225

the means in the original feature space (x-space):

226

(𝝁𝟏 ‒

𝝁𝟐)𝟐 = 𝑽𝒕𝑺𝑩𝑽

(11)

227 228

Finally the Fisher criterion (eq. 4) can be expressed in terms of 𝑆𝑊 and 𝑆𝐵 as:

229

𝑱(𝑽) =

230

the

𝑽𝒕𝑺𝑩𝑽

(12) In order to find

𝑽𝒕𝑺𝑾𝑽

maximum

of 𝑱(𝑽),

the

derivative

of

𝑱(𝑽)

should

be

equal

to

zero:

231 232

𝒅 𝑱(𝑽) = 𝟎 𝒅𝑽

233

generalized eigen value problem:

234

𝑺 ‒𝑾𝟏𝑺𝑩𝑽 = 𝝀𝑽

235

Where 𝝀 = 𝑱(𝑽)

236

It can be shown that for c classes, the most favorable projection matrix 𝑽 is the one whose

237

columns are the eigenvectors corresponding to the largest eigen values of the generalized eigen

238

value problem (eq. 13).

(13) To solve the

(14)

239 240

Partial least squares (PLS)

241

Partial least squares (PLS) is a multivariate statistical linear regression technique, based on

242

estimated latent variables which extracts the relationship between dependent variables (output

243

array) and independent variables (input array). The aim of the PLS regression is to build a linear

244

model enabling the prediction of a desired characteristic (Y) from a measured data matrix (X). 11

ACCEPTED MANUSCRIPT

245

In this method, dimension reduction of the raw data is based on the independent variable matrix

246

(X) as well as on the dependent variable matrix (Y). The key point of the method is the

247

simultaneous decomposition of X and Y which can be performed as follows (Geladi, 1988):

248

𝑿 = 𝑻𝑷' + 𝑫

(15)

249

𝒀 = 𝑼𝑸' + 𝑭

(16)

250

where T and U are the X- and Y-block score matrices; P and Q are the X and Y loadings; and D

251

and F are the residuals. The construction of a PLS model involves the creation of a relationship

252

between projections of the input and output variables, U and T, respectively, as below:

253

𝑼 = 𝑩𝑻

254

In other words, through the PLS modeling process the regression model is build between scores

255

of the input variables and scores of the output variables.

(17)

256 257

Support Vector Machine (SVM)

258

Support vector machine (SVM) was introduced by Vapnik ( 1998) and has been the focus of

259

many studies in recent years in a wide range of applications. To apply SVM in a regression

260

problem one should find the optimal hyperplane, in such a way that the distances between the

261

data points and the hyperplane are minimal (Smola & Schölkopf, 2004). The next paragraphs

262

describe the basic fundamental aspects of the LS-SVM modeling.

263

G  ( x i , y i ) iN1 represents a data set including N data points where each independent variable,

264

xi , should be linked with the corresponding dependent variable, y i . Suppose that the data set

265

is generated by an unknown function g (x) , then a function f that approximates g (x) should

266

be build based on the knowledge of data set G . Therefore, xi is first mapped into a higher 12

ACCEPTED MANUSCRIPT

267

dimensional space, F , by using a nonlinear mapping and then making a linear regression in

268

this space. The SVM approximates the function as: n f ( x)    i  i ( x)  b with  :   F ,   F

269

(18)

i 1

270

where  i are coefficients and b is the value of threshold. This approximation can be

271

assumed as a hyperplane in the feature space F (with D-dimensions) which can be identified

272

by the functions  i (x) . Since  is fixed, the coefficients (  i ) can be obtained from the data

273

by minimizing the sum of the empirical risk and a complexity term, which have been defined

274

in the following risk function:

R 

275

1 N

N

y i 1

i

 f ( xi )  

1  2

2

(19)

276

where  is a parameter that has to be determined based on earlier information, so error

277

values less than  may be ignored, as it is shown in the following error function.

278

0 if yi  f ( xi )   yi  f ( xi )    otherwise.  yi  f ( xi )

279

(20)

280 281

The first term in Eq. (19) describes the loss function and the second is related to the function

282

flatness. The SVMs construct a linear regression in a high-dimensional feature space by

283

using the loss function and reduce model complexity by minimizing  . The regularization

284

parameter (  ) is a constant which determines the trade-off between the model flatness and

285

the calibration error. Then the SVM regression model is build by minimizing the following

286

term:

2

13

ACCEPTED MANUSCRIPT

287 288

N 1 2     ( i   i* ) 2 i 1

(21)

289

290

 f ( xi )  b  y i     i  *  y i  f ( xi )  b     i Subject to   i*  0  i ,

(22)

291 292

The above optimization problem could be solved by constructing a Lagrangian function and

293

transforming it into the dual problem (Vapnik, 1998):

294 N

f ( x,  ,  )   ( i*   i ) K ( x,xi )  b *

295

(23)

i 1

296

where  i and  i* are the Lagrange multipliers, related to each input data point xi , and which

297

are subject to the following constraints:

298

0   i* ,  i   and

N

 ( i 1

* i

i )  0

299

Support vectors are actually the training points with non-zero Lagrange multipliers and K (.)

300

is the kernel function describing an inner product in the D-dimensional space as identified

301

below:

302 D

303

K ( x, xi )    i ( x) i ( xi )

(24)

i 1

304 14

ACCEPTED MANUSCRIPT

305

The Lagrange coefficients  ,  * are obtained by maximizing the following term subject to

306

the constraints stated earlier:

307 N

N

i 1

i 1

R( ,  * )   yi ( i*   i )    ( i*   i ) 

308

1 N ( i*   i )( *j   j ) K ( xi , x j )  2 i , j 1

(25)

309

Whenever the coefficients are determined, the regression estimate is given by Eq. (23). The

310

value of threshold b can be calculated from the constraints in Eq. (21) considering the fact

311

that the first constraint becomes an equality with  i  0

312

constraint becomes an equality with  i* if 0   i*   . The generalization performance is

313

influenced by the parameters  ,  and the kernel type. Linear, radial basis function (RBF),

314

and polynomial kernels are the most common kernel types. The radial basis function,

315

exp  xi  x j

316

Gaussian function. In order to obtain the support vectors,  2 and  should be optimized.



2

/ 2 2

if 0   i   , and the second

 is the most frequently used kernel function with 

2

the width of a

317 318

Results and discussion

319

To develop a robust model for detection and quantification of adulteration, the concentrations of

320

different elements were selected as variables. Analytical characteristics of the employed ICP-

321

OES procedure after the sample digestion via a Teflon lined stainless steel autoclave are

322

tabulated in Table 2. The limit of detection (LOD) was determined as the lowest concentration of

323

analyte that produces a signal equal to three times the standard deviation of ten replicates of the

324

blank solution. Additionally, LODs were calculated in the original samples (µg g-1) taking into

325

account the amount of sample and the volume of solvents employed during the procedure. The

326

LOD values of the ICP-OES are satisfactory for determination of the elements in almond and 15

ACCEPTED MANUSCRIPT

327

peanut samples, ranging from 0.3 µg L-1 (0.015 µg g-1) to 103 µg L-1 (5.15 µg g-1) for Be and Sr,

328

respectively. Relative standard deviation (%RSD) values were also calculated for five

329

independent analyses of multi-elemental standard solution to characterize the repeatability of the

330

method; obtaining RSD values ranging from 0.03% for Ni to 3.29% for Te.

331

Table 3 summarizes the mean concentration values (µg g-1) of the element compositions found in

332

the almond and peanut samples. From the initial 22 elements investigated only ten, including B,

333

Na, Mg, K, Ca, Fe, Cu, Cu, Zn and Sr, are presented in detectable concentrations in peanut and

334

almond samples.

335 336 337

Differentiation of pure and adulterated almond samples with class-modeling methods

338

Principal component analysis (PCA) is the best known variable-reduction method which is

339

widely used in chemometrics. PCA decomposes the data matrix with m rows (samples) and p

340

columns (variables) into the product of a score and a loading matrix. The scores are the positions

341

of the samples in the space of the principal components (PCs), which gives information about the

342

similarity of samples. All PCs are orthogonal and each successive PC contains a smaller fraction

343

of the total variability of the data set. Principal component analysis was applied to provide an

344

overview of the ability of the elemental composition data to characterize almond and peanut

345

samples. The data set consisted of 50 samples, including 25 pure almond samples and 25 pure

346

peanut samples. Auto-scaling was applied as pretreatment technique. By using the first two

347

principal components, the total explained variation is equal to 97.96% (PC1 = 92.49%, PC2 =

348

5.47%) indicating that the amount of information which would be lost taking into account the

349

first two PCs, is not considerable. The scores for the first two PCs are plotted as a scatter

16

ACCEPTED MANUSCRIPT

350

diagram in Fig. 1a. It is clear that two distinct clusters are formed, corresponding to the peanut

351

and almond samples which indicates that the multi-elemental fingerprinting is able to

352

discriminate almond and peanut samples.

353

Since the data structure analysis using PCA gave a good sample characterization, PCA-LDA

354

analysis was applied in order to find a predictive classification model, which would be able to

355

separate the three described classes i.e. almond, peanut and adulterated samples. A PCA-LDA

356

model was constructed using 100 samples (25 pure almond, 25 pure peanut, 50 adulterated

357

samples) as calibration set for training. Classification was evaluated using the cross-validation

358

(leave-one-out) method. The ability of distinguishing adulterated samples by the LDA

359

classification model was considered based on the validation set consisting of 50 adulterated

360

samples. The scatter plot of projected samples of the second discriminate function vs. the first

361

shows a clear class separation (Fig. 1b). The pure almond group showed negative values in the

362

first dimension and positive in the second; the pure peanut had high positive values in the first

363

dimension and negative in the second, while the adulterated points are distributed around zero in

364

both dimensions. The almond samples formed a distinct group on the left side of the plot. The

365

peanut and adulterated group showed a significant difference in the first dimension but did not

366

show a big difference in the second. The relative contributions of the first and second

367

discriminant functions to the total variance are 81.2% and 16.2%, respectively. The eigenvalues

368

also showed that the first discriminant function is more effective for classifying the samples.

369

The likelihood of the model was evaluated using cross-validation of the training set and also by

370

diagnosing the test set that has not used to build the model. Finally percentages of correct

371

classification were calculated. In Table 4, the results of LDA classification are reported. The

372

LDA model, correctly classified 100% of the calibration, 100% of the cross-validated cases and

17

ACCEPTED MANUSCRIPT

373

the recognition percentage was 98% for the validation set with only one wrong assignment. This

374

means that PCA-LDA is an efficient technique for the qualitative analysis of almond samples

375

and distinguishing adulterated almond samples from the pure ones.

376 377

Quantitative determination of adulteration with PLS and LS-SVM

378

Besides the detection of almond samples adulterated with peanut, the other important problem

379

focuses on quantitative determination of peanut in adulterated almond samples.

380

The investigated range of relative adulterant concentrations in binary mixtures was 5–95 %

381

weight ratios of adulteration. In this study, auto-scaling was applied as the pretreatment

382

technique. The modeling and prediction of mixture composition have been done in a first

383

approach by partial least squares (PLS) regression, which is a linear calibration method widely

384

used in the chemometric literature for multivariate calibration. In this work, the elemental

385

compositions measured by ICP-OES were used for constructing the calibration matrix. Fifty of

386

the 100 adulterated almond samples were randomly selected as calibration set and the remaining

387

50 were used for validation. To select the optimal number of factors for the PLS models, the

388

leave-one-out cross-validation method was employed. The prediction error (RMSE) on the

389

amount of peanut in the samples of calibration and prediction sets was estimated as follows:

390

1 𝑅𝑀𝑆𝐸 = 𝑛

𝑛

∑ (𝐶 ‒ 𝐶 ) 𝑖

𝑖

2

𝑖=1

391

where 𝐶𝑖 is the percentage of the peanut in the ith sample, 𝐶𝑖 is the predicted percentage and n is

392

the number of samples considered. RMSE values were calculated for calibration, cross-

393

validation and prediction sets. The number of latent variables was subsequently optimized by

394

considering a compromise between the root mean squared error of calibration (RMSEC), which 18

ACCEPTED MANUSCRIPT

395

represents the fit of the model to the calibration points, and the root mean squared error of cross-

396

validation (RMSECV), to prevent overfitting (Fig. 2a). The PLS model selected used two latent

397

variables, resulting in an RMSEC of 1.30% (w/w), and RMSECV of 2.15%, which demonstrates

398

the adequacy of the model. The scatter plot of the actual peanut concentration as adulterant

399

versus the predicted peanut content for both the calibration and the test sets are illustrated in Fig.

400

2 (b,c). The determination coefficient (R2) of 0.995 obtained for the calibration set, as shown in

401

Fig. 2b, reflects the generally adequate fit of the PLS regression model. The data from the

402

external validation set consisting of fifty samples were analyzed using the model that was

403

obtained with the training set and the accuracy of the model was evaluated based on the root

404

mean square error of prediction (RMSEP) and determination coefficient of the validation set.

405

The predicted values of adulteration with peanut vs. the experimental ones are shown in Fig. 2c.

406

The numerical values of the statistical parameters of the PLS model are listed in Table 5. An

407

RMSEP of 3.81 (w/w) and R2 value of 0.986 were obtained for the validation set. The overall

408

results show that the proposed method has a suitable reliability for estimating the concentration

409

of peanut adulterant in almond samples.

410

In comparison to the linear model, a nonlinear approach was performed. LS-SVM was used in

411

this work as a nonlinear modeling method. The basic principle of LS-SVM regression is

412

mapping the nonlinearity of the input training data onto a higher dimensional linear feature space

413

via the mapping function. In this study, the data of the validation and the calibration sets are the

414

same as for PLS model building. Optimization of the parameters is an important issue in building

415

an LS-SVM model. Kernel selection and setting appropriate kernel parameters can greatly

416

improve the SVM regression performance. In this study, the radial basis function (RBF) was

417

used as the kernel function. The RBF kernel is often used for regression analysis because of its

19

ACCEPTED MANUSCRIPT

418

good generalization performance, effectiveness and speed in the training process (Janik & Lobos,

419

2006). This function is also able to model nonlinear relationships between the independent

420

parameter and the response variables (Qu, Shih, Jing, & Wang, 2003).

421

Furthermore, variations of the hyper-parameters, including the regularization parameter, γ, and

422

the kernel parameter, σ2, have influence on model performance. Grid search, random search

423

(Bergstra & Bengio, 2012) and genetic algorithms (Wu, Tzeng, Goo, & Fang, 2007) are the most

424

widely used strategies for hyper-parameter optimization. Grid search is a reliable and efficient

425

global optimization method which was used in this study. A robust model was attained by

426

selecting parameters that give the lowest RMSEC error. This grid search was performed on the

427

original training data set. To find the optimal values of the parameters, a grid search was

428

performed for different combinations of γ and σ2 parameters and then the levels with smallest

429

values of RMSECV were selected as the optimum.

430

Fig. 3(a) illustrates the optimization process. The first searching step is a 10×10 grid (.) and is

431

considered as large steps, while in the second search, the grids (×) is again 10×10 and the

432

searching step is smaller. Error contour lines determine the most favorable search area. The

433

results indicated that an LS-SVM with 𝛾 ≃ 860.71, and 𝜎2 ≃ 165.04 resulted in the best LS-

434

SVM implementation. The nonlinear regression method was trained using the objects in the

435

training set and it was evaluated by the test set data. Fig. 3 (b,c) shows the plots of observed

436

versus predicted values for the training and the test sets. The data are distributed around the

437

bisecting line with a determination coefficient of 0.999 for the calibration set and 0.997 for the

438

test set. The statistical parameters, such as R2 between the predicted and experimental values

439

obtained using the LS-SVM for training and test set also are listed in Table 5. The RMSEC and

440

RMSECV values reveal the good performance of the LS-SVM model for prediction of the

20

ACCEPTED MANUSCRIPT

441

adulteration ratio. The high value of R2 cross-validation, 0.996, and the low value of RMSECV

442

of 0.92, as compared with the RMSEC of 0.54 and RMSEP of 1.66, suggest a good internal

443

consistency and the high-quality predictive capability of the LS-SVM model. The estimated

444

adulteration ratios are in excellent quantitative agreement with the experimental values (Fig. 3c).

445

The values and statistical parameters shown in Table 5 allow a comparison between both

446

methods. According to the calibration and cross-validation statistics and the prediction

447

performance, the LS-SVM method is superior to the PLS model.

448 449

Effect of the variables

450

In order to investigate the effect of the different elements in differentiating the samples, the LS-

451

SVM and PLS models were evaluated. The models were constructed using ten variables

452

including B, Na, Mg, K, Ca, Fe, Cu, Cu, Zn and Sr, and then, the effect of each variable was

453

evaluated by omitting it from the model. The RMSECV was calculated for the constructed

454

models. Fig. 4 shows the results of this process. For both LS-SVM and PLS models, the

455

maximum increment of RMSECV was due to Ca, Mg and K followed by Na and B. This means

456

that Ca, Mg and K are the most effective parameters in differentiating almond and peanut

457

samples, while Cu, Zn, Mn and Sr have relatively little effect in this process.

458 459

Conclusion

460

This study has shown the ability of multi-elemental fingerprinting coupled with chemometric

461

methods to qualitatively and quantitatively analyze peanut adulteration in almond powder

462

samples. PCA has been performed using the contents of B, Na, Mg, K, Ca, Fe, Cu, Cu, Zn and

463

Sr (in µg g-1) as variables, in order to investigate the discrimination ability of multi-elemental

21

ACCEPTED MANUSCRIPT

464

fingerprinting for differentiation of the peanut and almond samples. On the other hand, PCA-

465

LDA discrimination analysis allows separating the almond, peanut and adulterated almond

466

samples. The first two weight vectors were used to show the ability of LDA to group three

467

classes. The adulteration of almond was detected using the PCA-LDA method showing 100%

468

correct classification for both calibration and cross-validation and 98% accuracy for validation

469

set samples. Whenever adulteration has been recognized, quantitative analysis of peanut

470

adulterant was carried out by PLS and LS-SVM models. A determination coefficient of 0.995 for

471

training and 0.986 for validation set, RMSEC, 1.30, and RMSEP, 3.81, were achieved by the

472

PLS model. The LS-SVM was also applied as a non-linear regression method to determine the

473

level of adulteration in mixture samples. The adulteration ratios were determined successfully in

474

the range of 5-95% (w/w) with higher determination coefficients (R2 calibration, 0.999, and

475

validation, 0.997) and lower RMSE values (RMSEC, 0.54, and RMSEP, 1.66) when using the

476

LS-SVM model.

477

The proposed multi-elemental fingerprinting method combined with multivariate analysis

478

methods is a straightforward and effective tool for the determination and identification of

479

adulteration of almond powder with peanut.

480

In order to be able to verify the authenticity of almond powders originating from different areas

481

or the food industry by using the above-discussed sensitive, easy and rapid method, we will

482

develop the models for a wider variety of almond and peanut samples, covering the relevant

483

geographic origins and harvesting times.

484 485

Acknowledgement

486

The authors gratefully acknowledge the University of Zanjan for financial support of this work.

22

ACCEPTED MANUSCRIPT

487 488 489 490 491

Ali, M. E., Rahman, M. M., Hamid, S. B. A., Mustafa, S., Bhassu, S., & Hashim, U. (2014).

492

Canine-Specific PCR Assay Targeting Cytochrome b Gene for the Detection of Dog Meat

493

Adulteration in Commercial Frankfurters. Food Analytical Methods, 7(1), 234–241.

494

http://doi.org/10.1007/s12161-013-9672-y

495 496 497

Bergstra, J., & Bengio, Y. (2012). Random Search for Hyper-Parameter Optimization. Journal of Machine Learning Research, 13, 281–305. Chen, H., Fan, C., Chang, Q., Pang, G., Hu, X., Lu, M., & Wang, W. (2014). Chemometric

498

determination of the botanical origin for Chinese honeys on the basis of mineral elements

499

determined by ICP-MS. Journal of Agricultural and Food Chemistry, 62(11), 2443–2448.

500

http://doi.org/10.1021/jf405045q

501

Coetzee, P. P., Van Jaarsveld, F. P., & Vanhaecke, F. (2014). Intraregional classification of wine

502

via ICP-MS elemental fingerprinting. Food Chemistry, 164, 485–492.

503

http://doi.org/10.1016/j.foodchem.2014.05.027

504

Cordella, C., Moussa, I., Martel, A. C., Sbirrazzouli, N., & Lizzani-Cuvelier, L. (2002). Recent

505

developments in food characterization and adulteration detection: Technique-oriented

506

perspectives. Journal of Agricultural and Food Chemistry. http://doi.org/10.1021/jf011096z

507

D’Archivio, A. A., Giannitto, A., Incani, A., & Nisi, S. (2014). Analysis of the mineral

508

composition of Italian saffron by ICP-MS and classification of geographical origin. Food

509

Chemistry, 157, 485–489. http://doi.org/10.1016/j.foodchem.2014.02.068

23

ACCEPTED MANUSCRIPT

510

Dourado, F., Barros, A., Mota, M., Coimbra, M. a, & Gama, F. M. (2004). Anatomy and cell

511

wall polysaccharides of almond (Prunus dulcis D. A. Webb) seeds. Journal of Agricultural

512

and Food Chemistry, 52(5), 1364–1370. http://doi.org/10.1021/jf030061r

513

Drivelos, S. A., Higgins, K., Kalivas, J. H., Haroutounian, S. A., & Georgiou, C. A. (2014). Data

514

fusion for food authentication. Combining rare earth elements and trace metals to

515

discriminate “fava Santorinis” from other yellow split peas using chemometric tools. Food

516

Chemistry, 165, 316–322. http://doi.org/10.1016/j.foodchem.2014.03.083

517

Ellis, D. I., Brewster, V. L., Dunn, W. B., Allwood, J. W., Golovanov, A. P., & Goodacre, R.

518

(2012). Fingerprinting food: current technologies for the detection of food adulteration and

519

contamination. Chemical Society Reviews, 41(17), 5706. http://doi.org/10.1039/c2cs35138b

520

Esteki, M., Rezayat, M., Ghaziaskar, H. S., & Khayamian, T. (2010). Application of QSPR for

521

prediction of percent conversion of esterification reactions in supercritical carbon dioxide

522

using least squares support vector regression. Journal of Supercritical Fluids, 54(2), 222–

523

230. http://doi.org/10.1016/j.supflu.2010.04.007

524

Gan, Z., Yang, Y., Li, J., Wen, X., Zhu, M., Jiang, Y., & Ni, Y. (2015). Using sensor and

525

spectral analysis to classify botanical origin and determine adulteration of raw honey.

526

Journal of Food Engineering. http://doi.org/10.1016/j.jfoodeng.2016.01.016

527

Geladi, P. (1988). Notes on the history and nature of partial least squares (PLS) modelling.

528

Journal of Chemometrics, 2(January), 231–246.

529

http://doi.org/http://doi.wiley.com/10.1002/cem.1180020403

530

Gupta, N., & Panchal, P. (2009). Extent of awareness and food adulteration detection in selected

531

food items purchased by home makers. Pakistan Journal of Nutrition.

532

http://doi.org/10.3923/pjn.2009.660.667

24

ACCEPTED MANUSCRIPT

533

Heidarbeigi, K., Mohtasebi, S. S., Foroughirad, A., Ghasemi-Varnamkhasti, M., Rafiee, S., &

534

Rezaei, K. (2015). Detection of adulteration in saffron samples using electronic nose.

535

International Journal of Food Properties, 18(7), 1391–1401.

536

http://doi.org/10.1080/10942912.2014.915850

537

Janik, P., & Lobos, T. (2006). Automated classification of power-quality disturbances using

538

SVM and RBF networks. IEEE Transactions on Power Delivery, 21(3), 1663–1669.

539

http://doi.org/10.1109/TPWRD.2006.874114

540

Jiménez-Sotelo, P., Hernández-Martínez, M., Osorio-Revilla, G., Meza-Márquez, O. G., García-

541

Ochoa, F., & Gallardo-Velázquez, T. (2016). Use of ATR-FTIR spectroscopy coupled with

542

chemometrics for the authentication of avocado oil in ternary mixtures with sunflower and

543

soybean oils. Food Additives & Contaminants: Part A, 33(7), 1105–1115.

544

http://doi.org/10.1080/19440049.2016.1203073

545

Kamruzzaman, M., Makino, Y., Oshita, S., & Liu, S. (2015). Assessment of Visible Near-

546

Infrared Hyperspectral Imaging as a Tool for Detection of Horsemeat Adulteration in

547

Minced Beef. Food and Bioprocess Technology, 8(5), 1054–1062.

548

http://doi.org/10.1007/s11947-015-1470-7

549

Karabasanavar, N. S., Singh, S. P., Kumar, D., & Shebannavar, S. N. (2014). Detection of pork

550

adulteration by highly-specific PCR assay of mitochondrial D-loop. Food Chemistry (Vol.

551

145). Elsevier Ltd. http://doi.org/10.1016/j.foodchem.2013.08.084

552

Karunathilaka, S. R., Farris, S., Mossoba, M. M., Moore, J. C., & Yakes, B. J. (2016).

553

Characterising variances of milk powder and instrumentation for the development of a non-

554

targeted, Raman spectroscopy and chemometrics detection method for the evaluation of

555

authenticity. Food Additives & Contaminants: Part A, 33(6), 921–932.

25

ACCEPTED MANUSCRIPT

556 557

http://doi.org/10.1080/19440049.2016.1188437 Laursen, K. H., Schjoerring, J. K., Kelly, S. D., & Husted, S. (2014). Authentication of

558

organically grown plants - advantages and limitations of atomic spectroscopy for multi-

559

element and stable isotope analysis. TrAC - Trends in Analytical Chemistry.

560

http://doi.org/10.1016/j.trac.2014.04.008

561

López, M. I., Trullols, E., Callao, M. P., & Ruisánchez, I. (2014). Multivariate screening in food

562

adulteration: Untargeted versus targeted modelling. Food Chemistry, 147, 177–181.

563

http://doi.org/10.1016/j.foodchem.2013.09.139

564

Moore, J. C., Spink, J., & Lipp, M. (2012). Development and Application of a Database of Food

565

Ingredient Fraud and Economically Motivated Adulteration from 1980 to 2010. Journal of

566

Food Science. http://doi.org/10.1111/j.1750-3841.2012.02657.x

567

Mousavi, S.-F., Esteki, M., Mostafazadeh-Fard, B., Dehghani, S., & Khorvash, M. (2012).

568

Linear and Nonlinear Modeling for Predicting Nickel Removal from Aqueous Solutions.

569

Environmental Engineering Science, 29(8), 765–775. http://doi.org/10.1089/ees.2011.0233

570

Mozhayeva D, Zhakupbekova A, Kenessov B, Akmoldayeva S, C. L. (2014). New method for

571

the discovery of adulterated cognacs and brandies based on solid-phase, 76, 53–66.

572

Qu, M., Shih, F. Y., Jing, J., & Wang, H. (2003). Automatic solar flare detection using MLP,

573

RBF, and SVM. Solar Physics, 217(1), 157–172. http://doi.org/10.1023/A:1027388729489

574

Ruiz Orduna, A., Husby, E., Yang, C. T., Ghosh, D., & Beaudry, F. (2015). Assessment of meat

575

authenticity using bioinformatics, targeted peptide biomarkers and high-resolution mass

576

spectrometry. Food Additives & Contaminants: Part A, 32(10), 1709–1717.

577

http://doi.org/10.1080/19440049.2015.1064173

578

Smola, A. J., & Schölkopf, B. (2004). A Tutorial on Support Vector Regression. Statistics and

26

ACCEPTED MANUSCRIPT

579 580

Computing, 14(3), 199–222. http://doi.org/10.1023/B:STCO.0000035301.49549.88 Alonso, J. M., Kodad, A., & Gradziel, T. M. (2012). Almond. In M. L. Badenes & D. H. Byrne

581

(Eds) Fruit Breeding (p. 875) Springer, New York. http://doi.org/10.1007/978-1-4419-

582

0763-9

583

Suykens, J. A. K. (2001). Nonlinear modelling and support vector machines. IMTC 2001

584

Proceedings of the 18th IEEE Instrumentation and Measurement Technology Conference

585

Rediscovering Measurement in the Age of Informatics Cat No01CH 37188, 1, 287–294.

586

http://doi.org/10.1109/IMTC.2001.928828

587

Vapnik, V. N. (1998). Statistical Learning Theory. Adaptive and Learning Systems for Signal

588

Processing, Communications and Control, 2, 1–740. http://doi.org/10.2307/1271368

589

Varmuza, K., & Filzmoser, P. (2010). Introduction to multivariate statistical analysis in

590

chemometrics. Applied Spectroscopy (Vol. 64).

591

http://doi.org/10.1366/000370210791114185

592

Wang, Z., & Jablonski, J. E. (2016). Targeted and non-targeted detection of lemon juice

593

adulteration by LC-MS and chemometrics. Food Additives & Contaminants: Part A, 33(3),

594

560–573. http://doi.org/10.1080/19440049.2016.1138547

595

Welke, J. E., Damasceno, F. C., Nicolli, K. P., Mentz, L., Caramao, E. B., Pulgatid, F. H., &

596

Zini, C. A. (2015). Preliminary investigation of medicinal herb adulteration using

597

comprehensive two-dimensional gas chromatography and chemometric analysis. Journal of

598

the Brazilian Chemical Society, 26(6), 1232–1246. http://doi.org/10.5935/0103-

599

5053.20150089

600

Wu, C. H., Tzeng, G. H., Goo, Y. J., & Fang, W. C. (2007). A real-valued genetic algorithm to

601

optimize the parameters of support vector machine for predicting bankruptcy. Expert

27

ACCEPTED MANUSCRIPT

602 603

Systems with Applications, 32(2), 397–408. http://doi.org/10.1016/j.eswa.2005.12.008 Zhang, J., Zhang, X., Dediu, L., & Victor, C. (2011). Review of the current application of

604

fingerprinting allowing detection of food adulteration and fraud in China. Food Control.

605

http://doi.org/10.1016/j.foodcont.2011.01.019

606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623

28

ACCEPTED MANUSCRIPT

624 625 626 627

Figure captions

628

Fig. 1 a) PC1-PC2 score plot, b) Discriminant scores on the two-dimension space defined by the

629

first and second discriminant functions of pure samples and mixtures of almond and peanut (in

630

the range 30/70 – 70/30, m/m).

631

Fig. 2 PLS modeling based on the multi-elemental fingerprinting to predict adulteration of

632

almond with peanut a) RMSEC and RMSECV versus the number of latent variables in the PLS

633

model; Prediction results of best PLS model for calibration (b) and prediction (c) sets.

634

Fig. 3 LS-SVM modeling based on the multi-elemental fingerprinting to predict adulteration of

635

almond with peanut. (a) Grid search on γ and σ2; Prediction results of the best LS-SVM model

636

for calibration (b) and prediction (c) sets.

637

Fig. 4 RMSECV values after omitting different variables from the PLS and SVM models.

638 639 640 641 642 643

29

ACCEPTED MANUSCRIPT Highlights    

Detection of peanut adulteration in almond powder by multi-elemental fingerprinting. PCA and LDA were able to differentiate the adulterated samples. Detection of adulteration up to the level of 5% was possible using LS-SVM model. The effective elements in differentiating almond and peanut samples were determined.

ACCEPTED MANUSCRIPT

Table 1. The ICP-OES operating parameters.

Instrumental Parameter

Level

Generator Frequency (MHz)

40

RF Power (kW)

1.3

Plasma gas flow rate (L min-1)

15

Auxiliary gas flow rate (L min-1)

0.2

Nebulizer gas flow rate (mL min-1)

800

Integration time (s)

3

Stabilization time (s)

15

ACCEPTED MANUSCRIPT

Table 2. Analytical figures of merit from the ICP-OES determination of the elements in almond samples, digested in a teflon lined stainless steel autoclave. Spectral

RSD(%)a

LOD Spectral RSD(%)a LOD In solution line In solution In sample In sample Element line (nm) Element (µg g-1)c (µg L-1)b (nm) (µg L-1)b (µg g-1)c Li 413.256 0.14 50 2.5 Ni 221.648 0.03 10 0.5 Be 234.86 0.08 0.3 0.015 Cu 213.597 1.24 12 0.6 B 182.528 1.40 50 2.5 Zn 202.548 0.70 4.0 0.2 Na 588.995 2.41 29 1.45 Ga 294.364 1.68 46 2.3 Mg 279.077 0.88 30 1.5 Se 196.026 1.47 75 3.75 Al 308.215 1.14 45 2.25 Sr 232.235 0.52 103 5.15 K 766.490 0.27 40 2 Cd 214.44 0.32 2.5 0.125 Ca 315.88 3.06 10 0.5 Te 214.281 3.29 41 2.05 Cr 205.560 0.98 6.1 0.30 Ba 230.42 0.56 4.1 2.05 Mn 257.610 1.54 1.4 0.07 Tl 190.801 3.12 40 2.0 Fe 234.349 1.09 13 0.65 Pb 217.000 0.98 40 2.0 Co 228.616 0.88 7 0.35 Bi 190.171 1.17 34 1.7 a RSD data were calculated as percentage corresponding to (s/x ) * 100 mean b LODs were calculated as the concentration corresponding to signals equal to three times the standard deviation of a blank solution. c LODs data in µg g-1 of almond sample were calculated taken into consideration the amount of almond digested and the appropriate dilution.

ACCEPTED MANUSCRIPT

Table 3. Mean concentration and the range of the elements (µg g-1) in almond and peanut samples. Almond Element Mean Range SD Al
Peanut Mean
Range 16.4 - 22.3

673.0 - 787.1

17.3 - 22.7 20.7 - 29.2 4931.1 - 5432.0 1874.7 - 2048.8 11.6 - 13.3

11.1 - 14.7 27.8 - 32.4

SD
ACCEPTED MANUSCRIPT

Table 4. The calibration, cross-validation and validation results of the LDA classification model. Classes Almond Peanut Adulterated

Predicted/Assigned Almond Peanut 25 0 0 25 0 0

Adulterated 0 0 50

Total 25 25 50

Cross-validation

Almond Peanut Adulterated

25 0 0

0 25 0

0 0 50

25 25 50

Validation set

Adulterated

0

1

49

50

Calibration

ACCEPTED MANUSCRIPT

Table 5. Calibration and validation results of PLS and LS-SVM models from using multielemental fingerprints. Method

RMSEC

RMSECV

RMSEP

R2calibration

R2 cross-validation

R2 validation

PLS LS-SVM

1.30 0.54

2.15 0.92

3.81 1.66

0.995 0.999

0.989 0.996

0.986 0.997