- Email: [email protected]
Spectroscopic determination of the photodegradation of monovarietal extra virgin olive oils and their binary mixtures through intelligent systems
Talanta 144 (2015) 363–368
Contents lists available at ScienceDirect
Talanta journal homepage: www.elsevier.com/locate/talanta
Spectroscopic determination of the photodegradation of monovarietal extra virgin olive oils and their binary mixtures through intelligent systems José S. Torrecilla a,n, Sara Vidal a, Regina Aroca-Santos a, Selina C. Wang b, John C. Cancilla a a b
Departamento de Ingeniería Química, Facultad de Ciencias Químicas, Universidad Complutense de Madrid, 28040 Madrid, Spain University of California-Davis, Olive Center, Davis, CA 95616, United States
art ic l e i nf o
a b s t r a c t
Article history: Received 8 April 2015 Received in revised form 11 June 2015 Accepted 17 June 2015 Available online 24 June 2015
A common phenomenon that takes place in bottled extra virgin olive oil (EVOO) is the photooxidation of its pigments, especially chlorophyll, which acts as a singlet-oxygen sensitizer. This translates into a severe decrease of quality, potentially leading to oxidized and rancid olive oils by the time they reach to the consumers. In this current research, the photochemical degradation has been monitored for 45 days in binary mixtures of four monovarietal EVOOs (Arbequina, Hojiblanca, Cornicabra, and Picual) through UV–Visible spectroscopy. A multilayer perceptron-based model was optimized to estimate the photodegradation suffered by the samples, in terms of photodegradation time, relying on the spectroscopic information gathered and attaining an error rate of 2.43 days (5.3%) in the determination of this parameter. & 2015 Published by Elsevier B.V.
Keywords: Extra virgin olive oil Chlorophyll Carotenoids Pigments Spectroscopy Artificial neural networks
1. Introduction Extra virgin olive oil (EVOO) is one of the key components of the Mediterranean diet [1,2], and it presents a series of traits that make it beneficial for health due to its characteristic composition [1,3], which is typically divided into two main fractions: the saponifiable fraction (comprising 98–99% of the weight of EVOO) and the non-saponifiable fraction (containing the remaining 1–2%) [1]. The first one consists of triglycerides, diglycerides, monoglycerides, and free fatty acids, while the non-saponifiable fraction is formed by minor components, such as pigments, volatile compounds, polyphenols, tocopherols, and sterols [1,3]. It is important to remark the role of the pigments, particularly chlorophylls and carotenes, which provide its characteristic color to EVOO [2,4]. Additionally, they have been found to be somewhat more than colorful molecules, as pigments have shown a relation with the mitigation of the effects of certain degenerative diseases, as well as possessing anticarcinogenic and antimutagenic qualities [5,6]. Moreover, pigments play an essential part in the oxidation and degradation of EVOO due to the effect of both light and temperature [5,6]. In the case of chlorophylls, the demetalation and loss of the magnesium atom in the center of the tetrapyrrole n
Corresponding author. Fax: þ 34 91 394 42 43. E-mail address: [email protected] (J.S. Torrecilla).
http://dx.doi.org/10.1016/j.talanta.2015.06.042 0039-9140/& 2015 Published by Elsevier B.V.
macrocycle is the first step of its degradation mechanism, and transforms these molecules into pheophytins (Fig. 1) [7]. Although this first stage does not involve any change of color, there is a light-favored pathway that leads to the photolytic cleavage of the chromophore of pheophytin through its photooxidation, generating a colorless molecule as a product [6]. It has been observed that chlorophyll pigments act as photosensitizers as they can transfer energy from electromagnetic radiation to triplet oxygen, thus producing singlet oxygen, a highly oxidative species [8], which would eventually react with the unsaturated fatty acids of EVOO, degrading the oil and lowering its quality [8]. On the contrary, carotenoids (which are mainly represented in EVOO by lutein and β-carotene, Fig. 2) have a remarkable antioxidant activity, protecting the oil against photooxidation by acting as singlet oxygen-quenchers and light filters [5,6,9]. Although the pigment-related compounds are the major responsible molecules of the photooxidation process that takes place in EVOO, the loss of organoleptic characteristics and quality affects the whole product. Therefore, it is crucial during bottling, packaging, transportation, and storage to limit the exposure of the EVOO to light and changes in temperature to ensure that the initial high quality bottled oil reaches the consumer in these same (or almost same) conditions [10]. As the EVOO photodegradation process leads to a cascade of
364
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
carry out the data treatment, providing a tool that is able to evaluate whether the product has been satisfactorily treated during distribution or not.
2. Materials and methods In the following subsection, the samples considered, the preparation steps applied, the analytical procedures followed, and the mathematical tools employed are described. 2.1. Sample preparation Fig. 1. Demetalation of the chlorophyll molecule (pheophytinization).
Four monovarietal EVOOs have been used to prepare the samples analyzed: Arbequina, Hojiblanca, Cornicabra, and Picual, all of which are within their expiration date, and stored in the dark at room temperature. A total of 22 samples were prepared: four of them corresponding to the single monovarietal EVOOs, and the rest, to binary mixtures of 25%, 50%, and 75% (v/v), leading to six samples for each possible mixture of olive oils. Once prepared, the samples were exposed to indirect sunlight for a total of 45 days. 2.2. UV–vis absorption measurements
Fig. 2. Chemical 2D structure of lutein (a) and β-carotene (b).
physicochemical changes that affect the chromophore molecules, it can be supervised and registered by optical techniques. More specifically, these modifications could be detected through UV– Visible absorption spectroscopy, as the monitored compounds are active in this area of the electromagnetic spectrum. In fact, several methods that rely on this technique to control the degradation of EVOO have been already reported: the determination of the thermal degradation of olive oil [11], the analysis of the compositional changes that take place during the storage of EVOO [10], the study of the kinetics that govern the thermal degradation of the pigments in olive oil [6], or even the objective authentication of monovarietal EVOOs [12]; thus showing that the employment of this spectroscopic technique is useful to determine compositional and quality-related changes in EVOO over time due to the degradation of its pigments. On the other hand, using UV–Visible absorption spectroscopy provides a relevant amount of data points, making necessary the use of mathematical and statistical procedures to process and adequate the data, such as principal components analysis (PCA), linear discriminant analysis (LDA), or soft independent modeling of class analogy (SIMCA) [13–15]. Nevertheless, sometimes these linear procedures do not offer the expected results and more sophisticated models are required. One of these advanced mathematical tools are artificial neural networks (ANNs), non-linear models which can determine relationships between independent and dependent variables within databases through non-linear interpolation [16]. They have been already employed in the study of olive oil and the olive sector, offering accurate estimations [17,18]. As the EVOO is photodegraded, it loses its organoleptic characteristics as well as quality grade, so the product that is provided to the consumers may not possess the initial nutritional nor healthy effects mentioned before if the product is not handled adequately. Therefore, the goal of this work is to offer a userfriendly and straightforward procedure to estimate the photodegradation time of four monovarietal EVOOs (Arbequina, Hojiblanca, Cornicabra, and Picual) and their binary mixtures by measuring their UV–vis absorption spectra and using ANNs to
The samples and a reference (stored in the dark) were daily measured five times per week (Monday through Friday) for a period of 45 days, in order to monitor the changes experimented by the pigment profile due to the presence of sunlight, thus evaluating the extent of the photodegradation suffered by the different pigments of the samples of EVOO. A total of 30 data points per sample were obtained. The absorption spectra of the samples were collected between 350 and 800 nm using a UV–vis spectrophotometer (Varian Cary 50 Conc). The samples were introduced in a quartz cuvette of 1 cm path length and the samples were analyzed three times in order to obtain an averaged spectrum of each one. This led to spectra containing one absorption value per wavelength, which originates 450 potential variables. The first phase of the mathematical treatment of the spectra was reducing the dimensionality of the system. This approach enables the elimination of redundant and/or non-informative variables (absorbance–wavelength pairs). In the case considered, the calculation of the areas under the curve (AUCs) of the different absorption bands observed was the procedure selected to both reduce the number of independent variables and extract the useful spectroscopic information from the bands, as a combination of AUCs might contain more information than absorption peaks at discrete wavelengths [19]. The mathematical treatment of the data was performed with OriginPro 8.0 and Matlab 2013 software packages. 2.3. Database The final database contains a total of 658 spectra (after discarding two defective measurements which qualified as outliers in the preliminary analysis), corresponding to the measurement of the different mixtures of EVOO over time. The information gathered within the database possesses compositional, spectroscopic, and chronological data regarding the different samples analyzed. 2.4. Artificial neural networks ANNs were the selected mathematical tool to assist in the estimation of the degradation time of the samples. These algorithms are able to determine non-linear relations between dependent and independent variables within databases through non-linear
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
365
data distributions to train MLPs [19,20]. The training phase takes place firstly, where the training dataset is processed by the model. The MLP performs the estimation and calculates the error attained (comparing the estimated response with the actual target values). In order to decrease this error value, the weights are modified towards obtaining the lowest possible error [16,19,20]. If this process keeps going on, the error would potentially reach a value of zero, as the weights are strictly improved to achieve a perfect fit to the training data available. However, following this training procedure will lead to the design of an over-fit model, unable to perform properly with new sets of data [22,23]. The verification stage is in charge of avoiding this phenomenon, and occurs just after the training step. In this case, the weight values remain fixed and the verification dataset is used to test the generalization capability of the model, or, in other words, the capability of the model to correctly perform accurately when facing external data [25]. Once this whole process has finished, a training cycle, or epoch, is completed. The accuracy achieved by the MLP is studied regarding the error obtained with the verification dataset, in terms of mean absolute error (MAE).
Fig. 3. Scheme of the MLP designed. AUCi corresponds to the different AUCs calculated.
interpolation. Even though there are different kinds of ANNs, the supervised multilayer perceptron (MLP) is one of the most successfully employed ANNs in scientific research [20,21], and it was the one selected to assist in the case proposed. As it is a supervised-learning ANN, there must be both input and target data so the MLP can be correctly trained and optimized [22]. Since the estimation mechanism of ANNs is based on non-linear interpolation, they strongly depend on the range of data covered by the database, because if the model were forced to extrapolate, the error attained in this case would increase [16,23]. Additionally, they highly depend on the statistical quality of the data available (in terms of low deviation and good signal-to-noise ratio) to perform accurate estimations [23]. MLPs are formed by units that are arranged in layers. There are three different types of layers: the input, the hidden, and the output layers. The first one is solely formed by nodes, which only introduce the value of the independent variables into the network. The hidden and output layers contain neurons, the actual calculation centers of the system, (Fig. 3). The quantity of neurons that occupy the output layer depends on the number of dependent variables considered in the system modeled, while the number of units in the hidden layer requires further optimization (vide infra) [21]. On the other hand, as the input nodes present the value of the independent variables to the model, there must be as many nodes as independent variables. Every unit in the MLP is connected to every unit in the neighboring layers. The importance of these connections is controlled by weighted parameters, named weights. They are essential for the final performance of the model, and their value must be optimized to attain a low error or high accuracy in the estimations attempted [24]. 2.4.1. ANN optimization In order to be able to appropriately operate, ANNs require the optimization of their weights [19,20,23]. This process comprises two distinct stages: the first one is known as the training phase, and the second one, the verification phase. To correctly fulfill these two steps, the database must be previously randomly divided: 85% of the data is used in the training stage (training dataset), while the remaining 15% is employed in the verification stage (verification dataset). This is one of the most typical training–verification
MAE =
1 N
N
∑
rk − yk
k=1
(1)
In Eq. (1), n stands for the number of data points of the verification dataset, rk is the already known real value, and yk is the response given by the model. Unlike the training error, the verification error will eventually start rising. In that moment, the model can be thought as optimized and the training cycles conclude [25]. In addition, there are considerations that must be taken into account before thinking of the designed MLP as fully optimized. They are the training and transfer functions, the hidden neuron number (HNN), the learning coefficient (Lc), the learning coefficient increase (Lci), and the learning coefficient decrease (Lcd). The training function must be carefully selected in order to obtain accurate and reliable estimations, as it is in charge of the learning process. Depending on the training function chosen, the training algorithm will vary and, thus, offer a different estimation performance and accuracy [16]. Also, the transfer function should be fixed. It provides the ability to carry out the non-linear interpolations as well as limiting the range of the results emitted by a neuron in order to make results comparable within the MLP. In this case, the selected transfer function was the sigmoid function (Eq. (2)), which normalized the results between 0 and 1.
⎛ ⎞ 1 ⎟ yk = ⎜ ⎝ 1 + e− x k ⎠
(2)
where xk is the result reached by a neuron and yk the solution of the sigmoid transfer function [26]. On the other hand, the HNN must be also optimized to obtain the combination of neurons that provides the lowest estimation error. The quantity of these neurons is extremely important for the final performance of the model: if the HNN employed is too high, it can produce an over-fit network; but if the number is too low, it would lead to a decrease in the learning capability of the model [16,19]. The procedure chosen for the optimization of the HNN was a trial-and-error or heuristic method, considering that the number of weights must be the lowest possible, and always smaller than the number of data points available. Finally, the optimization process must contemplate the different Lcs. The Lc parameter is responsible for the modification of the weights, while the Lcd and Lci parameters control the extent of the variation of Lc [16]. In order to study their combined effect and determining the optimum combination of these coefficients, they
366
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
were optimized following an experimental design based on the “Box–Wilson Central Composite Design 23 þstar points”. The ranges studied for the different coefficients were from 0.001 to 1 in the case of Lc and Lcd, while Lci was studied between 2 and 100 [19]. The design of the ANN-based model developed, as well as its optimization process, was carried out employing Matlab 7.0 and Statgraphics Centurion XVI software packages. 2.4.2. ANN validation In addition to the tests performed during the selection of the HNN, it is vital to check and ensure that the model operates properly throughout the whole range considered by the database, and that it can generalize well and not result in an over-fitting model. To do so, a statistical procedure known as k-fold crossvalidation test was performed. This method is based on the division of the whole database in k datasets which are employed sequentially as verification dataset while the remaining k–1 groups are used as training dataset [27]. Therefore, every data point available is both used in the training and verification datasets in different moments, so the model can be tested throughout the whole range considered by the database.
3. Results and discussion In this section, the mathematical processing of the spectra, the final MLP model designed, and the results achieved throughout the optimization process are reported. 3.1. Analysis of the spectra To begin analyzing the absorption spectra, a dimensionality reduction phase was carried out to eliminate redundant and/or unrepresentative information. To do so, the AUCs of the five main absorption bands have been selected as input information because these magnitudes have been successfully used previously in the same kind of models with positive results [19]. The AUCs were calculated from 440 to 472 nm (AUC1), 473 to 526 nm (AUC2), 527 to 551 nm, (AUC3), 593 to 630 nm (AUC4), and 631 to 703 nm (AUC5), so the bands corresponding to the absorption of the pigments were individually selected to avoid any overlapping (Fig. 4)
Fig. 4. AUC calculation limits represented in the UV–vis absorption spectra of three different EVOO samples: non-degraded Hojiblanca monovarietal EVOO (dashes), 15-days degraded Hojiblanca EVOO (dots), and a non-degraded binary mixture 50% Arbequina, 50% Picual EVOOs (line).
[12,28]. It can be observed that the spectra shown in Fig. 4 can be distinguished, both in terms of photodegradation time and composition. In the case of the Hojiblanca monovarietal EVOO sample, a decrease in the absorption bands due to the photodegradation of the pigments over time (both chlorophylls and carotenoids) can be attributed mainly to the demetalation and succeeding degradation of chlorophylls, which would then act as singlet-oxygen sensitizers, accelerating the photodegradation and photooxidation of the rest of pigments and antioxidants that form the minor fraction of EVOO, and thus decreasing the nutritional value and quality of the oil due to the oxidation of free fatty acids and other compounds [8]. However, it must be noted that the degradation of the samples might be also affected by the presence of oxygen in the head-space of their vials (constant for every sample), as it contributes to the oxidative processes [8,29]. It is possible to link the decrease in absorption to photodegradation, as the reference stored in the dark maintained a stable spectrum with time. The AUCs calculated have been used as input information in the ANN model designed with the intention of determining the degradation time (sunlight exposure) that each sample has suffered, regardless of its composition or olive oil varietal. The remaining spectra, corresponding to the different measurements of the samples analyzed, can be found in the Supplementary Information section. 3.2. ANN optimization As stated above, in addition to the optimization of the value of the weights that take place during the training phase, several parameters (HNN, Lc, Lcd, and Lci) and the training function were optimized. All the results, error rates, and correlation coefficients shown have been calculated from the verification datasets, not the training datasets, as the last ones are supposed to perfectly fit the target values and provide no information regarding the actual behavior and performance of the MLP. In other words, the verification datasets also acted as prediction datasets, as they were employed to evaluate the statistical performance of the non-linear models. In order to design a functional MLP, able to accurately estimate the photodegradation time suffered by EVOO samples through the value of the AUCs of the spectroscopic bands corresponding to the absorption of the pigments, the considerations specified before were assessed. The training function selected was the Levenberg– Marquardt training function (TrainLM), as it provides good results when dealing with a high number of data points. It possesses, in addition, a memory reduction feature that permits it to work effectively even when the computational requirements are elevated. This algorithm has already been applied with success in similar cases in the past [16,30]. Once the training function is selected, different HNN were tested, from 9 to 30 neurons, to determine the combination which leads to achieve the lowest estimation error. The attained results are summarized in Table 1. It can be seen that the error decreased when the HNN was tested from 15 to 21 neurons, while R2 followed a more irregular trend. However, from this point, the error rose again and the R2 slightly decreased. In light of these results, the HNN selected was 18, as it was the combination that provided the lowest error and the highest R2 values. Although the number of data points would allow the use of a higher HNN (up to 90 hidden neurons), the trend found from 21 to 30 neurons suggests that the subsequent tests would not improve the already attained results. Therefore, the following optimization steps were carried out with a HNN of 18. Finally, to improve the network performance, the training
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
thus enabling the control of the quality of EVOO when it arrives to the consumers and compare it to its initial properties. This would allow the study of the different photo-induced phenomena that occurs during the different stages prior to the EVOO sale, such as transportion and storage to assess how the product has been handled during this entire distribution chain.
Table 1 Results attained in the HNN tests, in terms of MAE (in days) and R2. The option selected for the model designed appears in bold. HNN
MAE (days)
R2
9 10 12 15 16 17 18 19 20 21 25 30
3.24 2.74 3.02 2.87 2.72 2.95 2.55 2.81 3.09 3.06 2.72 2.72
0.89 0.91 0.90 0.90 0.91 0.90 0.92 0.90 0.90 0.89 0.91 0.91
367
4. Conclusions
Table 2 Statistical results of the performance of the model when the k-fold cross-validation tests are performed. K
MAE (days)
R2
1 2 3 4 5 6 Average
2.55 2.51 2.32 2.46 2.59 2.13 2.43
0.92 0.92 0.95 0.94 0.92 0.96 0.94
parameters Lc, Lcd, and Lci were optimized following the procedure mentioned above. The values of the optimum combination found were 0.001 for Lc, 0.001 for Lcd and 2 in the case of Lci, and the error attained when they were applied was the lowest, 5%, which implies an error of 2.43 days in the estimation of time of degradation. These statistical performances, alike all the other optimization steps, were obtained with the prediction dataset, and not the training dataset, as these samples are used to solely optimize the weights. 3.3. Validation of the network As noted above, the generalization capability of the MLP designed was tested following a k-fold cross-validation statistical procedure, with k ¼6, allowing to determine the true applicability of the final model in the whole range of data, instead of limiting this test to a localized, randomly selected verification set. The statistical results can be seen in Table 2. The final model designed offered an error of 2.43 days in the estimation of the degradation time of the samples of EVOO analyzed, which represents a 5.3% of error regarding the whole time span analyzed (45 days). In addition, the R2 coefficient shows a high value (0.94), confirming a suitable performance of the model. These results show that ANNs-based models, such as the one here proposed, are able to accurately and reliably aid in the determination of the photodegradation suffered by single and binary mixtures of monovarietal EVOO samples (in terms of degradation time), starting from spectroscopic information regarding their pigment profile (chlorophyll, pheophytin, and carotenoid content). Therefore, the procedure proposed permits the evaluation of the photodegradation process of EVOOs by monitoring the modifications of the UV–vis spectra of the main pigments of this oil when exposed to the light. Also, a user-friendly, reliable, and straightforward chemometric tool able to control the changes and modifications suffered by the final product has been developed,
The estimation of the photodegradation process using binary mixtures of four monovarietal EVOOs has been achieved through the analysis of the UV–vis absorption spectra and the mathematical support of an ANN-based model. The AUC of the absorption bands which correspond to the chlorophyll and carotenoid pigments have been obtained and employed by the optimized MLP to determine the photodegradation time suffered by the different samples, attaining an average error rate of 5.3% during a 45-day period of time. The accuracy attained suggests that the proposed tool could be employed to control the modifications suffered by EVOO since its production in the olive mill until its sale, analyzing the loss of quality and pigment photodegradation that occurs during the transport, storage, and shelf stages. Although this distribution time is in occasions greater than 45 days, a proof-ofconcept has been set, and potentially larger time scales can also be accurately monitored. Therefore, this procedure would allow determining the influence of each step in the EVOO quality and photodegradation extent, as well as the improvement of the strategies and methods designed to preserve the quality of the oil so the final consumer can enjoy the expected high quality product.
Appendix A. Supplementary material Supplementary data associated with this article can be found in the online version at doi:10.1016/j.talanta.2015.06.042.
References [1] I. Vasilescu, S.A.V. Eremia, C. Albu, A. Radoi, S.C. Litescu, G.L. Radu, Food Chem. 166 (2015) 324–329. [2] S. Dabbou, F. Brahmi, A. Taamali, M. Issaoui, Y. Ouni, M. Braham, M. Zarrouk, M. Hammami, J. Am. Oil Chem. Soc. 87 (2010) 1199–1209. [3] M.N. Franco, T. Galeano-Díaz, Ó. López, J.G. Fernández-Bolaños, J. Sánchez, C. De Miguel, M.V. Gil, D. Martín-Vertedor, Food Chem. 163 (2014) 289-289. [4] A. Fernández-Gutiérrez, A. Segura-Carretero, El. Aceite de Oliva Virgen Tesoro de Andalucía, Fundación Unicaja, 2010, p. 17. [5] R. Aparicio-Ruiz, M.I. Mínguez-Mosquera, B. Gandul-Rojas, J. Food Compos. Anal. 24 (2011) 811–820. [6] R. Aparicio-Ruiz, M.I. Mínguez-Mosquera, B. Gandul-Rojas, J. Agric. Food Chem 58 (2010) 6200–6208. [7] Y. Saga, H. Tamiaki, Chem. Biodivers. 9 (2012) 1659–1683. [8] E. Psomiadou, M. Tsimidou, J. Agric. Food Chem. 50 (2002) 716–721. [9] E. Psomiadou, M. Tsimidou, J. Agric. Food Chem. 50 (2002) 722–727. [10] J.R. Morelló, M.J. Motilva, M.J. Tovar, M.P. Romero, Food Chem. 85 (2004) 357–364. [11] I.P. Paz-Antolín, M. Molero-Meneses, Grasas Aceites 51 (2000) 424–428. [12] M.I. Mínguez-Mosquera, L. Rejano-Navarro, B. Gandul-Rojas, A.H. SánchezGómez, J. Garrido-Fernandez, J. Am. Oil Chem. Soc. 68 (1991) 332–336. [13] C. Pizarro, S. Rodríguez-Tejedor, N. Pérez-del-Notario, J.M. González-Saiz, J. Chormatogr. A 1218 (2011) 518–523. [14] N. Sinelli, M. Casale, V. DiEgidio, P. Oliveri, D. Bassi, D. Tura, E. Casiraghi, Food Res. Int. 43 (2010) 2126–2131. [15] R. Bucci, A.D. Magrí, A.L. Magrí, D. Marini, F. Marini, J. Agric. Food Chem. 50 (2002) 413–418. [16] H. Demuth, M. Beale, M. Hagan, Neural Network Toolbox for Use with MATLABs User’s Guide, Version 5, Ninth printing Revised for Version 5.1 (Release 2007b), Natick, MA, USA, 2007 (online only). [17] J.S. Torrecilla, J.C. Cancilla, G. Matute, P. Díaz-Rodríguez, Int. J. Food Sci. Technol. 48 (2013) 2528–2534. [18] J.S. Torrecilla, J.C. Cancilla, G. Matute, P. Díaz-Rodríguez, A.I. Flores, J. Food Eng. 118 (2013) 400–405.
368
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
[19] P. Díaz-Rodríguez, J.C. Cancilla, K. Wierzchoś, J.S. Torrecilla, Sensor Actuat. B— Chem. 206 (2015) 139–145. [20] J.S. Torrecilla, J.C. Cancilla, G. Matute, P. Díaz-Rodríguez, Int. J. Food Sci. Technol. 48 (2013) 2528–2534. [21] S. Curteanu, H. Cartwright, J. Chemometr. 25 (2011) 527–549. [22] J.S. Torrecilla, F. Rodríguez, J.L. Bravo, G. Rothenberg, K.R. Seddon, I. LópezMartin, Phys. Chem. Chem. Phys. 10 (2008) 5826–5831. [23] L. Cakir, N. Yilmaz, Measurement 57 (2014) 148–153. [24] J. De-Ona, C. Garrido, Neural Comput. Appl. 25 (2014) 859–869. [25] J.S. Torrecilla, C. Tortuero, J.C. Cancilla, P. Díaz-Rodríguez, Talanta 116 (2013) 122–126. [26] K. Knoerzer, P. Juliano, P. Roupas, C. Versteeg, Innovative Food Processing
[27] [28] [29] [30]
Technologies: Advances in Multiphysics Simulation, Wiley-Blackwell, Oxford, UK, 2011. M. Zounemat-Kermani, O. Kisi, T. Rajaee, Appl. Soft Comput. 13 (2013) 4633–4644. V. Domenici, D. Ancora, M. Cifelli, A. Serani, C.A. Veracini, M. Zandomeneghi, J. Agric. Food Chem. 62 (2014) 9317–9325. R. Aparicio-Ruiz, M. Roca, B. Gandul-Rojas, J. Agric. Food Chem. 60 (2012) 7040–7049. A. Bassam, R.A. Conde-Gutiérrez, J. Castillo, G. Laredo, J.A. Hernández, Fuel 124 (2014) 158–167.
Contents lists available at ScienceDirect
Talanta journal homepage: www.elsevier.com/locate/talanta
Spectroscopic determination of the photodegradation of monovarietal extra virgin olive oils and their binary mixtures through intelligent systems José S. Torrecilla a,n, Sara Vidal a, Regina Aroca-Santos a, Selina C. Wang b, John C. Cancilla a a b
Departamento de Ingeniería Química, Facultad de Ciencias Químicas, Universidad Complutense de Madrid, 28040 Madrid, Spain University of California-Davis, Olive Center, Davis, CA 95616, United States
art ic l e i nf o
a b s t r a c t
Article history: Received 8 April 2015 Received in revised form 11 June 2015 Accepted 17 June 2015 Available online 24 June 2015
A common phenomenon that takes place in bottled extra virgin olive oil (EVOO) is the photooxidation of its pigments, especially chlorophyll, which acts as a singlet-oxygen sensitizer. This translates into a severe decrease of quality, potentially leading to oxidized and rancid olive oils by the time they reach to the consumers. In this current research, the photochemical degradation has been monitored for 45 days in binary mixtures of four monovarietal EVOOs (Arbequina, Hojiblanca, Cornicabra, and Picual) through UV–Visible spectroscopy. A multilayer perceptron-based model was optimized to estimate the photodegradation suffered by the samples, in terms of photodegradation time, relying on the spectroscopic information gathered and attaining an error rate of 2.43 days (5.3%) in the determination of this parameter. & 2015 Published by Elsevier B.V.
Keywords: Extra virgin olive oil Chlorophyll Carotenoids Pigments Spectroscopy Artificial neural networks
1. Introduction Extra virgin olive oil (EVOO) is one of the key components of the Mediterranean diet [1,2], and it presents a series of traits that make it beneficial for health due to its characteristic composition [1,3], which is typically divided into two main fractions: the saponifiable fraction (comprising 98–99% of the weight of EVOO) and the non-saponifiable fraction (containing the remaining 1–2%) [1]. The first one consists of triglycerides, diglycerides, monoglycerides, and free fatty acids, while the non-saponifiable fraction is formed by minor components, such as pigments, volatile compounds, polyphenols, tocopherols, and sterols [1,3]. It is important to remark the role of the pigments, particularly chlorophylls and carotenes, which provide its characteristic color to EVOO [2,4]. Additionally, they have been found to be somewhat more than colorful molecules, as pigments have shown a relation with the mitigation of the effects of certain degenerative diseases, as well as possessing anticarcinogenic and antimutagenic qualities [5,6]. Moreover, pigments play an essential part in the oxidation and degradation of EVOO due to the effect of both light and temperature [5,6]. In the case of chlorophylls, the demetalation and loss of the magnesium atom in the center of the tetrapyrrole n
Corresponding author. Fax: þ 34 91 394 42 43. E-mail address: [email protected] (J.S. Torrecilla).
http://dx.doi.org/10.1016/j.talanta.2015.06.042 0039-9140/& 2015 Published by Elsevier B.V.
macrocycle is the first step of its degradation mechanism, and transforms these molecules into pheophytins (Fig. 1) [7]. Although this first stage does not involve any change of color, there is a light-favored pathway that leads to the photolytic cleavage of the chromophore of pheophytin through its photooxidation, generating a colorless molecule as a product [6]. It has been observed that chlorophyll pigments act as photosensitizers as they can transfer energy from electromagnetic radiation to triplet oxygen, thus producing singlet oxygen, a highly oxidative species [8], which would eventually react with the unsaturated fatty acids of EVOO, degrading the oil and lowering its quality [8]. On the contrary, carotenoids (which are mainly represented in EVOO by lutein and β-carotene, Fig. 2) have a remarkable antioxidant activity, protecting the oil against photooxidation by acting as singlet oxygen-quenchers and light filters [5,6,9]. Although the pigment-related compounds are the major responsible molecules of the photooxidation process that takes place in EVOO, the loss of organoleptic characteristics and quality affects the whole product. Therefore, it is crucial during bottling, packaging, transportation, and storage to limit the exposure of the EVOO to light and changes in temperature to ensure that the initial high quality bottled oil reaches the consumer in these same (or almost same) conditions [10]. As the EVOO photodegradation process leads to a cascade of
364
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
carry out the data treatment, providing a tool that is able to evaluate whether the product has been satisfactorily treated during distribution or not.
2. Materials and methods In the following subsection, the samples considered, the preparation steps applied, the analytical procedures followed, and the mathematical tools employed are described. 2.1. Sample preparation Fig. 1. Demetalation of the chlorophyll molecule (pheophytinization).
Four monovarietal EVOOs have been used to prepare the samples analyzed: Arbequina, Hojiblanca, Cornicabra, and Picual, all of which are within their expiration date, and stored in the dark at room temperature. A total of 22 samples were prepared: four of them corresponding to the single monovarietal EVOOs, and the rest, to binary mixtures of 25%, 50%, and 75% (v/v), leading to six samples for each possible mixture of olive oils. Once prepared, the samples were exposed to indirect sunlight for a total of 45 days. 2.2. UV–vis absorption measurements
Fig. 2. Chemical 2D structure of lutein (a) and β-carotene (b).
physicochemical changes that affect the chromophore molecules, it can be supervised and registered by optical techniques. More specifically, these modifications could be detected through UV– Visible absorption spectroscopy, as the monitored compounds are active in this area of the electromagnetic spectrum. In fact, several methods that rely on this technique to control the degradation of EVOO have been already reported: the determination of the thermal degradation of olive oil [11], the analysis of the compositional changes that take place during the storage of EVOO [10], the study of the kinetics that govern the thermal degradation of the pigments in olive oil [6], or even the objective authentication of monovarietal EVOOs [12]; thus showing that the employment of this spectroscopic technique is useful to determine compositional and quality-related changes in EVOO over time due to the degradation of its pigments. On the other hand, using UV–Visible absorption spectroscopy provides a relevant amount of data points, making necessary the use of mathematical and statistical procedures to process and adequate the data, such as principal components analysis (PCA), linear discriminant analysis (LDA), or soft independent modeling of class analogy (SIMCA) [13–15]. Nevertheless, sometimes these linear procedures do not offer the expected results and more sophisticated models are required. One of these advanced mathematical tools are artificial neural networks (ANNs), non-linear models which can determine relationships between independent and dependent variables within databases through non-linear interpolation [16]. They have been already employed in the study of olive oil and the olive sector, offering accurate estimations [17,18]. As the EVOO is photodegraded, it loses its organoleptic characteristics as well as quality grade, so the product that is provided to the consumers may not possess the initial nutritional nor healthy effects mentioned before if the product is not handled adequately. Therefore, the goal of this work is to offer a userfriendly and straightforward procedure to estimate the photodegradation time of four monovarietal EVOOs (Arbequina, Hojiblanca, Cornicabra, and Picual) and their binary mixtures by measuring their UV–vis absorption spectra and using ANNs to
The samples and a reference (stored in the dark) were daily measured five times per week (Monday through Friday) for a period of 45 days, in order to monitor the changes experimented by the pigment profile due to the presence of sunlight, thus evaluating the extent of the photodegradation suffered by the different pigments of the samples of EVOO. A total of 30 data points per sample were obtained. The absorption spectra of the samples were collected between 350 and 800 nm using a UV–vis spectrophotometer (Varian Cary 50 Conc). The samples were introduced in a quartz cuvette of 1 cm path length and the samples were analyzed three times in order to obtain an averaged spectrum of each one. This led to spectra containing one absorption value per wavelength, which originates 450 potential variables. The first phase of the mathematical treatment of the spectra was reducing the dimensionality of the system. This approach enables the elimination of redundant and/or non-informative variables (absorbance–wavelength pairs). In the case considered, the calculation of the areas under the curve (AUCs) of the different absorption bands observed was the procedure selected to both reduce the number of independent variables and extract the useful spectroscopic information from the bands, as a combination of AUCs might contain more information than absorption peaks at discrete wavelengths [19]. The mathematical treatment of the data was performed with OriginPro 8.0 and Matlab 2013 software packages. 2.3. Database The final database contains a total of 658 spectra (after discarding two defective measurements which qualified as outliers in the preliminary analysis), corresponding to the measurement of the different mixtures of EVOO over time. The information gathered within the database possesses compositional, spectroscopic, and chronological data regarding the different samples analyzed. 2.4. Artificial neural networks ANNs were the selected mathematical tool to assist in the estimation of the degradation time of the samples. These algorithms are able to determine non-linear relations between dependent and independent variables within databases through non-linear
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
365
data distributions to train MLPs [19,20]. The training phase takes place firstly, where the training dataset is processed by the model. The MLP performs the estimation and calculates the error attained (comparing the estimated response with the actual target values). In order to decrease this error value, the weights are modified towards obtaining the lowest possible error [16,19,20]. If this process keeps going on, the error would potentially reach a value of zero, as the weights are strictly improved to achieve a perfect fit to the training data available. However, following this training procedure will lead to the design of an over-fit model, unable to perform properly with new sets of data [22,23]. The verification stage is in charge of avoiding this phenomenon, and occurs just after the training step. In this case, the weight values remain fixed and the verification dataset is used to test the generalization capability of the model, or, in other words, the capability of the model to correctly perform accurately when facing external data [25]. Once this whole process has finished, a training cycle, or epoch, is completed. The accuracy achieved by the MLP is studied regarding the error obtained with the verification dataset, in terms of mean absolute error (MAE).
Fig. 3. Scheme of the MLP designed. AUCi corresponds to the different AUCs calculated.
interpolation. Even though there are different kinds of ANNs, the supervised multilayer perceptron (MLP) is one of the most successfully employed ANNs in scientific research [20,21], and it was the one selected to assist in the case proposed. As it is a supervised-learning ANN, there must be both input and target data so the MLP can be correctly trained and optimized [22]. Since the estimation mechanism of ANNs is based on non-linear interpolation, they strongly depend on the range of data covered by the database, because if the model were forced to extrapolate, the error attained in this case would increase [16,23]. Additionally, they highly depend on the statistical quality of the data available (in terms of low deviation and good signal-to-noise ratio) to perform accurate estimations [23]. MLPs are formed by units that are arranged in layers. There are three different types of layers: the input, the hidden, and the output layers. The first one is solely formed by nodes, which only introduce the value of the independent variables into the network. The hidden and output layers contain neurons, the actual calculation centers of the system, (Fig. 3). The quantity of neurons that occupy the output layer depends on the number of dependent variables considered in the system modeled, while the number of units in the hidden layer requires further optimization (vide infra) [21]. On the other hand, as the input nodes present the value of the independent variables to the model, there must be as many nodes as independent variables. Every unit in the MLP is connected to every unit in the neighboring layers. The importance of these connections is controlled by weighted parameters, named weights. They are essential for the final performance of the model, and their value must be optimized to attain a low error or high accuracy in the estimations attempted [24]. 2.4.1. ANN optimization In order to be able to appropriately operate, ANNs require the optimization of their weights [19,20,23]. This process comprises two distinct stages: the first one is known as the training phase, and the second one, the verification phase. To correctly fulfill these two steps, the database must be previously randomly divided: 85% of the data is used in the training stage (training dataset), while the remaining 15% is employed in the verification stage (verification dataset). This is one of the most typical training–verification
MAE =
1 N
N
∑
rk − yk
k=1
(1)
In Eq. (1), n stands for the number of data points of the verification dataset, rk is the already known real value, and yk is the response given by the model. Unlike the training error, the verification error will eventually start rising. In that moment, the model can be thought as optimized and the training cycles conclude [25]. In addition, there are considerations that must be taken into account before thinking of the designed MLP as fully optimized. They are the training and transfer functions, the hidden neuron number (HNN), the learning coefficient (Lc), the learning coefficient increase (Lci), and the learning coefficient decrease (Lcd). The training function must be carefully selected in order to obtain accurate and reliable estimations, as it is in charge of the learning process. Depending on the training function chosen, the training algorithm will vary and, thus, offer a different estimation performance and accuracy [16]. Also, the transfer function should be fixed. It provides the ability to carry out the non-linear interpolations as well as limiting the range of the results emitted by a neuron in order to make results comparable within the MLP. In this case, the selected transfer function was the sigmoid function (Eq. (2)), which normalized the results between 0 and 1.
⎛ ⎞ 1 ⎟ yk = ⎜ ⎝ 1 + e− x k ⎠
(2)
where xk is the result reached by a neuron and yk the solution of the sigmoid transfer function [26]. On the other hand, the HNN must be also optimized to obtain the combination of neurons that provides the lowest estimation error. The quantity of these neurons is extremely important for the final performance of the model: if the HNN employed is too high, it can produce an over-fit network; but if the number is too low, it would lead to a decrease in the learning capability of the model [16,19]. The procedure chosen for the optimization of the HNN was a trial-and-error or heuristic method, considering that the number of weights must be the lowest possible, and always smaller than the number of data points available. Finally, the optimization process must contemplate the different Lcs. The Lc parameter is responsible for the modification of the weights, while the Lcd and Lci parameters control the extent of the variation of Lc [16]. In order to study their combined effect and determining the optimum combination of these coefficients, they
366
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
were optimized following an experimental design based on the “Box–Wilson Central Composite Design 23 þstar points”. The ranges studied for the different coefficients were from 0.001 to 1 in the case of Lc and Lcd, while Lci was studied between 2 and 100 [19]. The design of the ANN-based model developed, as well as its optimization process, was carried out employing Matlab 7.0 and Statgraphics Centurion XVI software packages. 2.4.2. ANN validation In addition to the tests performed during the selection of the HNN, it is vital to check and ensure that the model operates properly throughout the whole range considered by the database, and that it can generalize well and not result in an over-fitting model. To do so, a statistical procedure known as k-fold crossvalidation test was performed. This method is based on the division of the whole database in k datasets which are employed sequentially as verification dataset while the remaining k–1 groups are used as training dataset [27]. Therefore, every data point available is both used in the training and verification datasets in different moments, so the model can be tested throughout the whole range considered by the database.
3. Results and discussion In this section, the mathematical processing of the spectra, the final MLP model designed, and the results achieved throughout the optimization process are reported. 3.1. Analysis of the spectra To begin analyzing the absorption spectra, a dimensionality reduction phase was carried out to eliminate redundant and/or unrepresentative information. To do so, the AUCs of the five main absorption bands have been selected as input information because these magnitudes have been successfully used previously in the same kind of models with positive results [19]. The AUCs were calculated from 440 to 472 nm (AUC1), 473 to 526 nm (AUC2), 527 to 551 nm, (AUC3), 593 to 630 nm (AUC4), and 631 to 703 nm (AUC5), so the bands corresponding to the absorption of the pigments were individually selected to avoid any overlapping (Fig. 4)
Fig. 4. AUC calculation limits represented in the UV–vis absorption spectra of three different EVOO samples: non-degraded Hojiblanca monovarietal EVOO (dashes), 15-days degraded Hojiblanca EVOO (dots), and a non-degraded binary mixture 50% Arbequina, 50% Picual EVOOs (line).
[12,28]. It can be observed that the spectra shown in Fig. 4 can be distinguished, both in terms of photodegradation time and composition. In the case of the Hojiblanca monovarietal EVOO sample, a decrease in the absorption bands due to the photodegradation of the pigments over time (both chlorophylls and carotenoids) can be attributed mainly to the demetalation and succeeding degradation of chlorophylls, which would then act as singlet-oxygen sensitizers, accelerating the photodegradation and photooxidation of the rest of pigments and antioxidants that form the minor fraction of EVOO, and thus decreasing the nutritional value and quality of the oil due to the oxidation of free fatty acids and other compounds [8]. However, it must be noted that the degradation of the samples might be also affected by the presence of oxygen in the head-space of their vials (constant for every sample), as it contributes to the oxidative processes [8,29]. It is possible to link the decrease in absorption to photodegradation, as the reference stored in the dark maintained a stable spectrum with time. The AUCs calculated have been used as input information in the ANN model designed with the intention of determining the degradation time (sunlight exposure) that each sample has suffered, regardless of its composition or olive oil varietal. The remaining spectra, corresponding to the different measurements of the samples analyzed, can be found in the Supplementary Information section. 3.2. ANN optimization As stated above, in addition to the optimization of the value of the weights that take place during the training phase, several parameters (HNN, Lc, Lcd, and Lci) and the training function were optimized. All the results, error rates, and correlation coefficients shown have been calculated from the verification datasets, not the training datasets, as the last ones are supposed to perfectly fit the target values and provide no information regarding the actual behavior and performance of the MLP. In other words, the verification datasets also acted as prediction datasets, as they were employed to evaluate the statistical performance of the non-linear models. In order to design a functional MLP, able to accurately estimate the photodegradation time suffered by EVOO samples through the value of the AUCs of the spectroscopic bands corresponding to the absorption of the pigments, the considerations specified before were assessed. The training function selected was the Levenberg– Marquardt training function (TrainLM), as it provides good results when dealing with a high number of data points. It possesses, in addition, a memory reduction feature that permits it to work effectively even when the computational requirements are elevated. This algorithm has already been applied with success in similar cases in the past [16,30]. Once the training function is selected, different HNN were tested, from 9 to 30 neurons, to determine the combination which leads to achieve the lowest estimation error. The attained results are summarized in Table 1. It can be seen that the error decreased when the HNN was tested from 15 to 21 neurons, while R2 followed a more irregular trend. However, from this point, the error rose again and the R2 slightly decreased. In light of these results, the HNN selected was 18, as it was the combination that provided the lowest error and the highest R2 values. Although the number of data points would allow the use of a higher HNN (up to 90 hidden neurons), the trend found from 21 to 30 neurons suggests that the subsequent tests would not improve the already attained results. Therefore, the following optimization steps were carried out with a HNN of 18. Finally, to improve the network performance, the training
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
thus enabling the control of the quality of EVOO when it arrives to the consumers and compare it to its initial properties. This would allow the study of the different photo-induced phenomena that occurs during the different stages prior to the EVOO sale, such as transportion and storage to assess how the product has been handled during this entire distribution chain.
Table 1 Results attained in the HNN tests, in terms of MAE (in days) and R2. The option selected for the model designed appears in bold. HNN
MAE (days)
R2
9 10 12 15 16 17 18 19 20 21 25 30
3.24 2.74 3.02 2.87 2.72 2.95 2.55 2.81 3.09 3.06 2.72 2.72
0.89 0.91 0.90 0.90 0.91 0.90 0.92 0.90 0.90 0.89 0.91 0.91
367
4. Conclusions
Table 2 Statistical results of the performance of the model when the k-fold cross-validation tests are performed. K
MAE (days)
R2
1 2 3 4 5 6 Average
2.55 2.51 2.32 2.46 2.59 2.13 2.43
0.92 0.92 0.95 0.94 0.92 0.96 0.94
parameters Lc, Lcd, and Lci were optimized following the procedure mentioned above. The values of the optimum combination found were 0.001 for Lc, 0.001 for Lcd and 2 in the case of Lci, and the error attained when they were applied was the lowest, 5%, which implies an error of 2.43 days in the estimation of time of degradation. These statistical performances, alike all the other optimization steps, were obtained with the prediction dataset, and not the training dataset, as these samples are used to solely optimize the weights. 3.3. Validation of the network As noted above, the generalization capability of the MLP designed was tested following a k-fold cross-validation statistical procedure, with k ¼6, allowing to determine the true applicability of the final model in the whole range of data, instead of limiting this test to a localized, randomly selected verification set. The statistical results can be seen in Table 2. The final model designed offered an error of 2.43 days in the estimation of the degradation time of the samples of EVOO analyzed, which represents a 5.3% of error regarding the whole time span analyzed (45 days). In addition, the R2 coefficient shows a high value (0.94), confirming a suitable performance of the model. These results show that ANNs-based models, such as the one here proposed, are able to accurately and reliably aid in the determination of the photodegradation suffered by single and binary mixtures of monovarietal EVOO samples (in terms of degradation time), starting from spectroscopic information regarding their pigment profile (chlorophyll, pheophytin, and carotenoid content). Therefore, the procedure proposed permits the evaluation of the photodegradation process of EVOOs by monitoring the modifications of the UV–vis spectra of the main pigments of this oil when exposed to the light. Also, a user-friendly, reliable, and straightforward chemometric tool able to control the changes and modifications suffered by the final product has been developed,
The estimation of the photodegradation process using binary mixtures of four monovarietal EVOOs has been achieved through the analysis of the UV–vis absorption spectra and the mathematical support of an ANN-based model. The AUC of the absorption bands which correspond to the chlorophyll and carotenoid pigments have been obtained and employed by the optimized MLP to determine the photodegradation time suffered by the different samples, attaining an average error rate of 5.3% during a 45-day period of time. The accuracy attained suggests that the proposed tool could be employed to control the modifications suffered by EVOO since its production in the olive mill until its sale, analyzing the loss of quality and pigment photodegradation that occurs during the transport, storage, and shelf stages. Although this distribution time is in occasions greater than 45 days, a proof-ofconcept has been set, and potentially larger time scales can also be accurately monitored. Therefore, this procedure would allow determining the influence of each step in the EVOO quality and photodegradation extent, as well as the improvement of the strategies and methods designed to preserve the quality of the oil so the final consumer can enjoy the expected high quality product.
Appendix A. Supplementary material Supplementary data associated with this article can be found in the online version at doi:10.1016/j.talanta.2015.06.042.
References [1] I. Vasilescu, S.A.V. Eremia, C. Albu, A. Radoi, S.C. Litescu, G.L. Radu, Food Chem. 166 (2015) 324–329. [2] S. Dabbou, F. Brahmi, A. Taamali, M. Issaoui, Y. Ouni, M. Braham, M. Zarrouk, M. Hammami, J. Am. Oil Chem. Soc. 87 (2010) 1199–1209. [3] M.N. Franco, T. Galeano-Díaz, Ó. López, J.G. Fernández-Bolaños, J. Sánchez, C. De Miguel, M.V. Gil, D. Martín-Vertedor, Food Chem. 163 (2014) 289-289. [4] A. Fernández-Gutiérrez, A. Segura-Carretero, El. Aceite de Oliva Virgen Tesoro de Andalucía, Fundación Unicaja, 2010, p. 17. [5] R. Aparicio-Ruiz, M.I. Mínguez-Mosquera, B. Gandul-Rojas, J. Food Compos. Anal. 24 (2011) 811–820. [6] R. Aparicio-Ruiz, M.I. Mínguez-Mosquera, B. Gandul-Rojas, J. Agric. Food Chem 58 (2010) 6200–6208. [7] Y. Saga, H. Tamiaki, Chem. Biodivers. 9 (2012) 1659–1683. [8] E. Psomiadou, M. Tsimidou, J. Agric. Food Chem. 50 (2002) 716–721. [9] E. Psomiadou, M. Tsimidou, J. Agric. Food Chem. 50 (2002) 722–727. [10] J.R. Morelló, M.J. Motilva, M.J. Tovar, M.P. Romero, Food Chem. 85 (2004) 357–364. [11] I.P. Paz-Antolín, M. Molero-Meneses, Grasas Aceites 51 (2000) 424–428. [12] M.I. Mínguez-Mosquera, L. Rejano-Navarro, B. Gandul-Rojas, A.H. SánchezGómez, J. Garrido-Fernandez, J. Am. Oil Chem. Soc. 68 (1991) 332–336. [13] C. Pizarro, S. Rodríguez-Tejedor, N. Pérez-del-Notario, J.M. González-Saiz, J. Chormatogr. A 1218 (2011) 518–523. [14] N. Sinelli, M. Casale, V. DiEgidio, P. Oliveri, D. Bassi, D. Tura, E. Casiraghi, Food Res. Int. 43 (2010) 2126–2131. [15] R. Bucci, A.D. Magrí, A.L. Magrí, D. Marini, F. Marini, J. Agric. Food Chem. 50 (2002) 413–418. [16] H. Demuth, M. Beale, M. Hagan, Neural Network Toolbox for Use with MATLABs User’s Guide, Version 5, Ninth printing Revised for Version 5.1 (Release 2007b), Natick, MA, USA, 2007 (online only). [17] J.S. Torrecilla, J.C. Cancilla, G. Matute, P. Díaz-Rodríguez, Int. J. Food Sci. Technol. 48 (2013) 2528–2534. [18] J.S. Torrecilla, J.C. Cancilla, G. Matute, P. Díaz-Rodríguez, A.I. Flores, J. Food Eng. 118 (2013) 400–405.
368
J.S. Torrecilla et al. / Talanta 144 (2015) 363–368
[19] P. Díaz-Rodríguez, J.C. Cancilla, K. Wierzchoś, J.S. Torrecilla, Sensor Actuat. B— Chem. 206 (2015) 139–145. [20] J.S. Torrecilla, J.C. Cancilla, G. Matute, P. Díaz-Rodríguez, Int. J. Food Sci. Technol. 48 (2013) 2528–2534. [21] S. Curteanu, H. Cartwright, J. Chemometr. 25 (2011) 527–549. [22] J.S. Torrecilla, F. Rodríguez, J.L. Bravo, G. Rothenberg, K.R. Seddon, I. LópezMartin, Phys. Chem. Chem. Phys. 10 (2008) 5826–5831. [23] L. Cakir, N. Yilmaz, Measurement 57 (2014) 148–153. [24] J. De-Ona, C. Garrido, Neural Comput. Appl. 25 (2014) 859–869. [25] J.S. Torrecilla, C. Tortuero, J.C. Cancilla, P. Díaz-Rodríguez, Talanta 116 (2013) 122–126. [26] K. Knoerzer, P. Juliano, P. Roupas, C. Versteeg, Innovative Food Processing
[27] [28] [29] [30]
Technologies: Advances in Multiphysics Simulation, Wiley-Blackwell, Oxford, UK, 2011. M. Zounemat-Kermani, O. Kisi, T. Rajaee, Appl. Soft Comput. 13 (2013) 4633–4644. V. Domenici, D. Ancora, M. Cifelli, A. Serani, C.A. Veracini, M. Zandomeneghi, J. Agric. Food Chem. 62 (2014) 9317–9325. R. Aparicio-Ruiz, M. Roca, B. Gandul-Rojas, J. Agric. Food Chem. 60 (2012) 7040–7049. A. Bassam, R.A. Conde-Gutiérrez, J. Castillo, G. Laredo, J.A. Hernández, Fuel 124 (2014) 158–167.