Introduction

Olive (Olea europaea sp. europaea var. sativa) is a historically capital crop in the Mediterranean Basin [19]. Using automated or concrete methods, virgin olive oil is the merely edible oil of great production in the worldwide [44]. Chemically, olive oil has several constituents with antioxidant characteristics, especially polyunsaturated fatty acids, carotenoids, tocopherols, chlorophylls, and other phenolic complexes [61]. The most significant categories of phenolic compounds present in olive fruits comprise phenolic alcohols (hydroxytyrosol and tyrosol), flavonoids (luteolin, apigenin, and methyl luteolin), sterols (e.g., stigmasterol, β-sitosterol, and sitostanol), and phenolic acids (e.g., o- and p-coumaric acid, ferulic acid and cinnamic acid, and dihydroxybenzoic derivatives such as vanillic acid and vanillin) [51, 64].

Olive oil has been extensively studied for its benefits on coronary heart diseases, principally for its aptitude to decrease low-density lipoprotein (LDL) cholesterol levels and blood pressure [74]. Other health advantages of olive oil consumption described in numerous studies comprise decreasing probability of developing several diseases including inflammatory ones such as atherosclerosis [58] and asthma [79], cardiovascular diseases and hypertension [57] as well as metabolic syndromes such as diabetes [32] and obesity [11] of olive oil. Moreover, many works have shown that olive oil compounds are characterized by the ability to fight cancer cells [40, 60]. The antioxidants present in olive oil can diminish oxidative damage due to free radicals, which is supposed to be one of the leading causes of cancer [54]. Studies suggest that olive oil can fight several age-related diseases such as Alzheimer’s disease [2] and Parkinson’s disease [50]. Furthermore, olive oil can help treat rheumatoid arthritis, olive oil supplementation shows to increase inflammatory markers and decrease oxidative stress in individuals having rheumatoid arthritis [43].

Olive crop growing, extensive throughout the Mediterranean Basin area, is essential for the local tradition, rural market, and environment [80]. In North Africa and especially Tunisia, the olive oil segment plays a considerable role in the economy, offering employment and also exports profits [16]. Olive trees cover a wide region of 1,611,200 ha with an annual oil production of 170,000 tons [33].

In Tunisia, the oleoculture depends principally on two varieties: Chetoui and Chemlali [10]. In addition, there are also other secondary local varieties specific to minor regions such as “Sayali” in the northern regions, “Oueslati” in Kairouan, “Zalmati”, and “Zarrazi” in Zarzis [8].

In Tunisia, olive orchards are characterized by the cultivar population Chemlali planted in the north, the Sahel, the centre and the south covering approximately 80% of the Tunisian olive grove [10]. Similarly, Chemlali cultivar endures an important lack of rainfall unhelpful for growth and reproductive expansion in cultivated dry areas [16]. The presence of different factors makes difficult to characterize olive oils using a small number of chemical compounds or a simple data manipulation. Therefore, samples should be identified for their chemical profiles such as: major components (fatty acid), minor components, and/or sensory descriptors, and data should be analyzed by statistical techniques and artificial intelligence algorithms.

The bootstrap is a computational method that employs accurate resampling with replacement, in the aim to reduce uncertainty [22]. In fact, the bootstrap approach is constructed on statistical resampling procedure, to generate varied training sets that are used to train the members making a group [22]. This practice is replicated and every network member is produced with a different arbitrary sampling of the original training set. It is imperative to mention that when resorting to bootstrap sampling, mainly on a relative little data size [42], it is essential that all combinations of the networks be well trained, to avoid that some biased network components of the bootstrap ensemble noticeably influence the estimate values [6, 23, 26]. In the bootstrap method of uncertainty examination, the total example set is splitted into two various sets: validation sets and training sets. In addition, evaluating these data is difficult with univariate statistical method, such as one-way ANOVA and t test, multivariate statistical methods are applied to numerous variables to obtain significant interpretation [47]. These statistical techniques, such as multiple linear regressions (MLR) and artificial neural network (ANN), could be applied to the classification of olive oils with respect to their different characteristics such as geographical origin or olive cultivar using various spectroscopic or chemical parameters. Eventually, this study examines the effect of growing areas on numerous quality characteristics of Chemlali and Zalmati olive oil cultivars based on different parameters such as oxidative stability, fatty acids and on minor components such as phenolic acids and sterols. Ultimately, this study allowed us to establish many relationships between oxidative stability and fatty acids, phenolic acids, and sterols by investigative statistical techniques.

Materials and methods

Olive oil samples

Olive fruits were collected during three successive harvests (from 2014 to 2016), from entirely ripened olive trees belonging to several Tunisian cultivars: Chemlali Sfax (N: 34.444°, E: 10.451°), Chemlali Medenine and Zalmati Medenine (N: 33.524°, E: 10.056°). After harvesting, olive fruits were immediately transported to laboratory, where oil was extracted within 30 h by extraction equipment (oleodoseur). The extracted oil was conserved into dark glass bottles and saved at 4 °C to later analyses.

Quality parameter: oxidative stability

Oxidative stability was estimated in term of hour by Rancimat (Metrohm, Switzerland). The temperature was diversified between 50 and 220 °C with temperature stability less than 0.1 °C. Five grams of experiment were put into the glass reaction vessel for the analysis. For equipment columns, the reaction temperature and air-flow phases were kept constant at 100 °C in air flow of 20 l/h as defined by Uncu and Ozen [76]. All measurements were performed in triplicate.

Fatty acid methyl ester analysis

Fatty acid methyl ester (FAMEs) was prepared as defined by the European Union standard method [29]. FAMEs had been prepared by melting vigorously a 0.2 g of studied oil in 3 ml of hexane and 0.4 mL of methanolic potassium hydroxide at 2 N, and analyzed by gas chromatography (Shimadzu) with a fused silica column made with 50% cyanopropylmethyl and 50% phenylmethyl-polysiloxane. 1 μl was injected using nitrogen as carrier gas with a flow rate of 1 ml/min. The injector and detector temperatures were anchored at 220 °C, while the oven temperature was held to 180 °C. Ten fatty acids including palmitic acid (C16:0), palmitoleic acid (C16:1), margaric acid (C17:0), margaroleic acid (C17:1), stearic acid (C18:0), oleic acid (C18:1), linoleic acid (C18:2), linolenic acid (C18:3), arachidic acid (C20:0), and eicosenoic acid (C20:1) were recognized from the retention intervals. Each sample was analyzed in triplicate.

Phenolic profiles of olive oils

Phenolic extraction

Fourteen grams of each olive experiment were extracted by employing 4 × 14 ml of methanol/water (80:20 v/v). Then, the methanol was discarted, and 15 ml of acetonitrile were included to the residue and rinsed with hexane (3 × 20 ml). The resulting acetonitrile solution was evaporated, flushed with nitrogen, and dissolved in 1 ml of methanol/water [63]. Final extract was refined through a 0.45 μm pore-size membrane filter (MF-Millipore®, Merck). 20 μl of extract was instantly filled to HPLC and the gallic acid (Sigma Chemical Company, St Louis, MO) was employed as internal reference [53]. Sampling was done in triplicate.

HPLC analysis of phenolic acids

To investigate phenolic acids, an HPLC system (JASCO) was employed. Preliminary concentrations of the dynamic phases were 90% for water/acetic acid and 10% for MeOH and the concentrations had been adjusted according to the following procedure: in the beginning, the concentration of MeOH was increased to 30% in 10 min and kept for 20 min. Next, MeOH% was enhanced to 40% in 10 min, and kept for another 5 min. Followed by increasing it up to 50% in 5 min, and preserving it at this concentration for another 5 min, MeOH was enhanced to 60, 70 and 100% in 5 min. At last, primary conditions were achieved at the end of 85 min. Chromatograms were attained at 280 and 320 nm and diverse phenolic acids were identified by comparing their retention time with those of reference. Phenolic acids had been evaluated by employing their respective five-point calibration curve and the results were determined in term of mg/kg. All samples were performed in triplicate.

Sterol determination

The essay of sterolic parts extracted from olive oil was characterized by a modified method used by Abood et al. [1] and in agreement with the European Union Commission Regulation EEC 2568/91 [53]. Five grams of olive oil from each variety was initially joined to α-cholestanol advised as internal standard and then saponified with potassium hydroxide solution in ethanol at 2 M. After 1 h of swelling, 100 ml of water was joined and the extraction of unsaponifiable was carried-out by 200 ml of ethyl ether. Subsequently, 20 mg of unsaponifiable fraction was melted in 0.5 ml of chloroform and again eliminated by silica gel plate chromatography. The elution was completed by a blend of hexane and ether (65/35, v/v); then, the plate was crushed by 2,7-dichlorofluorescein (0.2% in the ethanol). The band corresponded to sterols was afterwards fragmented. Sterols recovered from the plate were dissolved in chloroform and purified. The solvent was vanished under N2 and the sterols were adapted into trimethylsilyl ethers by the accumulation of pyridine–hexamethyldisilzane–trimethylchlorosilane (9:3:1, v/v/v), left for 15 min and subsequently centrifuged. The assortment was then analyzed with the aid of gas chromatography via a gas chromatograph system (model 7890A, Agilent Technologies) with flame ionization detector. The employed column was a capillary HP-5 (5% phenyl: 95% dimethylpolysiloxane), and the operating circumstances for the analysis were as follows: the oven temperature was isothermal and constant at 260 °C, while injector and detector temperatures were maintained at 280 °C and 290 °C, respectively. Helium was used as carrier gas, with a flow rate through the column of 1 ml/min and 1:50 Split ratio. The injection amount was set to 5 µl and samples were tested in triplicate.

Statistical analysis

All factors were evaluated in triplicate for each trial. Analysis of variance (ANOVA) was done by Statistical Package for the Social Sciences (SPSS) software (Version 19.0 SPSS Ltd). The significance of differences at 5% level between means was established by one-way ANOVA, using Tukey’s test. Analysis of variance (ANOVA) was employed to assess the potency of growing region conditions on Tunisian olive oil cultivars.

Linear regression model

The general purpose of multiple regressions is to learn more about the relationship between several independent or predicted variables and a dependent or criterion variable [28, 48]. The general form of the regression equations is

$$Y = A_{0} + A_{1} X_{1} + A_{2} X_{2} + \cdots + A_{n} X_{n} ,$$

where Y is the dependent variable, A0 is the intercept, A1…An are regression coefficients, and X1–Xn are independent variables referring to measured indices. Determination coefficient (R2) was used as predictive success criteria for regression model [26, 28]. To establish prediction equations, the multiple regression analyses were conducted (Minitab Inc., State College, PA).

Neural networks

The artificial neural network (ANN) can be defined as a machine-learning forecast algorithm that is similar to human nervous system. ANN adapted from a theoretical way to the widely employed knowledge with useful appliances to diverse problems [24, 70, 77]. The most important of these appliances are applied to the prediction and ANN are frequently employed as non-linear estimators that reveal an elevated tolerance to inaccuracy and achieve well despite noisy situations [49]. Those advantages also make ANN suitable for biotechnological problems, and principally for food science [13]. The most employed representation in ANN uses is the multilayer perceptron (MLP) because of a high ability to estimate the functional construction in the non-linear structures. MLP is composed of an input layer, output layer, and hidden layers [55]. Variables that were related to the oxidative stability were selected as candidates for input into the final ANN model. For statistical analysis, an exploratory three-layer multiplayer perceptron (MLP) ANN model with a back-propagation algorithm was assembled [5, 27]. The multilayer perceptron (MLP) algorithm is an influential structure of an artificial neural network that is usually employed for regression, and also classification. The MLP algorithm is an excellent algorithm to apply for the mapping and regression. It can be employed to transform an N-dimensional input indicator to an M-dimensional output indicator; this transformation can as well be non-linear. The validation data are used only when the predictor model is organized and are dissimilar from those used for training data. In general, 80% of all experimental data are reserved for training, and the remaining 20% are used for validation. The ANN model is tested on the bootstrapped training set with unchanged initial weights; the remaining samples in the training set independently from bootstrapped training set are used for split test validation so as to avoid any over-fitting. Only one node was used in the output layer, corresponding to the oxidative stability. The hidden layer consisted of several nodes ranged from 5 to 20. The feed-forward topology was implemented in all organizations. In this study, K-fold cross-validation procedure was used for the validation and testing of ANN model, consisting on dividing the original data into K subsets. In turn, each of the K sets is used to validate the model fit on the rest of the data, fitting a total of K models. Cross validation is a statistical technique employed to test the ability of machine-learning designs. It is generally employed in applied machine learning to evaluate and choose a model for a specified predictive modeling complication for the reason that it is simple to understand, simple to execute, and results in ability evaluates that usually have a minor bias when compared with other techniques [81]. Besides, cross validation is known as a resampling procedure employed to calculate machine-learning prototypes. The procedure has a unique factor called k that is attributed to the groups that a specified data example is to be contained into [7]. Then, the procedure is frequently named k-fold cross validation [78]. Cross validation is a very powerful device. It aids us better employ the data, and it provides us much more details about the algorithm performance. Values of R2 from training and validation data were used as selection criteria for the best predictor model. The neural module from the Statistical Analysis System (SAS) package (version 9; SAS Institute Inc., Cary, NC) was used for training and validation of the networks.

Results

Oxidative stability

The oxidative stability is a major element in the characterization of vegetable oils [12]. In fact, this parameter is a quality index to confirm the preservation status of the oils as well as their predictive resistance to oxidative phenomena [12]. Oxidative stability of all olive oils tested is presented in Table 1.

Table 1 Oxidative stability and fatty acids composition of Chemlali Sfax, Chemlali Medenine, and Zalmati Medenine virgin olive oils
Full size table

It was evident that olive oils belonging to Zalmati Medenine cultivar have the highest average with 10.14 h. Statistically, olive oils from Chemlali Sfax had lower (P < 0.05) oxidative stability in comparison with samples of Chemlali Medenine and Zalmati Medenine (Table 1). In general, the oxidative stability provides precious information on oil’s hypothetical shelf life; similarly, the resistance to oxidative deterioration is usually endorsed to the lipid composition and the antioxidant content [39, 52].

Fatty acid composition

The specific fatty acid composition in olive oil varies depending on various factors such as olive cultivar, latitude, climate, and maturity stage [45, 67]. On the other hand, the composition of fatty acids influences the stability, nutritional value and is characteristic of olive oils [20, 36]. In this study, fatty acid estimation was besides performed on the olive oils, following the common product analyses. Table 1 reports the identified fatty acids: palmitic (C16:0), palmitoleic (C16:1), margaric (C17:0), margaroleic (C17:1), stearic (C18:0), oleic (C18:1), linoleic (C18:2), linolenic (C18:3), arachidic (C20:0), and eicosenoic acid (C20:1). In the entire experiments, the oleic acid is continually the most abundant. In addition, a considerable difference (P < 0.05) was discovered among oleic acid in Chemlali Sfax, Chemlali Medenine, and Zalmati Medenine olive oils. Similarly, the oleic acid mean values for the Chemlali Medenine cultivar had the best average (Table 1). This result is consistent with the finding of Servili et al. [68] who found that fatty acid composition could be related to the geographical location, the characteristics of the olive grove zones, and the salinity of soil. Likewise, our results were similar with those of Temime et al. [73] who suggested the importance of the environment parameter in defining the oil quality of the Tunisian Chetoui cultivar.

Phenolic composition

Total phenol content

As shown in Table 2, the results of the colorimetric determination of total phenol content in oil extracts were ranged between 187.86 ± 2.98 mg of gallic acid equivalent per kg (GAE/kg) for Chemlali Medenine and 266.68 ± 5.76 mg of gallic acid equivalent per kg (GAE/kg) for Chemlali Sfax. Moreover, no significant difference (P > 0.05) was detected in total phenol content between Chemlali Medenine and Zalmati Medenine olive oils (Table 2). These results are in agreement with many previous reports demonstrating that quantities of phenolic compounds in olive oil ranged from 50 to 1000 mg GAE/kg in function of geographical zone, cultivar origin, and oil extraction technique [9].

Table 2 Total phenols content and phenolic acids of Chemlali Sfax, Chemlali Medenine, and Zalmati Medenine virgin olive oils
Full size table

Phenolic acids profile

The foremost phenolic acids (P < 0.05) for olive oils collected from the Chemlali and Zalmati cultivars are: phenylacetic acid, p-hydroxyphenylacetic acid, protocatechuic acid, gallic acid, and trans-cinnamic acid (Table 2). Interestingly, Chemlali Medenine had essentially the most elevated intensities (P < 0.05) of phenylacetic acid (15.85 ± 0.37 ppm), p-hydroxyphenylacetic acid (17.82 ± 2.3 ppm), protocatechuic acid (7.52 ± 0.66 ppm), gallic acid (8.71 ± 0.25 ppm), and trans-cinnamic acid (7.33 ± 0.24 ppm).

Sterol composition

The sterol compositions of the Chemlali Sfax, Chemlali Medenine, and Zalmati Medenine oils are shown in Table 3.

Table 3 Sterol composition of Chemlali Sfax, Chemlali Medenine, and Zalmati Medenine virgin olive oils
Full size table

All experiments displayed high amounts of campesterol, with percentages ranging from 2.4 to 3.26%, which were in conformity among the limit (4%) established by European Union Regulations [18]. Another remarkable attribute of Chemlali and Zalmati varieties was their high Δ-5-avenasterol content, which ranged between 7.85 and 11.5% [75]. In fact, countless studies in the literature have connected this compound with antioxidant action 75]. Likewise, all olive oil experiments were distinguished to comprise greater than 1200 ppm of total sterols, which is the minimum value established by European Union Regulations with indication to the extra virgin olive oil class. All experiments had been additionally noted to exhibit expanded levels of β-sitosterol (84% of total sterols).

Linear regression analysis

In this section, the multiple linear regression analysis using the oxidative stability parameter as dependent variable was performed. Fatty acids values (C16:1, C17:1, C18:1, and C18:2), phenolic acids (hydroxyphenylacetic acid, o-coumaric acid, and gallic acid), sterols (campestanol, stigmasterol, and sitostanol) were considered as variables. The multiple linear regression analysis (Table 4A, B, C) revealed a high correlation (R2 > 62.7) between overall acceptability and all parameters.

Table 4 Multiple linear regression equations of oxidative stability versus measured parameters
Full size table

Relationship between oxidative stability and fatty acids

Concerning the regression model for oxidative stability determination of Chemlali Sfax, the studied fatty acids have positive effect, meaning an increase in the phenol contents, and lead to improve the oxidative stability (Table 4A). On the basis of these obtained results, we remark that margaroleic acid (C17:1) has the higher effect on oxidative stability, followed by stearic acid (C18:0), palmitoleic (C16:1), and linoleic (C18:2) acids with similar effects. Oleic acid (C18:1) has very low consequence on oxidative stability. On the other hand, the regression model for determination of oxidative stability of Chemlali Medenine olive oil is represented in the following equivalence in Table 4A. We observe that C17:1 has the higher effect on oxidative stability, followed by C16:1 and C18:0 acids with similar effects and finally by C18:2. C18:1 has a very low role on oxidative stability variation (Table 4A).

Relationship between oxidative stability and phenolic acids

The regression models for determination of oxidative stability were illustrated in the mentioned equality (Table 4B). The obtained results showed that Chemlali Medenine has the highest amount of p-hydroxyphenylacetic acid and the lowest amount of o-coumaric acid. In addition, p-hydroxyphenylacetic and gallic acids have a positive effect on improvement of oxidative stability. Moreover, these results describe that Zalmati Medenine has the lowest amount of gallic and p-hydroxyphenylacetic acids. All parameters have an active contribution on enhancement of oxidative stability. It is obvious that the effect of p-hydroxyphenylacetic acid is more important in the increase of oxidative stability of Chemlali Sfax olive oil, and this influence is less significant for Zalmati Medenine and Chemlali Medenine. Moreover, o-coumaric acid has more important effect in the oxidative stability of Chemlali Sfax when compared to Zalmati Medenine. Nevertheless, this phenol has no effect on oxidative stability of Chemlali Medenine. Finally, gallic acid has an essential role in the oxidative stability of Chemlali Sfax. This role is less intensive for oxidative stability of Chemlali Medenine and Zalmati Medenine.

Relationship between oxidative stability and sterols content

It can be seen, as shown in Table 4C, that the studied sterols have positive consequence, which signifies that an increase in the sterol contents leads to improve oxidative stability. In fact, Eussen et al. [30] reported that major phytosterols found usually in plants comprise sitosterol, stigmasterol, and campestanol. They have much stand out due to their nutritional value as natural components of regular diet. The regression model for determination of oxidative stability of Chemlali Medenine olive oil is shown in Table 4C. On the basis of regression model proposed for estimation of oxidative stability of Zalmati Medenine olive oil (Table 4C), it is obvious that all parameters have an active contribution on enhancement of oxidative stability.

ANN modeling of oxidative stability Analyses

Table 5 illustrates the coefficient of determination (R2) percentages for training and validation of established artificial neural network models. Two main quantitative measures were used to evaluate the modeling performance: R2 and cross validation R2. The cross validation R2 indicates the predictive ability of the model. The predictor model should be able to correctly estimate the other values, different from those used during adjustment. Data analysis of oxidative stability demonstrated that obtained ANN could be classified as a back-propagation feed-forward multilayer-perception neural net type. This neural network is considered by a set of non-linear equations that predict output variables (y) from input variables (x) in a flexible way using layers of linear regressions and S-shaped functions. From Table 5, Chemlali Sfax model (Model 1) had a better training performance R2 = 99.962% and validation performance R2 = 98.734%. Training data refer to the data set used to effectively adjust the predictor model to experimental data, whereas validation data refer to another set of data previously reserved and is not used to adjust the predictor model but to check if the model is properly tuned. In this model, configuration with five hidden nodes is illustrated in Fig. 1 and yielded the highest score.

Table 5 Coefficient of determination (R2) percentages for training and cross validation (CV R2) of established artificial neural network models
Full size table
Fig. 1
figure 1

Structure of artificial neural network (11-5-1) for predicting oxidative stability of Chemlali Sfax olive oil

Full size image

Moreover, Chemlali Medenine model (Model 2) had a better training performance R2 = 99.796% and validation performance R2 = 98.948%. Configuration with seven hidden nodes is represented in Fig. 2 and yielded the highest score.

Fig. 2
figure 2

Structure of artificial neural network (10-7-1) for predicting oxidative stability of Chemlali Medenine olive oil

Full size image

Thus, this model was considered the best one for prediction of oxidative stability based on olive oil components. Furthermore, Zalmati Medenine model (Model 3) had a better training performance R2 = 99.927% and validation performance R2 = 99.77%. Configuration with 4 hidden nodes illustrated in Fig. 3 and yielded the highest score. Thus, this model was considered the best predictor particularly for Zalmati Medenine.

Fig. 3
figure 3

Structure of artificial neural network (11-4-1) for predicting oxidative stability of Zalmati Medenine olive oil

Full size image

Discussion

Fatty acid composition

The differentiations detected between areas for the fatty acid composition might be proven by the variation in altitude and temperature between the chosen zones. In fact, region of Sfax is known as a coastal central province and Medenine is known as a south coastal province. This result is concordant with the finding of Pinelli et al. [59] who observed that fatty acid composition could be associated with geographical locality, the kind of the olive grove zones, and the salinity of soil. Similarly, obtained results were analogous with those of Kotti et al. [41] who proposed the effect of environment factor in defining the oil quality of Tunisian Chemlali cultivar. Truly, they showcase that virgin olive oils produced by Chemlali olive cultivar had been in most cases high in oleic acid.

Sterol composition

In the present study, the sterolic composition of three varieties of virgin olive oils had been compared: namely, Chemlali Sfax and Chemlali Medenine, which showed elevated oxidative stabilities, and Zalmati Medenine, which showed instable factors. Sterols have most commonly been suggested to compose essential factors entailed within the selection of olive oil quality and are frequently applied in olive oil authenticity assessments [62]. Actually, virgin olive oil is well recognized for its composition of minor contents that have health-promoting traits [56].

Relationship between oxidative stability and fatty acids

From established equations, we deduced that palmitoleic acid (C16:1), margaroleic acid (C17:1), and linoleic acid (C18:2) of Chemlali Medenine variety have the most important consequence on oxidative stability variation compared to Chemlali Sfax and Zalmati Medenine. Stearic acid (C18:0) of Chemlali Sfax variety is considered has the most important effect on oxidative stability when compared to two studied olive cultivars. It is noted that the geographical location enhances this difference with a preference for the south of Tunisia for major fatty acid contents. These differences could be explained by the diverse fatty acid composition of varieties grown in regions far apart. In fact, the geographical location had an influence on fatty acid composition despite the same Chemlali variety origin. The fatty acid values of olive oils present an extensive range of variability with genetic factors explaining from 70 to 80% of the differences between olive tree accessions [17]. These observations validate the obtained results with a significant difference (P < 0.05) between fatty acid composition of two different olive tree cultivars (Chemlali Medenine and Zalmati Medenine) grown in a common geographical area.

Relationship between oxidative stability and phenolic acids

It was demonstrated that coumaric acid has several biological actions such as antioxidant and radical scavenging activities, anti-inflammatory, and neuroprotective actions [38]. In this study, gallic acid was shown to have an essential role in the oxidative stability of Chemlali Sfax. This role is less important for oxidative stability of Chemlali Medenine and Zalmati Medenine. In agreement with our results, the previous studies reported that geographical location is tightly related to the variation of phenolic composition [3] and oxidative stability of olive oils [65]. It has been shown that phenolic compounds are more efficient than tocopherols in enhancing the stability of olive oil toward oxidation [65].

Relationship between oxidative stability and sterols content

On the basis of obtained results, the campestanol, stigmasterol, and sitostanol have a positive effect on improvement of oxidative stability. Virgin oil shows a very acceptable correlation between stability and sterols concentration [37]. On the basis of regression model proposed for estimation of oxidative stability of Zalmati Medenine olive oil (Table 4C), it is obvious that all parameters have an active contribution on enhancement of oxidative stability. The effect of campestanol appears to be more important in the oxidative stability enhancement of Chemlali Sfax and Chemlali Medenine olive oils when compared to Zalmati Medenine.

Furthermore, stigmasterol and sitostanol appear to be more effective in the oxidative stability of Chemlali Sfax compared to Zalmati Medenine and Chemlali Medenine. It should be observed that stigmasterol and sitostanol contribution in variation of oxidative stability is linked with geographical location. In fact, it appears that the center of Tunisia is more adequate for enhancement of stigmasterol and sitostanol. Ryan et al. [66] have shown that climatic factors such as temperature and precipitation are closely related with geographical location and have notable effect on plant physiology, and thus on vegetative components. Besides, the quality of this product can be established from geographical and varietal origins, which are immediately linked with olive oil distinctive characteristics [25].

ANN modeling of oxidative stability analyses

The model illustrated in Fig. 1 was characterized by the best ANN architecture and a more adequate structure. It presented high values of R2 for both training and validation data, as shown in Table 5. In fact, several studies deduced that a perfect model would result in a cross-validation R2 value of 1 [25, 69].

As shown in Table 5, all selected values presented high R2 (more than 98%) for both training and validation data despite the variation in the ANN architecture. This difference is located in the number of hidden nodes. The MLP architecture is a method supported on back-propagation algorithm, which is an extremely flexible and adjustable in designing and simulation [72]. It can be considered as a standardized algorithm for any supervised-learning pattern recognition process. However, it uses “black box” for classification that is very difficult to explain the performance of results in some cases [31].

The high coefficients of determination (R2) for the three established models at the training stage indicate a strong relationship between the output (oxidative stability) and the inputs (sterols, phenols, and fatty acids). Deciding the number of neurons in the hidden layer is a very important part of deciding the overall neural network architecture [46]. Hidden layers do not directly interact with the external environment, but still have a tremendous influence on the final output. This shows that both models have good predictive abilities and are suitable for predicting the oxidative stability for new predictions. On the other side, it is obvious that bootstrap technique has proved to be significantly useful and appropriate for predictors such as ANNs. In fact, Carney and Cunningham [14] studied bootstrap aggregation in the perspective of ANNs and deduced that model generalization aptitude can be considerably developed. It is the simplest approach, since it does not entail the complex calculation of derivatives and Hessian-matrix inversion implicated in linear techniques or the Monte Carlo solutions of the integrals concerned in the Bayesian method [21].

The MLR is an algorithmic usual model generally employed in biological sciences and technology [4, 71]. In comparison with other artificial intelligence approaches, MLR modeling offers a more clear vision to forecast [15]. ANNs are a category of non-linear numerical models that are distinguished by a complex construction of computational and interconnected components, the neurons. In comparison with MLR models, ANN provides a computational approach of deciding a non-linear connection between inputs and one or more outputs. ANN has employed for modeling, recognition, and forecast of complicated organizations [35]. In our study, ANNs as an alternative are proposed for predicting oil oxidative stability.

Our results indicated that one potential application of ANN may be evaluating the quality of MLR prediction models due to of their inherent flexibility [34], and ANN may perform better than MLR in many cases as shown in this study. In fact, the R2 values estimated from ANN method are higher than those from multiple linear regression method despite the number of parameters simultaneously implied in each analysis. Therefore, ANN technique may be a more appropriate tool when emphasis is put on the prediction itself and not on the underlying relations between independent and dependent variables.

Conclusion

Oxidative stability as well as fatty acid composition, phenolic acid, and sterols content of different Tunisian olive oils were investigated to study the differences among oils from different geographical locations. On the basis of the combination of ANOVA analysis and multiple linear regressions (MLR), different approaches were investigated to improve the geographical classification. This study demonstrates that fatty acid, phenolic acid, and sterol compositions were significantly (P < 0.05) influenced by both olive cultivar and geographical region. The findings showed also that phenolic acids content and sterols content displayed significant qualitative and quantitative differences among the cultivar and their geographical origins.

The models based on multiple linear regressions cannot predict oxidative stability of studied olive oils with similar accuracy as the obtained for the selected neural networks. Besides, the modeling based on artificial neural networks brings more advantages, because opens the possibility of the use of networks for prediction of more complex parameters. This approach gets easy doing artificial neural network models by connections between different variables and interpretation of obtained results by highlighting various effects of related explanatory variables.