The procedure for selection included a previous interview before the difference test addressed to the product (Dutcosky, 2007). Questions about allergies and availability for training were made, along with filling in the form required by the Ethics Committee.

The selection of assessors was performed through the triangle test, a modality of sensory analysis called discriminative, which differentiates two samples that received different treatments (ASTM, 2010). The probability of accuracy is one-third. It is recommended to use 20 to 40 subjects for a solid result (NÆS et al., 2010).

The test consisted of two samples: rancid and regular sunflower oil. The rancid oil was produced in an oven at 60 °C for 14 days (Houhg and Fiszman, 2005) with air circulation, within an open amber glass recipient, with 10% head space. Oils of the same brand were purchased at Pato Branco – PR local market. Three samples in random order were presented to the assessors, ten different times, where two samples were equal and they had to identify the different one by circling it. They had to taste it, advised to not smell it or try to differentiate by color somehow. The color was masked by black cups.

The main conditions were kept constant; 15 mL of oil (Borràs et al., 2015) at 50 °C ± 2 °C (Houhg and Fiszman, 2005) in a plastic cup of 50 mL, coded with three random digits. Each replicate contained two equal samples oil (regular oil, no rancidity) and a different (rancid) alternating with two equal (rancid) and a different (no rancidity). Warm distilled water kept at 40 °C and plain crackers were provided to clean the palate between samples (Houhg and Fiszman, 2005; Borràs et al., 2015). Each session lasted from 10 to 15 minutes, from 9 am to 12 pm.

The number of correct answers from the assessors so that there was a significant difference between the samples was found in a table based on the chi-square test; if the assessor reached the minimum of correct answers he was selected– 10 replicates requires 7 right answers (p < 0.05). Another statistical analysis applied to the selection was Wald Sequential Analysis, according to the graphical method (ISO, 2004) to further evaluate the assessors approved or rejected, and those who required training (Santana et al., 2006).

The decision system was obtained through hypothesis testing (ISO, 2004) H_o: p₁ ≤ p₀, and using the values p₀ = 0.33 (probability of a correct response when no perceptible difference exists), p₁ = 0.67 (probability of a correct response when a perceptible difference does exist), for α risk = 0:05 (probability of concluding that a perceptible difference exists when one does not) and β risk = 0.05 (probability of concluding that no perceptible difference exists when one does).

An unstructured scale of 10 cm was used for training, presented with the numbers 0 and 10 at the extremes (Houhg and Fiszman, 2005), where the assessors had a choice of where to place the intensity of the rancid defect of the sample on any point.

The training procedure consisted of three different days/stages of analysis, to calculate the accuracy of the answers and consistency of the team. On each day, four dilutions with rancid oil (0%, 10%, 50% and 100%) were provided to the assessors selected in a sufficient amount of 15mL (Borràs et al., 2015), using plastic cups coded with random digits with three numbers.

Dilutions of 0% and 100% were presented as the extremes of the scale, where 100% represented the sample at its maximum rancidity (14 days – 60 °C) and 0% represented the regular oil sample with no rancidity. The remaining, 10% and 50% dilutions were placed between 0 and 10 cm by assessors, corresponding to little–none/much rancid flavor. This procedure was repeated three times within the same day in order to have mean and standard deviations for each day. Each session lasted 10-15 minutes.

Assuming that samples were only 10% and 50%, a paired test was applied to check whether there was a difference between them, and those who inverted the order of samples on the scale (placed 50% before 10%) had their responses considered incorrect. To check the difference, the bilateral paired test table was consulted (p < 0.05) (ASTM, 2010).

The responses were measured in centimeters along the 10 cm scale. ANOVA statistical analysis evaluated individual results, means and standard variations, giving the three days’ precision using Tukey’s mean comparison test (p < 0.05) performed by Statistica^® software 12.7.

Similarly to the individual performance, the mean of each day’s responses was calculated, with respective standard deviations, to evaluate panel homogeneity. Assessors that did not differ statistically (p > 0.05) from each other, by Tukey’s mean comparison test, coinciding in the analysis of both samples, 10% and 50%, were selected for the final trained sensory panel.

Validation is important to test the panel reproducibility, which means that if the test is repeated after some time, or by another sensory panel trained exactly as in the present study, the results would not differ significantly (Lea et al., 1995).

The validation was performed eight times with the products under study, fish burgers which had been stored for 30 days, and soybean oil with two distinct antioxidants. The burgers were made with grass carp fish meat (79.00%), where 33% of the total fatty acids were polyunsaturated (Wu and Mao, 2008); ice (10.00%), vegetable fat (5.00%), textured soy protein - TSP (3.00%), spices (2.99%) and BHT (0.01%). The water to hydrate the TSP was discounted from the ice. They were vacuum-packed, and stored under refrigeration until the days of sampling, then frozen each day (initial – 0 days; 7, 14, 17, 21, 23, 25 and 30 days).

The burgers were thawed and grilled to serve to the assessors. The samples were cut into uniform sizes of about 1.5 cm³, and maintained at 75 ºC (internal center) to the time of delivery (Mitterer-Daltoé et al. 2012) using plastic cups coded with three-digit random numbers. Water at room temperature was provided to clean the mouth between samples.

Samples of soybean oil with the tertiary butylhydroquinone (TBHQ) antioxidant 200 mg/Kg, 100 and 200 mg/Kg of Quassia amara (Q.a) extract were tested after 96 hours of rancidity (60 ºC – oven) to detect any difference among them, regarding the rancid flavor.

An unstructured scale of 10 cm was applied again, for the distribution of burger and oil samples within range (different sheets), anchored in little–none/much rancid flavor. ANOVA was applied to the trained team’s results to check for differences between samples (p < 0.05). The recognition of the difference between samples were compared for equivalence with the training rancid oil to validate the trained panel.

Twenty-six people attended the selection (9 males;17 females; ages ranging from 20–50), all of which were assessors (A), 15 of which, got seven right responses or more of the ten replicates provided, based on the chi-square table for the triangle discriminatory test, and they were considered suitable for training. By means of Wald Sequential Analysis, eight more assessors were between the acceptance lines (a_x = 2.0789 + 0.5n) and rejection lines (r_x = 2.0789 + 0.5n). These assessors obtained results that made them eligible for training within the applied statistics. At this point, three people were excluded, as they were found below or at the rejection line (Fig. 1).

From the 23 assessors selected, 18 agreed to continue the training. According to the unilateral paired test table (p < 0.05) 13 assessors should set the right order of sample concentrations, 10% before 50%, in the unstructured 10 cm scale, so that, according to the paired test, the standard dilutions would present significant difference (Table 1) and become standards for the rest of the training process. Fourteen assessors got all the correct orders, verifying significant differences between dilutions.

To be approved, the assessors should have shown a total of nine correct responses, which means no change in the order of sample concentrations inside the triplicate, for every day of training.

According to the results of each assessor, the mean and the standard deviations of the responses were calculated in triplicate for each day of training through ANOVA, with the mean comparison analysis of p-values (Tukey), the mean of the tested three days which assessors presented homogeneity among the days (Table 2).

Assessor A1, whose day 3 differed significantly from the others (p ˂ 0.01), was eliminated at this stage of the statistical analysis. The remaining assessors exhibited homogeneity among days, with no significant differences among means.

For better visualization of this outcome, the F value of ANOVA (one-way) was calculated along with the p-value to test the individual performance of each assessor. F values higher than F_critical (5.1432) demonstrate significant differences among days of training.

The inability of A1 (F₁₀ = 45.87 and F₅₀ = 70.87) was also computed by the F value (Figure 2) which tested the ability and the homogeneity among the days of the other assessors in training.

The individual performance, in terms of homogeneity among the days and differentiation of samples, removed five assessors thus far. The thirteen remaining were evaluated according to panel homogeneity. Those who did not differ from each other for both samples (rancid oil; 10% and 50% standard solution) were considered part of the final trained panel (Tukey p > 0.05; n = 9).

By analyzing the p-values, only seven (A3, A4, A9, A12, A13, A17 and A18) of the thirteen remaining assessors showed homogeneity in their responses (Table 3).

Figures 3 and 4 exhibit a graphical representation through box-plots of the variation in assessors’ behavior, regarding the given dilutions. The trained panel (7 female assessors; ages ranging from 20 to 40) demonstrated proximity to 1 cm of the scale (dilution 10%) and 5cm (50% dilution), 1.70 ± 0.58 for 10% standard and 5.57 ± 0.51 for 50%. Through box-plot charts, it can be seen that the trained panel presented a combination of factors to be chosen: low variability of data, means (□) in the center of responses and proximity between means and medians (—–).

Although A3 showed a large standard deviation for the 50% dilution (Fig. 4), he presented the same consistency with the panel. The assessor with the best results (box-plot and Table 3) was A9, revealing very low data variability, high precision and accuracy. The box-plot not only allows to visualize the behavior of each assessor, but also to behold the influence of each standard used. A better data distribution was found with less variation and values closer to 1 cm for the 10% standard solution of rancid oil (Fig. 3), which was not observed in the 50% standard.

The validation of the trained panel was carried out in the analysis of samples: fish hamburgers and soybean oil with Quassia amara extract and synthetic antioxidant TBHQ.

Table 4 shows the equivalence of the rancid taste among samples and standards where burgers H1 and H2 did not differ from each other in rancid taste, nor when compared to H3, H4, H5. However, they showed a significant difference when compared to H6, H7 and H8, the last one considered more pronounced in this parameter. The sample H8, with 30 days of storage under refrigeration, manifested a significant difference with the other days except from the 25^th and equivalent to the 50% standard. H3, H4 and H5 have a rancid flavor deffect correnponding to a 10% rancid oil standard.

The assessors’ perception showed an equivalent degree of difference in standards and hamburger samples. At the first phase of training it became clear that 10% and 50% standards are different, proving that hamburguers are too. The trained panel was able to find differences equivalent to the training of samples, which have a more complex matrix, and therefore validates the assessors and the method.

For oil samples, after 96 hours in an oven at 60 °C, there was no significant difference among the samples tested (Soybean oil with TBHQ, soybean oil with 100 mg/Kg of Quassia amara and soybean oil with 200 mg/Kg of Quassia amara) regarding the rancid flavor (Table3). These results demonstrate that the concentration of Quassia amara extract used was effective in preventing rancidity, like the synthetic antioxidant. Even the lowest concentration tested, revealing that is possible to use 100 mg/Kg for economic purposes, with a similar effect and no significant difference. Also, it is evident that the rancidity found in all three samples corresponds to the 10% rancid oil standard.