Physicochemical Changes and Antioxidant Activity Prediction Model of Corn/Ginger-Based Extrudates during a Long Term Storage

Abstract

Texture characteristics and antioxidant activities (AOA) of extrudates were stored under different conditions that were detected by texture profile analysis (TPA) and DPPH* method in this study. The physicochemical properties of extrudates were significantly affected by storage time and temperature. The hardness values of extrudates stored at 0°C were the highest, while the crispness values of it were the lowest. The AOA decreased significantly from 20.23% to 14.87% with temperature increasing from −10°C to 25°C. The back propagation Artificial Neural Network (bp-ANN) was used to predict the AOA from hardness and crispness. The optimized model structures had two hidden layers, one with ten neurons per layer (R² ≥ 0.999) and another one with eight neurons per layer (R² ≥ 0.993). The ANN model was a better predictor of AOA from texture characteristics than linear fitting model (AOA vs. hardness: R² = 0.913; AOA vs. crispness: R² =0.952).

Introduction

Extrusion has been the most popular food processing method with its versatility, simple technology, low cost and high production capability since the early 1930s. It is widely used in manufacturing food products such as cereal breakfast, snack food, ready-to-eat food, children's diets and pet food (Camire et al., 2007; İbanoğlu et al., 2006; Stojceska et al., 2009). Nowadays, the extrusion was widely used to increase the nutritional value and additional value by adding some special ingredients in the products, such as high dietary fiber grain, bioactive functional components, and aroma compounds (Dehghan-Shoar et al., 2010; Holguín-Acuña et al., 2008; Menis et al., 2013; Ozer et al., 2006). Antioxidant capacity is traditionally used to evaluate the nutritive value of extruded food; the most pronounced changes of antioxidant activity present in food result from oxidation reactions occurring slowly during storage (Zielinski et al., 2012). The testing of antioxidant activity will be a heavy workload in the long period of monitoring.

Most of the extruded foods have a long-term storage period more than half a year. Many studies have been done on the changes of extruded foods during storage period, including odor, color, water transfer, texture characteristics and antioxidant activity (García-Tejeda et al., 2013; Kumar et al., 2012; Shaviklo et al., 2011). However, these studies were time consuming and costly. Therefore, efficient numerical tool has been improved more desirable. Also, an in-depth study on the quality changes of extruded foods during a long-term storage plays an important role on the on-line control and shelf life prediction.

Artificial Neural Network is a mathematical model to distribute parallel information processing algorithm based on the structure of biological neurons. It offers a simple approach to nonlinear problems without any historical information on the relationship of variables. Artificial Neural Network modeling has been increasingly accepted and widely used in various scientific and technological fields including food industry (Fan et al., 2013). ANN has been successfully applied for modeling, optimizing, predicting and process controlling of complex problems existing or occurring inside foods. In the area of food quality control, ANN has been successfully applied to predict the shelf life of pasteurized milk and soya milk (Vallejo-cordoba et al., 1995; Ko et al., 2000), process variables in food extrusion (Cubeddu et al., 2014), extrudate properties (Shankar and Bandyopadhyay, 2007), and product colour during food extrusion (Valadez-Blanco et al., 2007). However, it has been rarely used in the prediction of changes on extrudate properties during storage.

In this study, texture characteristics and antioxidant activities were measured. A polynomial model was used to predict the quality changes and a linear model and ANN model were used to determine the relationship between the textures and antioxidant activities. The objectives of this study were: 1. study the effects of three different storage temperatures (−10°C, 0°C, 25°C) on the physicochemical properties; 2. build an artificial network (ANN) model to predict the AOA value using texture characteristics.

Materials and Methods

Materials    Corn flour and Ginger powder were purchased from a local supermarket (shanghai, China). The moisture content was determined by using the drimeter (MB45, Ohaus, Shanghai, China). Based on our previous study, the optimal conditions were described as follows:

Corn flour and ginger powder were mixed in ratio 9:1 (0.1 g/g ginger powder). The moisture content of the raw materials was regulated to 180 g/kg. The materials were mixed and stored in self-styled bags to get a uniform moisture.

Extrusion was performed in a co-rotating twin-screw extruder (DS32II, Jinan, Saixin Food Machinery, Shandong, China), consisting of three independent zones with controlled temperature in the barrel without the die. The materials were fed into the extruder at a constant feed rate (15 kg/h) using a volumetric feeder. The length to diameter (L/D) ratio for the extruder was 18.9:1, and the maximum screw speed was 8 r/s. The temperature profiles in the feed, compression metering zones and the third zone were 90°C, 120°C and 150°C, respectively. The expansion index was the highest (3.97, D_extrudates/D_die) and bulk density was the lowest (0.113 g/cm³).

The extrudates were divided into two groups. One group was cut into 2.5 cm (length) for the texture profile determination, and the other one was ground into powder using a mill and screened through 60-mesh sieve (British, standard, µm) for the assessment of AOA. The hardness, crispness and AOA of extrudates were determined after being balanced at room temperature for 24h, which were marked as storage time of 0 day. Both the cylindrical extrudates and powder samples were divided into three series, and stored in low-density polyethylene plastic bags at different temperatures (−10, 0, and 25°C). The relative humidity storage environment was controlled according to Moraga et al. (2012), ranging from 20% – 30%. All the properties were measured each 15 days in the first 6 months and 30 days after 6 months.

Texture profile analysis    Texture analyzer (Stable Micro Systems, TA. XT. Plus, Surrey, UK) with a 50-kg load cell was used to evaluate the texture values of extrudates. The equipment was set to zero through the height correction. The samples were then carefully placed on a texture analysis table and each sample was compressed to a depth of 60% deformation at a constant speed of 1 mm/s and 50 N compression loads using a metal probe (Stable Micro Systems, P100, Surrey, UK). The probe started its upward stroke at a constant speed of 5 mm/s after the compression stroke was complete. Deriving from TPA curve, the maximum compressive stress was indicated as the hardness value (N), and the number of the peak of wave was indicated as the crispness. All samples were conducted at least six replicates, and the average values (excluding maximum and minimum ones) were used for further analysis.

Determination of Antioxidant activity by DPPH method    Antioxidant activity was measured using the DPPH method with little modifications (Sharma et al., 2012). Extruded samples (0.1 g) were mixed with 1 mL methanol, and centrifuged at 5000 rpm for 10 min after two hours of water bath. The supernatant (100 µL) was added to 3.9 mL of methanolic solution containing DPPH* (6×10⁻⁵ mol/L). The solution was vigorously shaken and the absorbance of supernatant was read at 0 min and 30 min at 515 nm using a methanol blank. The antioxidant activity was calculated as follows:

  

All tests were done in triplicate.

Artificial neural networks (ANNs)    Artificial neural networks were developed by using Matlab R2012a version, with the Neural Network Toolbox 7.0.3 (The Mathworks, Inc, Natick, MA, USA). ANNs can simulate human brain to store and process knowledge with input, hidden and output layers. An output node, subject to weighted transfer function, becomes one of the inputs to the nodes of the following level (Khajeh and Barkhordar, 2013). The neural weight shows the influence of the strength on neurons connection with the information transfer between neurons. ANN weights and biases are computed relied on pairing of experimental input and output (target) data. There are several common types of transfer functions, the equations are as follows:

Hyperbolic tangent transfer function (tansig)

  

Linear transfer function (purelin)

  

ANN training and simulation    The schematic structure of ANN had input layer, hidden layer and output layer (Figure 1). The inputs for the network include hardness and crispness, and the output was the AOA value. In this work, a multilayer feed-forward neural network trained with an error back propagation learning algorithm used for computing the ANN weights and biases was applied to model the change of AOA. Sixteen observations determined in the experiment were randomly divided into three subsets, namely a training subset composed of ten observations, a validation subset composed of three observations, and a simulating subset composed of three observations. A hyperbolic tangent transfer function was selected for each hidden layer to map the experimental data to the range of −1 to 1, while a purelin function was selected for output layer in the present study in order to simplify the algebraic equation when training the neural network (Dwivedi and Ramaswamy, 2010; Kerdpiboon et al., 2006). Performance of ANN with Levenberg-Marquardt algorithm approached to minimal error. The training process stopped when the performance error goal was less than 0.001.

Fig. 1.

Schematic structure of ANN for the prediction of hardness, crispness and AOA.

The input, hidden and output layers are the composition of ANN (Figure 1). The number of hidden layers and number of neurons of each hidden layer was decided with few nodes in this task (Shankar and Bandyopadhyay, 2007). The number of hidden layers was varied from 1 to 2, while the number of neurons was varied from 1 to 10. The back-propagation training algorithm could adjust the weights and biases automatically.

The average mean square error (MSE), percentage of relative mean square error (%MRE), standard deviation of %MRE (STD_R) and the coefficient of determination (R2) were used to compare the performance of various ANN configuration (Kerdpiboon et al., 2006; Niamnuy et al., 2012). All of them were defined as:

  

Where, is the average of the, ΔY_R, Y_E is the experimental outputs, Y_N is the network outputs, Y_A is the average of the experimental values, N is the experimental group number.

The hyperbolic tangent transfer function is bounded between −1 and 1 while the magnitudes vary from different parameters. Thus, the input and output data should be normalized to range between −1 and 1, according to equation below:

  

where x is variable, x_max is maximum value, x_min is minimum value, y_max and y_min are defaults chosen 1 and −1, respectively.

Statistical analysis    The data were analyzed using SPSS 17.0 (SPSS Inc., Chicago, USA). The polynomial fitting model and linear fitting model were performed using Origin Pro 8.0 software (OriginLab, Inc., USA). The modeling and prediction of ANN were carried out using MATLAB R2012a (The Math-works, Inc., USA).

Result and Discussion

Texture characteristics changes during storage    It can be observed that the hardness for all samples increased during a long-term storage (Figure 2A). The hardness of corn-ginger based extrudates stored at −10°C, 0°C and 25°C ranged from 484.53 N (0 month) to 697.06 N, 721.19N and 617.32N (9 months), respectively. The increasing of hardness for all samples during a long-term storage was similar to results for carrot, pulse and rice by-product based extrudates (Kumar et al., 2012), and corn-beef based extrudates (Badding-Smithey et al., 1995). It could be explained by the fact that the starch bonding caused by the increasing moisture content of extrudates with extended storage period led to a tight structure (Chanvrier et al., 2007). Significant difference was the effect of storage temperature on hardness (p < 0.05). The hardness of these extrudates stored at 0°C was the highest. Some researchers have reported the relationship between starch retrogradation and hardness of starch-based product (Baik and Chinachoti, 2000; Silverio et al., 2000). It indicated that the kinetics of starch retrogradation exhibited a strong temperature dependence because the nucleation rate increase exponentially with decreasing of temperature down to the glass transition temperature, while the propagation rate increased exponentially with increasing of temperature up to melting temperature (Silverio et al., 2000). The amylopectin recrystallization could increase the hardness significantly (Farhat et al., 2000).

Crispness is one of the most important characteristics of extrudates. The crispness decreased significantly as the storage period extended and temperature increased (Figure 2B). The crispness of samples stored at 25°C decreased faster than the other two, −10°C and 0°C, as the water is apt to diffuse at high temperature. The final crispness of samples under −10°C, 0°C and 25°C after 9 months storage were 21.33, 14.00 and 7.67, respectively, while the determined moisture content of these samples were 105.9 g/kg, 114.2 g/kg and 130.7 g/kg, respectively. The similar result had been done on the crispness of maize-soybean extrudates after 50 days storage (Caballero et al., 2013). Also, González et al. (2006) found that crispness of expanded maize would be lost when product moisture was beyond 9%. The loss of crispness could be explained by the migration of water vapor from storage environment into packaging materials deteriorated the physical properties of extrudates (Kumar et al., 2012; Shaviklo et al., 2011).

AOA changes during storage    Estimation of AOA on the extrudates was measured against DPPH* radicals (Figure 2C). The AOA significantly decreased at three different temperatures with a large reduction about 38.37% at 25°C, and 19.31% at −10°C and 21.78% at 0°C, respectively. The temperature had significantly effect on antioxidant capacity. The results were similar to extruded fish feed (Hernández et al., 2014), and almond kernels (Raisi et al., 2015). The reason was the oxygen transformation through the self-styled bags and the limitation of oxygen removed from both extrudates and self-styled bags (Jensen et al., 2011).

Polynomial fitting model    The parameter estimation revealed the positive effect of storage period on the hardness, and the negative effect of storage period on the crispness and AOA (Table 1). The hardness, crispness and AOA were highly correlated with storage period and storage temperature (Figure 2). An intuitionistic relationship between the variable X, which represented storage period, and independent variable Y, which represented hardness, crispness and AOA, was calculated using a polynomial equation constructed as:

Table 1. Hardness, crispness and AOA values of the samples stored during a period from 0 to 9 months under different temperature.

Time Month	Hardness (N)			Crispness			AOA (%)
	−10°C	0°C	25°C	−10°C	0°C	25°C	−10°C	0°C	25°C
0	484.53 ± 74.70	484.53 ± 74.70	484.53 ± 74.70	84.33 ± 2.43	84.33 ± 2.43	84.33 ± 2.43	25.07 ± 0.33	25.07 ± 0.33	25.07 ± 0.33
0.5	497.24 ± 69.08	492.96 ± 59.18	490.10 ± 53.51	82.67 ± 1.58	80.33 ± 3.50	77.33 ± 1.91	24.91 ± 0.21	24.64 ± 0.41	23.57 ± 0.31
1	503.10 ± 80.30	499.90 ± 53.58	496.42 ± 62.55	81.33 ± 4.73	70.67 ± 2.08	60.33 ± 2.66	24.39 ± 0.27	23.85 ± 0.31	22.51 ± 0.23
1.5	519.08 ± 35.92	528.27 ± 61.16	510.77 ± 82.38	75.33 ± 1.26	67.67 ± 4.43	55.67 ± 1.53	23.96 ± 0.14	23.07 ± 0.16	21.46 ± 0.29
2	533.36 ± 45.52	543.50 ± 57.76	515.67 ± 78.69	68.00 ± 1.69	61.00 ± 3.51	39.67 ± 2.31	23.29 ± 0.31	22.41 ± 0.32	19.81 ± 0.58
2.5	563.61 ± 48.99	569.31 ± 54.69	534.36 ± 61.16	64.33 ± 1.15	50.67 ± 3.51	35.00 ± 3.03	22.99 ± 0.45	21.90 ± 0.20	18.6 ± 0.42
3	575.34 ± 48.12	594.12 ± 54.03	550.28 ± 39.68	58.67 ± 1.84	41.33 ± 2.40	31.00 ± 2.19	22.04 ± 0.26	21.58 ± 0.51	17.61 ± 0.37
3.5	590.47 ± 37.16	617.93 ± 42.58	557.07 ± 80.01	50.33 ± 3.50	41.00 ± 3.72	25.33 ± 3.57	21.59 ± 0.57	21.07 ± 0.16	16.71 ± 0.42
4	620.65 ± 80.88	624.95 ± 45.52	565.69 ± 66.89	47.00 ± 4.57	38.33 ± 2.57	23.67 ± 4.15	21.5 ± 0.26	20.97 ± 0.37	15.9 ± 0.23
4.5	631.61 ± 76.92	652.68 ± 54.20	576.53 ± 81.29	46.33 ± 1.57	27.67 ± 2.14	18.33 ± 1.15	21.19 ± 0.44	20.74 ± 0.25	15.42 ± 0.19
5	642.49 ± 52.92	661.62 ± 42.58	582.71 ± 50.85	39.67 ± 2.31	28.67 ± 2.56	13.33 ± 3.11	21.05 ± 0.38	20.39 ± 0.33	15.56 ± 0.13
5.5	662.46 ± 44.91	665.63 ± 36.46	595.34 ± 47.84	33.33 ± 3.12	21.00 ± 1.53	10.00 ± 1.97	20.97 ± 0.23	20.19 ± 0.45	15.11 ± 0.26
6	666.88 ± 73.51	695.37 ± 52.21	595.90 ± 64.83	29.67 ± 1.69	17.67 ± 2.08	12.67 ± 3.42	20.57 ± 0.29	20.01 ± 0.31	14.85 ± 0.18
7	682.78 ± 52.80	701.30 ± 49.90	606.60 ± 55.83	27.33 ± 0.58	15.67 ± 2.65	9.67 ± 2.86	20.4 ± 0.23	19.81 ± 0.25	14.96 ± 0.33
8	686.18 ± 64.53	717.17 ± 43.00	609.64 ± 59.19	24.67 ± 2.22	14.67 ± 3.13	11.00 ± 2.07	20.29 ± 0.25	19.69 ± 0.22	14.99 ± 0.26
9	697.06 ± 56.01	721.19 ± 69.74	617.32 ± 69.62	21.33 ± 1.31	14.00 ± 1.53	7.67 ± 1.53	20.23 ± 0.31	19.61 ± 0.34	14.87 ± 0.24

Fig. 2.

Polynomial fitting model for analyzing the changes of hardness (A), crispness (B) and AOA (C) during a storage period under different temperatures.

  

Where X is period, Y is hardness, crispness and AOA. a, b, c is coefficient.

Average accuracy of the estimation was assessed using the R² and “accuracy factor, AF” (Table 2). The AF was refined by Baranyi et al. (1999) as Equation 10:

Table 2. Verification of polynomial models: R² value and accuracy factor.

	Regression model:Y = β0 + β1X + β2X2	R2	AF
Hardness
−10°C	Y = 461.06 + 45.97X − 2.12 X2	0.981	1.02
0°C	Y = 461.92 + 51.45X − 2.44 X2	0.988	1.02
25°C	Y = 474.48 + 28.95X − 1.45 X2	0.989	1.01
Crispness
−10°C	Y = 89.27 − 13.12X + 0.58 X2	0.978	1.07
0°C	Y = 89.48 − 17.72X + 1.04 X2	0.992	1.06
25°C	Y = 83.33 − 21.35X + 1.46 X2	0.987	1.28
AOA
−10°C	Y = 25.53 − 1.28X + 0.08 X2	0.988	1.01
0°C	Y = 25.02 − 1.43X + 0.09 X2	0.990	1.01
25°C	Y = 25.05 − 2.98X + 0.21 X2	0.988	1.02

  

Where C_k is calculated value for observation number k, M_k is the measured value for observation number k, and n is the number of observations.

All the equations had high R² values, which proved the goodness-of-fit of the polynomial model (Table 2). The lowest R² for hardness, crispness and AOA appeared at −10°C were 0.981, 0.978 and 0.988, respectively. The result showed that the polynomial model was best fit of AOA values with the highest R² = 0.992. AF values were used to estimate the calculated values, the lower the AF values were the more accurate average estimate was. The AF value indicated that the calculated values were accurate.

Relationship between textures and antioxidant activities

Linear fitting modeling    It observed that hardness and crispness were highly correlated with AOA values under all the temperatures (Table 3). The correlation was the highest at 0°C for hardness and AOA (R² = 0.944), and at 25°C for crispness and AOA (R² = 0.981) (Figure 3). The high correlation coefficient showed the internal chemical property (AOA) could be reflected through the physical properties (hardness and crispness). The rationale of this phenomenon has two possible explanations. First, the oxygen and water vapour could transfer through the self-styled bags could reduce the AOA value and crispness score, while the hardness score increased. The AOA value had negative effect on the hardest sample (Figure 4). Furthermore, the extent Maillard Reaction that could change the chemical composition and structure of extrudates might be another significant reason.

Table 3. Verification of linear fitting models: R² values and accuracy factor.

	Regression model: Y = aX + b	R2	AF
Hardness-AOA
−10°C	Y = −0.02X + 33.77	0.937	1.12
0°C	Y = −0.02X + 34.39	0.944	1.07
25°C	Y = −0.07X + 58.93	0.913	1.21
Crispness-AOA
−10°C	Y = 0.08X + 18.16	0.952	1.08
0°C	Y = 0.07X + 18.51	0.971	1.05
25°C	Y = 0.14X + 13.47	0.981	1.03

Fig. 3.

Linear curve fitting results for correlation analysis between AOA and hardness ( a: −10°C; b: 0°C; c: 25°C) and between AOA and crispness ( d: −10°C; e: 0°C; f: 25°C).

Fig. 4.

Predictability of the ANN model for AOA from hardness (a) and AOA from crispness (b).

ANN modeling    In reality, a trained neural network consists of input variables, connection weights, biases and transfer functions forming a complex calculation process. Equation 12 – 14 represent the output neural network in each layer, namely input layer to hidden layer, first hidden layer to second hidden layer and hidden layer to output layer (Dwivedi and Ramaswamy, 2010; Llave et al., 2012).

  

  

Where X is the input value; Y₁, Y₂, Y₃ are the output values in each transport layer; f₁ and f₂ are the transfer functions in the hidden and output layers respectively; W₁, W₂, W₃ are the connection weights and B₁, B₂, B₃ are the biases in each transport layer. The connection weights and biases were listed in the appendix.

The experimental data were used to train the ANN in order to give better predict model by optimizing the network configurations. The performance matrices for the ANN models predicted storage condition and quality changes of extrudates with different numbers of hidden layers and neurons per layer (Table 4). The hardness and crispness of extrudates were quite related with its AOA (R² ≥ 0.993) under each temperature. The ANN configuration that involved two hidden layers with ten neurons could predict AOA from hardness (%MRE ≤ 0.144; R² ≥ 0.999) very well. The best ANN configuration to predict AOA from crispness involved two hidden layers and eight neurons in each layer (%MRE ≤ 1.217; R² ≥ 0.993). In all cases, increasing the number of hidden layers and neurons per layer was not necessary to have a beneficial impact to the performance of ANN training. Regardless of the number of hidden layers, the R² (hardness vs. AOA, 0.992; crispness vs. AOA, 0.993) turned to be high and stable when the neurons increased to four. Consequently, the ANN model provided an ideal performance parameter for predicting the AOA value from hardness and crispness score.

Table 4. Errors in the prediction of AOA with different numbers of hidden layers and neurons per hidden layer.

No. of hidden layer	No. of neuron	Hardness-AOA
		−10°C			0°C			25°C
		%MRE	STDR	R2	%MRE	STDR	R2	%MRE	STDR	R2
1	1	0.451	0.004	0.993	0.561	0.006	0.992	1.768	0.011	0.982
	2	0.962	0.009	0.964	1.11	0.126	0.959	0.536	0.005	0.996
	4	0.263	0.002	0.998	0.201	0.004	0.996	0.598	0.006	0.995
	6	0.218	0.003	0.997	0.390	0.002	0.998	1.142	0.023	0.987
	8	0.365	0.002	0.996	0.09	0.001	0.999	3.983	0.054	0.898
	10	0.127	0.001	0.999	2.26	0.457	0.725	0.215	0.004	0.999
2	1	1.062	0.011	0.914	0.626	0.004	0.981	1.226	0.007	0.993
	2	1.191	0.019	0.901	0.262	0.003	0.996	0.849	0.008	0.993
	4	0.233	0.002	0.998	0.215	0.002	0.998	1.356	0.011	0.992
	6	0.314	0.001	0.997	0.153	0.002	0.999	0.591	0.011	0.996
	8	0.142	0.001	0.999	0.074	0.001	0.999	1.229	0.012	0.993
	10	0.095	0.001	0.999	0.129	0.001	0.999	0.144	0.002	0.999
		Crispness-AOA
		−10°C			4°C			25°C
		%MRE	STDR	R2	%MRE	STDR	R2	%MRE	STDR	R2
1	1	0.527	0.003	0.993	2.419	0.014	0.846	1.621	0.007	0.991
	2	0.575	0.004	0.994	0.365	0.004	0.994	1.492	0.011	0.989
	4	0.331	0.001	0.997	0.421	0.003	0.993	1.224	0.012	0.993
	6	0.094	0.002	0.999	0.145	0.002	0.997	1.173	0.014	0.992
	8	0.288	0.003	0.996	0.056	0.001	0.999	1.241	0.012	0.993
	10	0.346	0.004	0.991	0.052	0.001	0.999	1.398	0.010	0.993
2	1	0.592	0.005	0.989	0.596	0.007	0.985	1.407	0.011	0.992
	2	0.567	0.005	0.989	0.489	0.003	0.994	1.375	0.012	0.992
	4	0.223	0.003	0.997	0.257	0.002	0.997	1.246	0.012	0.994
	6	0.102	0.001	0.999	0.375	0.004	0.994	1.039	0.014	0.993
	8	0.083	0.001	0.999	0.038	0.001	0.999	1.217	0.013	0.993
	10	0.185	0.003	0.996	0.057	0.001	0.999	1.152	0.011	0.994

Comparing with the linear fitting model based on correlation analysis, the ANN model was a better predictor for both AOA from hardness (R² = 0.999 for ANN vs. R² = 0.913 for linear fitting model) and crispness (R² = 0.993 for ANN vs. R² = 0.952 for linear fitting model). The AOA could be accurately predicted from hardness and crispness with its high correlation.

Conclusion

Storage period and temperature had significant effect on the quality of extrudates. The relationship between storage conditions (period and temperature) and extrudates properties (hardness, crispness and AOA) was analyzed by correlation analysis, and a strong relationship was discovered. An ANN model was successfully trained for simulating and predicting the AOA value from texture score of extrudates during a long-term storage under different temperatures. The optimized ANN model was a better predictor for the AOA from hardness and crispness with high R² values (0.999 and 0.993, respectively).

Acknowledgements    This study was sponsored by the National Research Funds for the Central Universities and Baoshan Association for Science & Technology of Shanghai.

Appendix

1. (1)   Weights and biases of bp-1-10-10-1 for prediction of AOA from hardness under −10°C
   

2. (2)   Weights and biases of bp-1-10-10-1 for prediction of AOA from hardness under 0 °C
   

3. (3)   Weights and biases of bp-1-10-10-1 for prediction of AOA from hardness under 25 °C
   

4. (4)   Weights and biases of bp-1-8-8-1 for prediction of AOA from crispness under −10 °C
   

5. (5)   Weights and biases of bp-1-8-8-1 for prediction of AOA from crispness under 0 °C
   

6. (6)   Weights and biases of bp-1-8-8-1 for prediction of AOA from crispness under 25°C
   

Reference

•  Badding-Smithey,  S.L.,  Huffa,  H.E., and  Hesieh,  F. (1995). Processing parameters and product properties of extruded beef with nonmeat cereal binders. Food Sci. Tech., 28, 386-394.
•  Baik,  M.Y. and  Chinachoti,  P. (2000). Moisture redistribution and phase transitions during bread staling. Cereal Chem., 77, 484-488.
•  Baranyi,  J.,  Pin,  C., and  Ross,  T. (1999). Validating and comparing predictive models. Int. J. Food Microbiol., 48, 159-166.
•  Caballero,  María S.,  Drago,  Silvina R.,  Costa,  Silvia C.,  Sabbag,  Nora G., and  González,  Rolando J. (2013). Physicochemical characteristics and shelf life estimation of maize/soybean extrudates added with bovine haemoglobin. Int. J. Food Sci. Tech., 48, 1808-1814
•  Camire,  M.E.,  Dougherty,  M.P., and  Briggs,  J.L. (2007). Functionality of fruit powders in extruded corn breakfast cereals. Food Chem., 101, 765-770.
•  Chanvrier,  H.,  Uthayakumaran,  S.,  Appelqvist,  I. A.,  Gidley,  M. J.,  Gilbert,  E. P., and  Lopez-Rubio,  A. (2007). Influence of storage conditions on the structure, thermal behavior, and formation of enzyme-resistant starch in extruded starches. J. Agric. Food Chem., 55, 9883-9890.
•  Cubeddu,  A.,  Rauh,  C., and  Delgado,  A. (2014). Hybrid artificial neural network for prediction and control of process variables in food extrusion. Innov. Food Sci. Emerg., 21, 142-150.
•  Dehghan-Shoar,  Z.,  Hardacre,  A.K., and  Brennan,  C.S. (2010). The physico-chemical characteristics of extruded snacks enriched with tomato lycopene. Food Chem., 123, 1117-1122.
•  Dwivedi,  M. and  Ramaswamy,  H. S. (2010). Artificial neural network modelling of heat transfer to canned particulate fluids under axial rotation processing. Int. J. Food Eng., 6, Art.5.
•  Fan,  F.H.,  Ma,  Q.,  Ge,  J.,  Peng,  Q.Y.,  Riley,  W.W., and  Tang,  S.Z. (2013). Prediction of texture characteristics from extrusion food surface images using a computer vision system and artificial neural networks. J. Food Eng., 118, 426-433.
•  Farhat,  I.A.,  Blanshard,  M.V., and  Mitchell,  J.R. (2000). The retrogradation of waxy maize starch extrudates: Effects of storage temperature and water Content. Biopolymers, 53, 411-422.
•  García-Tejeda,  Y.V.,  López-González,  C.,  Pérez-Orozco,  J.P.,  Rendón-Villalobos,  R.,  Jiménez-Pérez,  A.,  Flores-Huicochea,  E.,  Solorza-Feria,  J., and  Bastida,  C.A. (2013). Physicochemical and mechanical properties of extruded laminates from native and oxidized banana starch during storage. LWT - Food Sci. Technol., 54, 447-455.
•  González,  R. J.,  Torres,  R. L.,  De Greef,  D. M., and  Guadalupe,  B. A. (2006). Effects of extrusion conditions and structural characteristics on melt viscosity of starchy materials. J. Food Eng., 74, 96-107.
•  Hernández,  A.,  García García,  B.,  Jordán,  M. J., and  Hernández,  M. D. (2014). Natural antioxidants in extruded fish feed: Protection at different storage temperatures. Anim. Feed Sci. Tech., 195, 112-119.
•  Holguín-Acuña,  A.L.,  Carvajal-Millán,  E.,  Santana-Rodríguez,  V.,  Rascón-Chu,  A.,  Márquez-Escalante,  J.A.,  Ponce de León-Renova,  N.E., and  Gastelum-Franco,  G. (2008). Maize bran/oat flour extruded breakfast cereal: A novel source of complex polysaccharides and an antioxidant. Food Chem., 111, 654-657.
•  İbanoğlu,  Ş.,  Ainsworth,  P.,  Özer,  E.A., and  Plunkett,  A. (2006). Physical and sensory evaluation of a nutritionally balanced gluten-free extruded snack. J. Food Eng., 75, 469-472.
•  Jensen,  S.,  Oestdal,  H.,  Clausen,  M.R.,  Andersen,  M.L., and  Skibsted,  L.H. (2011). Oxidative stability of whole wheat bread during storage. LWT - Food Sci. Technol., 44, 637-642.
•  Kerdpiboon,  S.,  Kerr,  W.L., and  Devahastin,  S. (2006). Neural network prediction of physical property changes of dried carrot as a function of fractal dimension and moisture content. Food Res. Int., 39, 1110-1118.
•  Khajeh,  M. and  Barkhordar,  A. (2013). Modelling of solid-phase tea waste extraction for the removal of manganese from food samples by using artificial neural network approach. Food Chem., 141, 712-717.
•  Ko,  S.H.,  Park,  E.Y.,  Han,  K.Y.,  Noh,  B.S., and  Kim,  S.S. (2000). Development of neural network analysis program to predict shelf life of soya milk by using 3 electronic nose. Food Eng. Prog., 4, 193-198
•  Kumar,  N.,  Sarkar,  B.C.,  Sharma,  H.K., and  Jha,  S.K. (2012). Colour kinetics and storage characteristics of carrot, pulse and rice by-product based extrudates. Brit Food J., 114, 1279-1296.
•  Llave,  Y.A.,  Hagiwara,  T., and  Sakiyama,  T. (2012). Artificial neural network model for prediction of cold spot temperature in retort sterilization of starch-based foods. J. Food Eng., 109, 553-560.
•  Menis,  M.E.C.,  Milani,  T.M.G.,  Jordano,  A.,  Boscolo,  M., and  Conti-Silva,  A.C. (2013). Extrusion of flavored corn grits: Structural characteristics, volatile compounds retention and sensory acceptability. LWT - Food Sci. Technol., 54, 434-439.
•  Moraga,  G.,  Igual,  M.,  García-Martínez,  E.,  Mosquera,  L. H.,  Martínez-Navarrete,  N. (2012). Effect of relative humidity and storage time on the bioactive compounds and functional properties of grapefruit powder. J. Food Eng., 112, 191-199.
•  Niamnuy,  C.,  Kerdpiboon,  S., and  Devahastin,  S. (2012). Artificial neural network modeling of physicochemical changes of shrimp during boiling. LWT - Food Sci. Technol., 45, 110-116.
•  Ozer,  E.A.,  Herken,  E.N.,  Guzel,  S.,  Ainsworth,  P., and  Ibanoglu,  S. (2006). Effect of extrusion process on the antioxidant activity and total phenolics in a nutritious snack food. Int. J. Food Sci. Tech., 41, 289-293.
•  Raisi,  M.,  Ghorbani,  M.,  Sadeghi  M. A.,  Kashaninejad,  M., and  Hosseini,  H. (2015). Effect of storage atmosphere and temperature on the oxidative stability of almond kernels during long term storage. J. Stored Prod. Res., 62, 16-21.
•  Shankar,  T.J. and  Bandyopadhyay,  S. (2007). Prediction of Extrudate Properties Using Artificial Neural Networks. Food Bioprod Process, 85, 29-33.
•  Sharma,  P.,  Gujral,  H.S., and  Singh,  B. (2012). Antioxidant activity of barley as affected by extrusion cooking. Food Chem., 131, 1406-1413.
•  Shaviklo,  G.R.,  Thorkelsson,  G.,  Rafipour,  F., and  Sigurgisladottir,  S. (2011). Quality and storage stability of extruded puffed corn-fish snacks during 6-month storage at ambient temperature. J. Sci. Food Agr., 91, 886-893.
•  Silverio,  J.,  Fredriksson,  H.,  Andersson,  R.,  Eliasson,  A.C., and  Aman,  P. (2000). The effect of temperature cycling on the amylopectin retrogradation of starches with different amylopectin unit-chain length distribution. Carbohydr Polym, 42, 175-184.
•  Stojceska,  V.,  Ainsworth,  P.,  Plunkett,  A., and  İbanoğlu,  Ś. (2009). The effect of extrusion cooking using different water feed rates on the quality of ready-to-eat snacks made from food by-products. Food Chem., 114, 226-232.
•  Valadez-Blanco,  R.,  Virdi,  A. I. S.,  Balke,  S. T., and  Diosady,  L. L. (2007). In-line colour monitoring during food extrusion: Sensitivity and correlation with product colour. Food Res. Int., 40, 1129-1139.
•  Vallejo-Cordoba,  B.,  Arteaga,  G.E., and  Nakai,  S. (1995). Predicting milk shelf life based on artificial neural networks and headspace gas chromatographic data. Journal of Food Sci., 60, 885-888.
•  Zielinski,  H.,  Castillo,  M. D.,  Przygodzka,  M.,  Ciesarova,  Z.,  Kukurova,  K., and  Zielinska,  D. (2012). Changes in chemical composition and antioxidative properties of rye ginger cakes during their shelf-life. Food Chem., 135, 2965-2973.