DOI:
10.1039/C5RA25375F
(Paper)
RSC Adv., 2016,
6, 4663-4672
Comparison of different chemometric methods in quantifying total volatile basic-nitrogen (TVB-N) content in chicken meat using a fabricated colorimetric sensor array†
Received
29th November 2015
, Accepted 22nd December 2015
First published on 5th January 2016
Abstract
Total Volatile Basic-Nitrogen (TVB-N) content is one of core measures in evaluating chicken freshness. This study reported the feasibility to quantify Total Volatile Basic-Nitrogen (TVB-N) content in chicken meat by a low cost colorimetric sensor array with the help of chemometric analysis. We fabricated a colorimetric sensor array by printing 12 chemically responsive dyes (i.e. 9 porphyrins/metalloporphyrins and 3 pH indicators) on a C2 reverse silica-gel flat plate for the fast and non-destructive quantitative determination of TVB-N content in chicken. A colour change profile for each sample was obtained by differentiating the image of the sensor array before and after exposure to volatile organic compounds (VOCs). Linear algorithm; partial least squares regression (PLSR) and nonlinear algorithms; back propagation artificial neural network (BPANN), Adaptive Boosting BPANN (BP-AdaBoost) and support vector machine regression (SVMR) methods based on particle swarm optimization (PSO) were used to build the TVB-N prediction model. Experimental results showed that the predictive precision of the PSO-SVMR model was superior to linear and classic non-linear models. The optimum PSO-SVMR model was obtained with 4 support vectors and Rp of 0.8981, RMSEP of 5.5255. The overall results are encouraging for the application of low cost colorimetric sensors combined with an appropriate chemometric method in the poultry industry for quality assessment because it is practical, non-invasive, rapid and simple.
Introduction
Poultry meat is one of the fastest growing food commodities in many parts of the world because of the low cost of production as compared to meat products such as beef, lamb or pork, low fat content, high nutritional value and distinct flavor.1 The demand for chicken is ever-rising and therefore its freshness is a major concern. The food-borne illnesses are mostly caused due to the poor quality control and test measures. Hence, there is an urgency for a simple, rapid and handy, low cost and accurate quality analysis and control equipment.2 At present, quality control measurement in the meat industry is mostly done by two methods in order to evaluate meat freshness. One is chemical and microbial measurement like Total Viable Bacterial Counts (TVC) and Total Volatile Basic-Nitrogen (TVB-N); the other is by sensory evaluation that involves the estimation of organoleptic attributes with the help of skillful experts.3 The former is a very objective method, but also a destructive method that takes 2–5 days to obtain results. This means that the method cannot simultaneously evaluate correct meat freshness by the time when the meat is sold. The latter is a very rapid but costly method. It is also very difficult for this method to evaluate slight differences in the meat freshness before the initial stage of putrefaction. The TVB-N content in chicken, as an important reference index, has been being used to evaluate chicken's freshness.4,5 TVB-N compounds in chicken contain mainly ammonia, trimethylamine (TMA) and dimethylamine (DMA) and the levels of TVB-N compounds increase with spoilage by either bacterial or enzymatic degradation. This method is however extremely time consuming, expensive and destructive, it is also not competent with modern industrial processing and production technologies. Therefore, the quality control in the meat industry demands the development of a freshness sensor, which can measure the meat freshness in situ, rapidly, simply, accurately and most preferentially in a non-destructive manner.
Electronic nose (E-nose), with the help of multivariate calibration techniques, represents an alternative approach for the evaluation of the freshness of meat or meat products.6 These techniques are alternative to traditional methods and are quick, easy to handle and do not require sample preparation or the use of chemical reagents. The sensors array in an E-nose system usually consists of numerous non-specific sensors and an odor stimulus generates characteristic fingerprint from the sensors array. Patterns of fingerprints from known odors are employed to construct a database and train a pattern recognition system so that unknown odors can subsequently be classified and identified. More detailed literature on E-nose can be referred to ref. 7. Most of E-nose systems consist of the Metal Oxide Semiconductor (MOS) sensors, which are usually conductometric in nature and their resistance changes (decreases/increases) when subjected to the odors vapor molecules. Therefore, this type of E-nose consisting MOS sensors array8 and carbon nanotubes (CNT)9 sensors are sensitive to the variation of humidity.
Recently, a novel low-cost colorimetric sensor array is being probed, which is not sensitive to humidity due to the hydrophobicity of the sensor materials and sensors plate. The design of the colorimetric sensor array is based on two fundamental requirements: (1) the chemo-responsive pigment must contain a centre (functional group) to interact strongly with analytes, and (2) this interaction centre must be strongly coupled to an intense chromophore. Chemo responsive pigments are those pigments that change color in either reflected or absorbed light, upon changes in their chemical environment.3,10 The basic principle of this method is the utilization of the color change induced by reaction between volatile compounds and an array of chemical dyes upon ligand binding for chemical vapor detection and differentiation. Chemical dyes are often selected according to their sensitivity to the specific Volatile Organic Compounds (VOCs).11,12 However, most researches focused on using the colorimetric sensors for the qualitative and quantitative study of meat freshness as in pork,13–17 fish.18 In our previous works, we have successfully developed novel artificial olfactory system based on colorimetric sensor array and implemented to identify spoilage bacteria,19 monitor vinegar acetic fermentation,20 classify Oolong tea,21 and rice wine,22 quantify chicken freshness.23,24
In order to quantify TVB-N content in chicken using colorimetric sensor array in our recent work, we have used linear and non-linear regression tools. Partial Least Squares Regression (PLSR) because it attempts to maximize the covariance thus capturing the variance and correlating the data together.25 When the experimental sensor data present non-linearity that might arise from the chemical nature of the target analytical parameter then the linear methods can lead to inaccurate predictions. These errors can be minimized by using non-linear multivariate techniques. One of the classic alternative is back-propagation artificial neural networks (BP-ANN),26 which can be effective when non-linear data responses are involved.27 We have introduced new algorithm in our previous works, by adaptive boosting BP-ANN and simply calling it BP-AdaBoost, with rather high performance prediction models.24,28 Another alternative is the Support Vector Machine Regression (SVMR), a relatively novel application of Support Vector Machine (SVM) learning algorithm proposed by ref. 29 is a promising alternative to existing linear and nonlinear multivariate statistical procedures. Hyper-parameters estimation in Support Vector Machines Regression (SVMR) is one of the major problems in the application of this type of algorithms to learning problems. At present, the most common optimization techniques are grid algorithm and genetic algorithm (GA).30 The grid algorithm is time consuming and does not perform well. Though the genetic algorithm (GA) is less time-consuming and can obtain the optimal solution well,31 the operation of genetic algorithm is difficult, the different types and rates of crossover and mutation need to be set for different optimal problem. PSO based SVMR is really seen to be a hot topic, not only because the vast areas are applying this technique, but also because the majority of works are very recent and interesting. PSO-SVMR approach has shown very good results in different applications as feature selection problem,32,33 biomedicine as identifying protein arginine methylation sites,34 prediction of cyanotoxin content,35 manufacturing, control, forest fire prediction,36 wind speed prediction,37 power load forecast,38 life of aircraft engines forecast39 and so on. However, evolutionary approach for parameter optimization as PSO has not been hybridized with SVMR to achieve best results in food quality assessment so far. Therefore, we focus the paper on the discussion of the application of PSO-SVMR to a real problem of quantifying chicken freshness, TVB-N content being the chemical reference measurement.
The objectives of this study are: (1) to establish the quantitative relationships between the colorimetric sensor data and TVB-N measurements of chicken breast fillets, and (2) to compare the predictive performances of calibration models established by linear i.e. partial least squares regression (PLSR) and non linear i.e. back-propagation artificial neural network (BP-ANN), AdaBoost BPANN, and Support Vector Regression (SVMR) by PSO.
Experimental materials and methods
Sample preparation
The chicken breast fillets were purchased from the local Auchan supermarket, Zhenjiang, Jiangsu Province, P.R. of China and transported to our laboratory within 20 min. The fillet was cut into pieces each weighing approximately 20 grams with the dimension of (4 × 4 × 1) cm. Such 75 samples were put in a sealed plastic bag and stored in a refrigerator at 4 °C before further analysis. For following 9 days, samples were randomly taken out from the refrigerator to determine its TVB-N content every another day (i.e. 1st, 3rd, 5th, 7th and 9th day).
Fabrication of colorimetric sensor array and data acquisition
We fabricated the sensor array first. For details of methods, see ESI.† The images of sensor array were captured by HP Scanjet 4890 flatbed scanner (Hewlett Packard Inc., Shanghai, China) as shown in Fig. 1. The scanner's resolution was set at 600 dpi.
 |
| Fig. 1 Colorimetric sensor array before exposure (a), colorimetric sensor array after exposure (b), and a difference image obtained after subtracting final sensor image from the initial image along with their corresponding data matrices (c). | |
Onto the data acquisition, the sensors array was captured by the flatbed scanner before exposure to the chicken sample that was considered as an ‘initial image’ (Fig. 1(a)). The typical diagram of the experimental setup of colorimetric sensory system for data acquisition is shown in Fig. 2. The sample was stored at 4 °C before data acquisition but the ambient temperature was controlled at 25 °C while sampling. In this experiment, we transferred the chicken sample from the sealed sample bag to the 250 mL glass beaker. The sensor array chip was mounted upside down in an inert platform i.e. thin plastic film wrapped over the lid of the glass beaker for its contact with the sample for exposure.
 |
| Fig. 2 Experimental setup for the colorimetric sensor array technique. | |
The uniform arrangements were made for parameters: the ambient temperature, the volume of sample, and the headspace time. On complete equilibration i.e. after 30 minutes of exposure, the sensor array was taken out to rescan and achieve a “final” image after exposure as shown in Fig. 1(b). This equilibration time of sensor reaction was determined by the preliminary experiments. According to the results of preliminary experiments, we found that the reaction between the dyes and VOCs were at complete equilibration after 30 min. To avoid factitious non-uniformity, the center of each dye spot (a round area consisting of 800 pixels) was averaged. Finally, a colorful difference image was obtained by simply subtracting the digital data matrix of “initial” image from that of “final” image; the difference image provided a color change profile that is a characteristic fingerprint to volatile oxidative compounds (VOCs) in chicken sample (for details, see ESI material†) as in Fig. 1(c).
Reference measurement
The reference measurement for the chicken freshness was carried out via TVB-N content in the sample. TVB-N content in chicken was measured by a steam distillation method, as per to the Chinese standard GB/T 5009.44.40 For details of methods, see ESI material.†
Multivariate calibrations
In our experiment, we divided 75 samples into 2 subsets: the calibration set, in which all samples were used for calibrating model and the prediction set, which we used to test the performance of the model. The division was as, calibration set
:
prediction set = 2
:
1 = 50
:
25 i.e. 1 sample out of every 3 samples were selected into the prediction set.
PLSR algorithm. Initially, the calibration set was subject to a classic linear regression tool, partial least squares regression (PLSR) with leave-one-out cross validation. The PLSR calibration models were evaluated using the coefficient of correlation (Rc) and the root-mean-square error cross validation (RMSECV). The performance of the final PLSR model was tested according to root mean square error of prediction (RMSEP) and the correlation coefficient (Rp) in prediction set. PLSR models were calculated using MATLAB R2010b (Math-Works, Natick, USA).
Back propagation-artificial neural network (BP-ANN). Back Propagation-Artificial Neural Network (BP-ANN) model was built. The back-propagating network architecture was subject to supervised training on selected samples known as the monitoring set. The BP-ANN was performed using the Neuroshell 2 software to extract best PC out of first 10 PCs latent variable. It was done based on the best Rp and RMSEP as shown in Fig. 5(b). Both learning rate factor and momentum factor were set 0.1; scale function was the ‘tanh’ function. The permitted regression error was set 0.01, and the maximal time of training was set 2000. For the further processing with the scatter diagrams of both the sets, data were processed through MATLAB R2010b (Math-Works, Natick, USA) on Windows 7.
BP-AdaBoost. The AdaBoost algorithm, short for adaptive boosting, introduced by Freund & Schapire (1995),41,42 solved many practical problems of the earlier boosting algorithms. Since boosting is, a method of combining performances of weak learners to build a strong classifier so that the performance would be better comparatively. We used the BP-ANN (i.e. an input layer, a hidden layer and an output layer) as the weak learning algorithm for AdaBoost and named it AdaBoost-BPANN prediction algorithm.
Support vector machine regression (SVMR) algorithm. The Support Vector Machine Regression (SVMR) algorithm has been focused in this work. SVMs are a popular machine learning method for classification, regression, and other learning tasks.43 The regression model was constructed by using a nonlinear mapping function (kernel function), which maps the input data (the extracted colorimetric sensor data) to a hyper dimensional feature space where the solution becomes linear and was generalized later to solve regression problems.44,45 The SVMR model tries to ignore the training data within a threshold ε to the model prediction so the model depends only on a subset of the training data, because the cost function for building the model. The SVM regression model can be expressed as:46,47 |
 | (1) |
where, k(x, xi) is the kernel function, xi is the input vector, αi is Lagrange multipliers called support value, β is the bias term. The radial basis function (RBF) has been used as a mapping function. Optimization of the meta-parameters C (regularization parameter), σ and ε (RBF kernel parameter) is the key step in SVMR as their combined values determine the boundary complexity and thus the regression performance.48 In order to perform this optimization, Particle Swarm Optimization (PSO) was implemented instead of using traditional GA and grid search method.
Particle swarm optimization (PSO) algorithm. A relatively new optimization technique, particle swarm optimization (PSO), is motivated by social behavior of organisms such as bird flocking and fish schooling.49 Here, a population (swarm) of N particles or proposed solutions evolves with each iteration, moving toward the optimal solution of the problem. A new population in the PSO algorithm is obtained shifting the positions of the previous one in each iteration. Each particle (composed of three parts, C, σ and ε) moves in the direction of its previously best position (pbest) and its best global position (gbest) to find the optimal solution. The method is one of the evolutionary algorithms and is very successful for it is easy to implement, understand and modify. There are just few parameters to adjust; maximum population was set to 20, maximum number of iterations was 200 and the constants c1 and c2 were set to 1.5 and 1.7 respectively.The swarm topology defines how particles are connected between them to interchange information with the global best. For details please refer to ref. 50. The corresponding results on Rp and RMSEP were compared to find out the best approach. The higher Rp and the lower RMSEP notifies that the prediction model is better. SVMR was performed by using MATLAB R2010b (Math-Works, Natick, USA) on Windows 7 using the Lib-SVM toolbox to derive the SVMR model.
Results and discussions
Reference measurements
The reference measurement result of TVB-N content for all samples is shown in Fig. 3. The bar graph shows the change in TVB-N content of chicken samples from day 1 until day 9. The tables show the details on chemical numerical values, its range as 5.74–42.661 and deviation as 10.9695. The TVB-N content with 15 mg/100 g is defined as the baseline of chicken freshness, the chicken sample with TVB-N content > 15 mg/100 g was defined as “stale” and otherwise the “fresh” ones according to Chinese Standard GB/T 5009.44 (2003).
 |
| Fig. 3 Bar graph showing chemical reference measurements of the chicken breast fillets with the statistics table. | |
Colorimetric sensor responses
Colorimetric sensor detects the odor changes of chicken samples during spoilage, which was produced by decomposition of the main internal chemical ingredients like protein, fat and carbohydrates. Fig. 1(c) shows the difference image of a sample. Each difference image has its specific colorific fingerprint. A color change profile for each sample can be obtained by differentiating the images of the sensor array before and after exposure to the VOCs of samples.12,51 The difference image is a RGB color image consisting of 3 color components images (i.e. R image, G image, and B image). The RGB image is an 8 bit image and the range of color values is [0 255]. A difference map is easily generated by digital subtraction, pixel by pixel, of the image of the array before and after exposure. As day passes by, microbial spoilage of chicken sample occurs during which a wide variety of volatile organic compounds (VOCs): hydrogen sulfide, dimethyl disulfide, indole, lactic acid, acetic acid, other fatty acids (propionic, isobutyric, isovaleric, n-butyric), C2–C5 alcohols, C6–C8 hydrocarbons, C3–C4 ketones, diacetyl-acetoin, putrescine, cadaverie, tyramine and other biogenic amines are produced.52 The metalloporphyrins dyes in the sensors array respond to most of the VOCs. The additional dyes of three pH indictors also respond to hydrogen sulfide and the organic acids. Microbial metabolites increased gradually along with the process of sample spoilage, and thus the sensors array has its unique colorific fingerprint to each sample corresponding to its freshness.
Model calibrations
PLSR model. Partial least square regression (PLSR) is one of the most classical and linear multivariate regression tools for modeling the relationship between dependent variables and independent Variables. The accuracy of the model was estimated using the parameters (Rp and RMSEP) obtained from the fitted equation. A good model should have the lowest RMSECV and RMSEP separated by a small difference and the highest Rp. PLSR method could somehow correlate the colorimetric sensor data to TVB-N content of chicken since Rp was 0.8293 and RMSEP was 6.13 as in Table 1. The scatter diagram showing the measured versus the predicted TVB-N content is shown in Fig. 4(a). The optimum number of Principle Component (PC) was found to be 7 based on the global histogram as in Fig. 5(a).
Table 1 Results of different linear and non-linear multivariate models for predicting TVB-N content (mg/100 g) of chicken breast fillets
S.n. |
Models |
Components/variables/support vectors |
Calibration set |
Prediction set |
RMSECV |
Rc |
RMSEP |
Rp |
1 |
PLSR |
7 |
8.32 |
0.7478 |
6.13 |
0.8293 |
2 |
BP-ANN |
6 |
5.18469 |
0.8972 |
5.1017 |
0.8627 |
3 |
BP-AdaBoost |
6 |
6.4702 |
0.8126 |
5.0282 |
0.8880 |
4 |
SVMR (by PSO) |
4 |
4.2227 |
0.9317 |
5.5255 |
0.8981 |
 |
| Fig. 4 Scatter diagram showing the measured versus predicted TVB-N content by linear (PLSR) (a) and non-linear multivariate regression algorithm, BP-ANN (b), BP-AdaBoost (c), and SVMR based on PSO (d). | |
 |
| Fig. 5 Parameter optimization by linear (PLSR) (a) and non-linear multivariate algorithm, BP-ANN (b), BP-AdaBoost (c), and SVMR based on PSO (d). | |
BP-ANN model. To improve the precision in the prediction of TVB-N content, the first 10 PC scores were used. The number of nodes in the input layer depended on the number of PCs, which was optimized by cross-validation, and the optimum number of PCs was determined according to the lowest RMSECV as in Fig. 5(b). The final BP-ANN model with PC of 6 in a 6:8:1 architecture with 8 nodes in the inner layer, resulted in a RMSEP = 5.1017 and R = 0.8627. The best results generated by BP-ANN are shown in Table 1. The performance of BP-ANN model was better than PLSR, which suggests that comparatively BP-ANN algorithm could establish a satisfactory quantitative relationship between chemical measurements and the colorimetric sensors data.
BP-AdaBoost model. AdaBoost is an ensemble method, which is possible to increase the accuracy of BP-ANN by averaging the decisions of ensemble of BP-ANN. The number of PCs and the prediction error threshold (Φ) have a significant effect on the AdaBoost-BPANN model, thus, they were optimized by cross-validation, and determined by the lowest RMSEP. The threshold (Φ) was optimized in a lager scope, and we find that when the parameter (Φ) was selected within 0.01–0.20, the model is ideal. Fig. 5(c) shows the RMSEP of AdaBoost-BPANN model with different PCs and thresholds. From Fig. 5(c), the lowest RMSEP could be obtained when the 6PCs and Φ = 0.03 were included. Eventually, the optimum AdaBoost-BPANN model was achieved with Rp = 0.8932, and RMSEP = 6.0766 g/100 mL in the prediction set. The scatter plot between reference measurement of TVB-N content and AdaBoost-BPANN predicted results is shown in Fig. 4(c).
PSO based SVMR model. In addition to PLSR, BP-ANN and BP-AdaBoost, we used the PSO module to optimize the SVMR parameters.53 The optimized condition would be the local parameter optimization.54,55 The PSO searches for the best C, σ and ε parameters by comparing the forecasting error in every iteration out of 200 iterations in our experiment. Search space is organized in three dimensions, one for each parameter. Table 1 shows the optimal prediction results in bold by the fitted SVMR model based on PSO technique.The goodness of fit of a statistical model describes how well it fits a set of observations. Indeed, it is important to select the model that best fits the experimental data. Here, in our research work, the goodness of fit was determined by the objective function as Mean Squared Error (MSE) (eqn (2)) to optimize the SVMR parameters C, σ and ε and it is completely based on the training data. The fitness of a statistical model describes how well it fits a set of observations.
The average relative error (ARE) is however used to evaluate the generalization performance of the SVMR model and is defined as in eqn (3), where the terms have the usual meaning, N is the test data.
|
 | (3) |
From the Fig. 5(d), we can see that the MSE and ARE values lead to roughly the same variation tendency in their curves when the iteration number was 200. The lowest MSE points with (C, σ and ε) have the lowest ARE values showing that the lowest MSE results in the best generalization performance (or lowest prediction error). ARE (in the table as correlation coefficient, Rc and Rp) can effectively reflect the generalization performance of regression models though its application is limited by the demand of test data different from the training set. As been mentioned earlier, c1 and c2 are the constants those were set 1.5 and 1.7 respectively. The optimal SVMR model was obtained with 4 support vectors when c was 3.2386, g = 0.01, and p = 0.01. The lowest MSE was 0.0846 with the elapsed time of 53.91 seconds and the Rp was 0.8981 along with the RMSEP of 5.5255 as shown in Table 1. The scatter diagram of measured against the predicted TVB-N content of chicken by PSO-based SVMR model is shown in the Fig. 4(d).
Discussions with comparison analysis
From the results and figures, it could be clearly analyzed that the non-linear model generated using PSO-SVMR algorithm outperformed the linear model by PLSR, as well as the non-linear models, BP-ANN and BP-AdaBoost. Therefore, the PSO-SVMR model was the best for our experiment and is a useful tool in the prediction of TVB-N content. PLSR produces a technique that is able to accept collinear data and separate out the sample noise in order to make linear combinations in the dependent concentration matrix.56 The correlation between the low cost sensor data and chemical reference data in chicken is nonlinear. Considering that, the spoilage of meat is rather complex process, where the nonlinear growth of microbiology generates the nonlinear accumulation of metabolites. The selected dyes in the sensors array have non-specific sensitivity and wide cross-sensitivity toward volatile compounds. Each dye in the sensors array could be simultaneously sensitive to numerous volatile compounds, and different dyes could be simultaneously sensitive to one of volatile compounds. Therefore, this sensors technique is not like the conventional component-by-component analyses (e.g., GC and GC-MS), and is difficult to assign specific colorific profile to a specific volatile compound. Moreover, the chemical reactions between the colorimetric dyes and the VOCs of metabolism are also extraordinarily complicated, thus leading to nonlinearity. Linear regression tools may not provide a complete solution to such nonlinear cases in this work. It justified that SVMR model as well as other non-linear regression models would be superior to PLSR model. Neural networks were chosen for the non-linear regression of our experimental data, BPANN and BP-AdaBoost. Although BP-ANN has proved its powerful capability in quantitative analysis, it also has its own deficiencies, which may lead to the following problems: (1) local minimum problem, (2) decreased rate problems, and (3) relatively low stability. BP-AdaBoost constructs a more powerful prediction system by developing a sequence composed by original forecasting algorithm, as has been justified by being better than BP-ANN from the results. AdaBoost algorithm takes the influence of weights into consideration and increases the iteration time apparently, hence the better results. The model produced by Lib-SVMR depends only on a subset of the training data, because the cost function for building the model ignores any training data close to the model prediction.56 SVMR is proved to be a better and intelligent approach for our work because SVMR neural network applies one hidden layer of non-linear neurons, one-output linear neuron and specialize learning procedure leading to the global minimum of the error function and excellent generalization ability of the trained network.57 When compared to the other neural networks, SVMR performed best in our research because principles wise, it adopted the structure risk minimization (SRM) principle. It is shown to be superior to the traditional empirical risk minimization (ERM) principle,58 employed by conventional neural networks like BP-ANN. SRM minimized an upper bound of the generalization error on the Vapnik–Chernoverkis dimension, as opposed to ERM, which minimized the training error. The SVMR method also had other interesting properties including an effective avoidance of over fitting, which improved its ability to build models using large numbers of molecular property descriptors with relatively few experimental results in the training set. This method had been proven to be very effective for addressing general purpose classification and regression problems.59 Therefore, non-linear approach is stronger than linear method in the level of self-learning and self-adjust. Besides, the topological architecture of SVMR might be more suitable for the solution to this regression problem in our research. However, the practicability of SVMR is affected due to the difficulty of selecting appropriate SVMR parameters. PSO has overcome the critical drawbacks of the traditional grid search method and the popular Genetic Algorithm (GA) method for SVMR parameter optimization. It is less time consuming than both the methods followed by simpler and easy parameters set up. Conclusively, non-linear model often has a simpler structure and a higher predictive precision of chemical data.60 This study shows that the colorimetric sensor array is one of the very low cost, rapid, and non-destructive quantitative measurement methods to predict the chicken freshness.
Conclusion
Colorimetric sensor array is a promising technique for the non-destructive quantification of TVB-N content in chicken. SVMR with PSO technique produced a global model capable of efficiently addressing non-linear relationships. The prediction performance was improvised using the SVMR model rather than other commonly used neural networks as well as linear regression tool. It can be concluded that colorimetric sensor technique coupled with SVMR regression tool combined with PSO has high potential to quantify TVB-N content in chicken in a rapid, intelligent and non-destructive manner.
Acknowledgements
This work has been financially supported by the National Natural Science Foundation of China (31371770). We are also grateful to many of our colleagues for stimulating discussion in this field.
References
- S. Barbut, World's Poult. Sci. J., 2010, 66, 399–410 CrossRef.
- N. Kaneki, T. Miura, K. Shimada, H. Tanaka, S. Ito, K. Hotori, C. Akasaka, S. Ohkubo and Y. Asano, Talanta, 2004, 62, 215–219 CrossRef CAS PubMed.
- C. Zhang, D. P. Bailey and K. S. Suslick, J. Agric. Food Chem., 2006, 54, 4925–4931 CrossRef CAS PubMed.
- P. Castro, J. C. P. Padrón, M. J. C. Cansino, E. S. Velázquez and R. M. D. Larriva, Food Control, 2006, 17, 245–248 CrossRef CAS.
- Ö. Fatih and Ö. Yeşim, Turk. J. Zool., 2000, 24, 113–120 Search PubMed.
- A. Loutfi, S. Coradeschi, G. K. Mani, P. Shankar and J. B. B. Rayappan, J. Food Eng., 2015, 144, 103–111 CrossRef CAS.
- M. Peris and L. Escuder-Gilabert, Anal. Chim. Acta, 2009, 638, 1–15 CrossRef CAS PubMed.
- J. H. Sohn, M. Atzeni, L. Zeller and G. Pioggia, Sens. Actuators, B, 2008, 131, 230–235 CrossRef CAS.
- Y. Ling, G. Gu, R. Liu, X. Lu, V. Kayastha, C. S. Jones, W.-S. Shih and D. C. Janzen, J. Appl. Phys., 2013, 113(2), 024312 CrossRef.
- K. S. Suslick, D. P. Bailey, C. K. Ingison, M. Janzen, M. E. Kosal, W. B. I. McNamara, N. A. Rakow, A. Sen, J. J. Weaver, J. B. Wilson, C. Zhang and S. Nakagaki, Quim. Nova, 2007, 30, 677–681 CrossRef CAS.
- M. C. Janzen, J. B. Ponder, D. P. Bailey, C. K. Ingison and K. S. Suslick, Anal. Chem., 2006, 78, 3591–3600 CrossRef CAS PubMed.
- N. A. Rakow and K. S. Suslick, Nature, 2000, 406, 710–713 CrossRef CAS PubMed.
- X. Huang, X. Zou, J. Zhao, J. Shi, Z. Li and T. Shen, Int. J. Food Sci. Technol., 2015, 50, 203–209 CrossRef CAS.
- X.-W. Huang, X.-B. Zou, J.-Y. Shi, Y. Guo, J.-W. Zhao, J. Zhang and L. Hao, Food Chem., 2014, 145, 549–554 CrossRef CAS PubMed.
- H. Li, Q. Chen, J. Zhao and Q. Ouyang, Anal. Methods, 2014, 6, 6271–6277 RSC.
- Y. Salinas, J. V. Ros-Lis, J.-L. Vivancos, R. Martinez-Manez, M. Dolores Marcos, S. Aucejo, N. Herranz, I. Lorente and E. Garcia, Food Control, 2014, 35, 166–176 CrossRef.
- H. Xingyi, Z. Fang and J. Feiyan, J. Agric. Mach., 2011, 42, 142–146 Search PubMed.
- M. K. Morsy, K. Zór, N. Kostesha, T. S. Alstrøm, A. Heiskanen, H. El-Tanahi, A. Sharoba, D. Papkovsky, J. Larsen, H. Khalaf, M. H. Jakobsen and J. Emnéus, Food Control, 2016, 60, 346–352 CrossRef CAS.
- Q. Chen, H. Li, Q. Ouyang and J. Zhao, Sens. Actuators, B, 2014, 205, 1–8 CrossRef CAS.
- Q. Chen, A. Liu, J. Zhao, Q. Ouyang, Z. Sun and L. Huang, Sens. Actuators, B, 2013, 183, 608–616 CrossRef CAS.
- Q. Chen, C. Sun, Q. Ouyang, Y. Wang, A. Liu, H. Li and J. Zhao, LWT--Food Sci. Technol., 2015, 60, 781–787 CrossRef CAS.
- Q. Ouyang, J. Zhao, Q. Chen and H. Lin, Food Chem., 2013, 138, 1320–1324 CrossRef CAS PubMed.
- Q. Chen, Z. Hui, J. Zhao and Q. Ouyang, LWT--Food Sci. Technol., 2014, 57, 502–507 CrossRef CAS.
- K. Urmila, H. Li, Q. Chen, Z. Hui and J. Zhao, Anal. Methods, 2015, 7, 5682–5688 RSC.
- X. Huang, X. Zou, J. Zhao, J. Shi, X. Zhang, Z. Li and L. Shen, Meat Sci., 2014, 98, 203–210 CrossRef CAS PubMed.
- J. Hertz, A. Krogh and R. G. Palmer, Introduction to the Theory of Neural Computation, Addison-Wesley, Redwood City, CA, 1991 Search PubMed.
- N. C. T. Mariani, R. C. Da Costa, K. M. G. de Lima, V. Nardini, L. C. Cunha Júnior and G. H. D. A. Teixeira, Food Chem., 2014, 159, 458–462 CrossRef CAS PubMed.
- Q. Chen, W. Hu, J. Su, H. Li, Q. Ouyang and J. Zhao, J. Food Eng., 2016, 168, 259–266 CrossRef.
- J. A. K. Suykens, T. van Gestel, J. de Brabanter, B. de Moor and J. Vandewalle, Least squares support vector machines, World Scientific Publishing Company, Singapore, 2002 Search PubMed.
- E. Avci, Expert Syst. Appl., 2009, 36, 1391–1402 CrossRef.
- S. H. Min, J. Lee and I. Han, Expert Syst. Appl., 2006, 31, 652–660 CrossRef.
- C.-L. Huang and J.-F. Dun, Applied Soft Computing, 2008, 8, 1381–1391 CrossRef.
- S.-W. Lin, K.-C. Ying, S.-C. Chen and Z.-J. Lee, Expert Syst. Appl., 2008, 35, 1817–1824 CrossRef.
- Y. Zhang, L. Tang, H. Zou, Q. Yang, X. Yu, J. Jiang, H. Wu and R. Yu, Chemom. Intell. Lab. Syst., 2015, 146, 102–107 CrossRef CAS.
- P. J. Garcia Nieto, J. R. Alonso Fernandez, V. M. Gonzalez Suarez, C. Diaz Muniz, E. Garcia-Gonzalo and R. Mayo Bayon, Appl. Math. Comput., 2015, 260, 170–187 CrossRef.
- Y. Wang, J. Wang, W. Du, C. Wang, Y. Liang, C. Zhou and L. Huang, in Advances in Neural Networks – Isnn 2009, Part 2, Proceedings, ed. W. Yu, H. B. He and N. Zhng, 2009, vol. 5552, pp. 382–390 Search PubMed.
- S. Salcedo-Sanz, E. G. Ortiz-García, Á. M. Pérez-Bellido, A. Portilla-Figueras and L. Prieto, Expert Syst. Appl., 2011, 38, 4052–4057 CrossRef.
- Q. Wu, Expert Syst. Appl., 2010, 37, 194–201 CrossRef.
- P. J. Garcia Nieto, E. Garcia-Gonzalo, F. Sanchez Lasheras and F. J. de Cos Juez, Reliab. Eng. Syst. Saf., 2015, 138, 219–231 CrossRef.
- GB/T, China National Standard: Method for analysis of hygienic standard of meat and meat products, 2003.
- M. H. Zhang, Q. S. Xu, F. Daeyaert, P. J. Lewi and D. L. Massart, Anal. Chim. Acta, 2005, 544, 167–176 CrossRef CAS.
- Y. Freund and R. Schapire, in Computational Learning Theory, ed. P. Vitányi, Springer, Berlin, Heidelberg, 1995, vol. 904, ch. 2, pp. 23–37 Search PubMed.
- C.-C. Chang and C.-J. Lin, ACM Trans. Intell. Syst. Technol., 2011, 2(3), 27 Search PubMed.
- H. R. Safavi and M. Esmikhani, Water Resour. Manag., 2013, 27, 2623–2644 CrossRef.
- A. Suarez Sanchez, P. J. Garcia Nieto, P. Riesgo Fernandez, J. J. del Coz Diaz and F. J. Iglesias-Rodriguez, Math. Comput. Model., 2011, 54, 1453–1466 CrossRef.
- B. Schölkopf and A. J. Smola, Learning with kernels: Support vector machines, regularization, optimization, and beyond, MIT press, Cambridge, Massachusetts, USA, 2002 Search PubMed.
- J. Shawe-Taylor and N. Cristianini, Kernel methods for pattern analysis, Cambridge university press, New York, 2004 Search PubMed.
- O. Devos, C. Ruckebusch, A. Durand, L. Duponchel and J.-P. Huvenne, Chemom. Intell. Lab. Syst., 2009, 96, 27–33 CrossRef CAS.
- J. Kennedy, R. C. Eberhart and Y. Shi, in Swarm Intelligence, ed. J. Kennedy and R. C. E. Shi, Morgan Kaufmann, San Francisco, 2001, pp. 287–325, DOI:10.1016/b978-155860595-4/50007-3.
- X. Wang, J. Wen, Y. Zhang and Y. Wang, Optik, 2014, 125, 1439–1443 CrossRef.
- K. S. Suslick, N. A. Rakow and A. Sen, Tetrahedron, 2004, 60, 11133–11138 CrossRef CAS.
- Y. Salinas, J. V. Ros-Lis, J.-L. Vivancos, R. Martinez-Manez, M. Dolores Marcos, S. Aucejo, N. Herranz and I. Lorente, Analyst, 2012, 137, 3635–3643 RSC.
- X. Shi and J. Zhou, Adv. Sci. Lett., 2012, 11, 238–243 CrossRef.
- Y. Ren, H. Liu, C. Xue, X. Yao, M. Liu and B. Fan, Anal. Chim. Acta, 2006, 572, 272–282 CrossRef CAS PubMed.
- F. Luan, R. Zhang, C. Zhao, X. Yao, M. Liu, Z. Hu and B. Fan, Chem. Res. Toxicol., 2005, 18, 198–203 CrossRef CAS PubMed.
- S. Qiu, L. Gao and J. Wang, J. Food Eng., 2015, 144, 77–85 CrossRef CAS.
- K. Brudzewski, S. Osowski and T. Markiewicz, Sens. Actuators, B, 2004, 98, 291–298 CrossRef CAS.
- C. J. C. Burges, Data Min. Knowl. Discov., 1998, 2, 121–167 CrossRef.
- J. Wang, H. Du, X. Yao and Z. Hu, Anal. Chim. Acta, 2007, 601, 156–163 CrossRef CAS PubMed.
- H. Jiang, Q. Chen and G. Liu, Process Biochem., 2014, 49, 583–588 CrossRef CAS.
Footnote |
† Electronic supplementary information (ESI) available. See DOI: 10.1039/c5ra25375f |
|
This journal is © The Royal Society of Chemistry 2016 |
Click here to see how this site uses Cookies. View our privacy policy here.