Prediction of total viable counts on chilled pork using an electronic nose combined with support vector machine

Meat Science 90 (2012) 373–377

Contents lists available at ScienceDirect

Meat Science j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / m e a t s c i

Prediction of total viable counts on chilled pork using an electronic nose combined with support vector machine Danfeng Wang a, Xichang Wang a, Taiang Liu b, Yuan Liu a,⁎ a b

College of Food Science and Technology, Shanghai Ocean University, Shanghai 201306, China Department of Chemistry, College of Science, Shanghai University, Shanghai 200444, China

a r t i c l e

i n f o

Article history: Received 28 September 2010 Received in revised form 2 July 2011 Accepted 31 July 2011 Keywords: Pork Electronic nose Total viable counts Prediction Support vector machine

a b s t r a c t The aim of this study was to predict the total viable counts (TVC) in chilled pork using an electronic nose (EN) together with support vector machine (SVM). EN and bacteriological measurements were performed on pork samples stored at 4 °C for up to 10 days. Bacterial numbers on pork were determined by plate counts on agar. Principal component analysis (PCA) was used to cluster EN measurements. The model for the correlation between EN signal responses and bacterial numbers was constructed by using the SVM, combined with partial least squares (PLS). Correlation coefficients for training and validation were 0.94 and 0.88, respectively, which suggested that the EN system could be used as a simple and rapid technique for the prediction of bacteria numbers in pork. © 2011 Elsevier Ltd. All rights reserved.

1. Introduction There is a general need in the food supply chain for rapid methods to monitor microbial quality, and to identify hygienic and safety conditions in order to enable necessary corrective actions at the appropriate time. Safety evaluations of pork products are also needed because the industry and regulatory bodies need to ensure that pork is safe for the market and meets the highest demands (El Barbri, Llobet, El Bari, Correigand Bouchikhi, 2008). Microorganism can be considered as the most serious factor affecting the quality and safety of pork and the number of microorganisms will directly affect the quality of pork. It has been recommended that the critical value in ralation to spoilage, total viable counts (TVC) have been set at 106 cfu/g (Gram et al., 2002; Panigrahi, Balasubramanian, Gu, Logue, & Marchello, 2006). The electronic nose (EN) is an instrument consisting of sensors, which in specific cases, are able to precisely detect and distinguish odors in complex food samples (Moy & Collins, 1996; Pregla et al., 2005; Peris & Escuder-Gilabert, 2009). The EN can detect and classify microorganism on the basis of Microbial Volatile Organic Compounds (MVOCs). It can also be used for microbial detection in environmental monitoring (Dutta, Hines, Gardner, & Boilot, 2002; Dutta, Morgan, Baker, Gardner, & Hines, 2005; Kuske, Romain, & Nicolas, 2005), clinical diagnosis (Parry, Chadwick, Simon, Oppenheim, & McCollum, 1995), and food analysis (Limbo, Sinelli, Torri, & Riva, 2009; Zhang et al., 2009; Zhang, Tong, Chen, & Lan, 2008).

⁎ Corresponding author. Tel.: + 86 21 61900380. E-mail address: [email protected] (Y. Liu). 0309-1740/$ – see front matter © 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.meatsci.2011.07.025

Earlier studies have shown the release of volatiles during storage of foods can result from rapid degradation by bacterial processes (F. Korel, 2001; Peris & Escuder-Gilabert, 2009; Du et al., 2002). In addition, the EN has been successfully used to monitor the shelf life/freshness of red meats (El Barbri, Llobet, El Bari, Correigand Bouchikhi, 2008; Horvath, Seregely, Dalmadi, Andrassy, & Farkas, 2007). This paper discusses the use of SVM for feature extraction and development of a prediction model for predicting the TVC in pork stored at 4 °C from the data obtained using an EN. The purpose of this paper is to evaluate the EN performance as an instrument for the quality control of pork meats and is part of a larger study on the use of EN for sensing food quality and safety.

2. Materials and methods 2.1. Sample preparation and sampling All analyses were performed using a local brand of chilled pork M. longissimus dorsi obtained from a market of Shanghai. The samples were divided into two parts, one was immediately cut into pieces of similar weight (25 g ± 0.1 g) and then placed in a white polystyrene foam tray and overwrapped with PE film (O2 transmission rate (OTR) = 18500± 40% cm3/ (m2·24 h·atm); Top Group CO., Ltd., China). It was then placed in a refrigerator kept at a constant temperature of (4 °C± 1 °C). The other half was cut into pieces 3–5 cm2 cubed and then further chopped and minced and samples were then transferred into 10 ml sterile vials suitable for the EN and were tightly capped and stored at 4 °C. Measurements were performed every day until the 10th day. On each sampling day, microbial analysis was performed in triplicate, and

374

D. Wang et al. / Meat Science 90 (2012) 373–377

0.4

50s

Intensity

0.3

0.2

5s 0.1

0.0

0

10

20

30

40

50

60

70

80

90

100

110

120

Times (s) Fig. 1. Time responses of an array of eighteen gas sensors at the first day.

six replicates were used for analysis with the EN. The experiments were repeated 4 times, and all data was used for statistical analysis. 2.2. Estimation of bacterial counts Three 25 g samples were taken aseptically and each was placed in a sterile stomacher bag containing 225 ml of 0.1% (w/v) peptone water (PW), (BEIJING LAND BRIDGE TECHNOLOGY Co., Ltd., China). Each sample and PW was stomached for 2 min. Ten-fold dilutions of the meat samples were prepared using PW. These dilutions were then plated on the surface of a Plate Count Agar (PCA), (BEIJING LAND BRIDGE TECHNOLOGY Co., Ltd., China). Colonies were counted following incubation at 37 °C ± 1 °C) for 48 h, and calculated as log10 colony forming units (cfu)/g of sample. 2.3. Electronic nose analysis

Total number of aerobic bacteria (log cfu/g)

The volatile components in the headspace of the samples were detected using an Alpha M.O.S. sensor array system (Alpha M.O.S., FOX4000. France). The instrument consisted of a sampling apparatus, a detector unit containing the array of sensors, and a pattern of recognition software for data recording and elaboration. The sensor array was composed of 18 metal oxide sensors (MOS). The sensors were divided into three systems: Sensor chamber 1: chamber CL (High Performance Controlled in temperature) LY2/LG, LY2/G, LY2/AA, LY2/ GH, LY2/gCTL, LY2/gCT; Sensor chamber 2: chamber A (High Performance Controlled in temperature) T30/1, P10/1, P10/2, P40/1, T70/2, PA/2; Sensor chamber 3: chamber B (High Performance Controlled in

8 7 6 5 4 3 0

2

4

6

8

10

Storage days Fig. 2. Kinetic curve of microbial growth in chilled pork at 4 °C.

temperature) P30/1, P40/2, P30/2, T40/2, T40/1, TA/2. The reference gas was filtered and dried air, with a purity quotient N99.999%. 2.3.1. Sample data, feature extraction and pre-processing The EN sensor signal responses of the meat samples were calculated using the following expression: R = ðR0 −Rt Þ = R0

where R is the EN sensor signal response, Rt is the value of the conductance of the MOS sensors (value of the sensor conductance in the presence of the meat sample) and R0 is the value of the MOS sensors at time 0 for each meat sample. In order to get a better correlation between EN regression and bacterial analysis, feature data were extracted. A typical example of the temporal responses of the sensor array at the first day is shown in Fig. 1. From the changes of these sensors, it can be seen that the values from 0 s to 5 s contain little useful information. In addition, the values after 50 s begin to level off and show a similar trend in all samples. So it can be understood that the data from 0 to 5 s and after 50 s can be negligible without affecting the original information. Therefore this redundant information was deleted and so the data from 6 to 50 s was extracted and is referred to as the feature extraction values. Correlation between EN signal responses and microbial loads were then established after feature extraction of data used by PLS and SVM. 2.3.2. Support vector machine (SVM) SVM is a relatively new machine learning technique for materials computation. It has been shown to perform well for classification and regression (Luts et al., 2010; Vapnik, 1998; Xu, Wencong, Chunrong, Qiang, & Jin, 2009). SVM has two advantageous features; a) ability to generalize and b) suitable for material optimization where sample size is limited. SVM has two distinct features which are high generalization ability and especial suitability for small size of training sets that are useful for materials optimization. Through multiple calculations, the optimal Support Vector Regression (SVR) model was determined utilizing a Gaussian kernel to radial basis function (RBF) with C = 10 and ε = 0.01. C is a regularized constant determining the tradeoff between the training error and the model flatness. ε is a prescribed parameter. Gaussian (RBF) Kernel function is: ! ‖xy‖2 K ðx; yÞ = exp σ2

D. Wang et al. / Meat Science 90 (2012) 373–377

375

0.4 0.3

6-10days

0.2

0-5days

0.1 -0.0

PC2 - 5.433%

-0.1

0D 1D 2D 3D 4D 5D 6D 7D 8D 9D 10D

-0.2 -0.3 -0.4 -0.5 -0.6 -0.7 -0.8 -0.9 -1.7 -1.6 -1.5 -1.4 -1.3 -1.2 -1.1 -1.0 -0.9 -0.8 -0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6

PC1 - 91.351% Fig. 3. PCA chart of chilled pork under different storage days.

corresponded to those samples having undergone up to 5 days of storage whereas the second group were those having undergone from 6 up to 10 days of storage. Thus there appeared to be a close correlation between the responses of EN sensor signal and the numbers of bacteria.

3. Results and discussion 3.1. Bacterial analysis The growth of aerobic bacteria (log10cfu/g) on pork samples at 4°C over 10 days was used as the basis for the prediction of the total bacteria numbers with the electronic nose (Fig. 2). Total viable counts (TVC) showed only a slight variation during the first four days (from 4 to 5 log10cfu/g) but then showed a large increase between days four and eight (from 5 to 7.25 log10cfu/g) and then remained at that level. The TVC on the first day represented the level of initial surface contamination. During the period of chilled storage, bacterial counts increase as nutrients become more readily available as a result of autolytic changes in the meat (Mahony, Green, Baylis, Fernandes, & Papkovsky, 2009). After 5 to 6 storage days, TVC increased to over 6 log10cfu/g.

3.3. Correlation between EN regression and bacterial analysis To investigate whether the results of the EN signal responses did in fact correlate well with those of microbial loads, a PLS-SVM calibration model was developed and validated. The underlying objective was to assess if the EN could predict microbial counts (TVC) on pork meat. In this work, the PLS method (Bastien, Vinzi, & Tenenhaus, 2005) was applied to dimension reduction. 15 components were used to build the SVR model as features. In order to assess the quality of the model, the leave-one-out cross-validation (LOOCV) (Amendolia et al., 2003; Holden, 1996; Kearns & R, 1997) was used. The optimal model was evaluated with the least root mean square error (RMSE), Mean

PCA has been widely used to analyze electronic nose data as it can reduce the dimensionality of original data and thus simplify the analysis without loss of original data (Costache, Corcoran, & Puslecki, 2009; Kim, Kim, & Bang, 2003; Zhao & Jiang, 2006). In order to give a more clear explanation, a PCA chart of chilled pork stored at 4 °C is shown (Fig. 3), using projection vectors determined by the first and second principal components. It can be seen that the values cluster mainly into two different groups, although there were several outliers, particularly for days 5 and 6 of storage. The first group

Table 1 RMSE, MRE and R of TVC of chilled pork of SVR.

TVC(predicted, log cfu/g)

3.2. PCA analysis

8 7 6 5 4 3 3

Items

RMSE

MRE (%)

R

Trained model of SVR LOOCV test of SVR

0.49 0.71

8.16 11.33

0.94 0.88

4

5

6

7

8

9

TVC(experimental, log cfu/g) Fig. 4. Experimental vs. predicted values of TVC of chilled pork by the model of SVR (r = 0.94).

D. Wang et al. / Meat Science 90 (2012) 373–377

TVC(predicted, log cfu/g)

376

4. Conclusions

8 7 6 5 4 3 3

4

5

6

7

8

9

TVC(experimental, log cfu/g) Fig. 5. Experimental vs. predicted values of TVC of chilled pork in LOOCV test of SVR (using RBF kernel, c = 10, ε = 0.1, and r = 0.88).

Relative Errors (MRE) and Correlation(R) between experimental and predicted.

RMSE =

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u n u u ∑ ðpi −ei Þ2 ti = 1

A positive correlation has been shown to exist between EN signal responses and bacteriological loads. The results are very promising considering the possible applications of an EN as a tool to predict the TVC in pork during chilled storage. The advantage of this analysis is that it is rapid and readily performed, simply by analyzing headspace from pork samples. However, the major disadvantage is that EN used is very large, heavy and expensive, which will limit its use for widespread testing. In order to make the method feasible for widespread use in meat industry, and for examples in supermarkets, there is a need to develop a smaller and cheaper EN device, specifically for meat samples. Such an EN would include only those appropriate sensors for to predicting the TVC in pork and would therefore incur cost less than larger EN devices.

Acknowledgments This work was supported by the Leading Academic Discipline Project of Shanghai Municipal Education Commission (Project No. J50704), the National Natural Science Foundation of China (Grant No. 30901125) and Innovation Program of Shanghai Municipal Education Commission (08YZ117). Thanks are expressed to Dr. Ron Tume from CSIRO of Australia for his kind help in checking the manuscript.

n

j

1 n pi −ei MRE = ∑ n i=1 ei

References

j

where ei is the experimental value of i sample, pi is the predicted value of i sample, n is the number of the whole samples in LOOCV. Table 1 lists the results of the SVR model. Plots of the predicted TVC versus the experimental TVC using the trained model of SVR with Gaussian (RBF) kernel function (where C = 10 and ε = 0.01) is shown in Fig. 4. Fig. 5 shows the plots of the predicted TVC versus the experimental TVC in LOOCV test by using SVR with Gaussian (RBF) kernel function (where C = 10 and ε = 0.01). Based on Table 1, Fig. 4 and Fig. 5, it was found that the SVR is an effective modeling tool for the data set in this work. A very good correlation was obtained between the EN signal responses and bacterial numbers. Correlation coefficients of 0.94 and 0.88 for training and validation, respectively were obtained. This good agreement can be explained because the electronic nose analyses the development of volatile components that result from the bacterial degradation of pork. Due to the high correlation existing between EN signal responses and bacterial numbers, the EN could be used as a rapid and alternative way for TVC prediction in pork meat. In order to test the validity of the trained SVR model, 4 additional pork samples were used and handled as the other samples. The predicted and the experimental results (Table 2) show that the values of deviation and relative deviation are very small, the maximum deviation and relative deviation being 0.24 and 5.67% respectively. The results indicated the potential of EN for predicting TVC in chilled pork during storage.

Table 2 Results of TVC prediction by using the SVR model for the independent test data of chilled pork. No.

TVC (Exp.)

TVC (Pred.)

D (Pred.-Exp.)

D/Exp. (%)

1 2 3 4

3.98 4.23 5.14 7.17

4.09 4.47 4.96 7.05

0.11 0.24 − 0.18 − 0.12

2.76 5.67 3.50 1.67

Amendolia, S. R., Cossu, G., Ganadu, M. L., Golosio, B., Masala, G. L., & Mura, G. M. (2003). A comparative study of K-Nearest Neighbour, Support Vector Machine and MultiLayer Perceptron for Thalassemia screening. Chemometrics and Intelligent Laboratory Systems, 69, 13–20. Bastien, P., Vinzi, V. E., & Tenenhaus, M. (2005). PLS generalised linear regression. Computational Statistics & Data Analysis, 48, 17–46. Costache, G. N., Corcoran, P., & Puslecki, P. (2009). Combining PCA-based datasets without retraining of the basis vector set. Pattern Recognition Letters, 30, 1441–1447. Du, W. X., Lin, C. M., Huang, T., Kim, J., Marshall, M., & Wei, C. I. (2002). Potential Application of the Electronic Nose for Quality Assessment of Salmon Fillets Under Various Storage Conditions. Journal of Food Science, 67(1), 307–313. Dutta, R., Hines, E. L., Gardner, J. W., & Boilot, P. (2002). Bacteria classification using Cyranose 320 electronic nose. Biomedical Engineering Online, 1, 4. Dutta, R., Morgan, D., Baker, N., Gardner, J. W., & Hines, E. L. (2005). Identification of Staphylococcus aureus infections in hospital environment: electronic nose based approach. Sensors and Actuators B: Chemical, 109, 355–362. El Barbri, N., Llobet, E., El Bari, N., Correig, X., & Bouchikhi, B. (2008). Electronic nose based on metal oxide semiconductor sensors as an alternative technique for the spoilage classification of red meat. Sensors, 8, 142–156. F. Korel, D. A. L., M.Ö. Balaban, JFS: Sens. Nutr. (2001). Qualities Food 66, 1018. Gram, L., Ravn, L., Rasch, M., Bruhn, J. B., Christensen, A. B., & Givskov, M. (2002). Food spoilage — interactions between food spoilage bacteria. International Journal of Food Microbiology, 78, 79–97. Holden, S. B. (1996). PAC-like upper bounds for the sample complexity of leave-one-out cross validation, Paper Presented at: Ninth Annual ACM Workshop on Computational Learning Theory. Jun. 28; Desenzano del Garda, Italy. Horvath, K. M., Seregely, Z., Dalmadi, I., Andrassy, E., & Farkas, J. (2007). Estimation of bacteriological spoilage of pork cutlets by electronic nose. Acta Microbiologica et Immunologica Hungarica, 54, 179–194. Kearns, Mj., & R, D. (1997). Algorithmic stability and sanity-check bounds for leave-oneout cross-validation. Proceedings of the Tenth Annual ACM Workshop on Computational Learning Theory, Nashville, Tennessee (pp. 152–162). New York: ACM Press. Kim, H. -C., Kim, D., & Bang, S. Y. (2003). An efficient model order selection for PCA mixture model. Pattern Recognition Letters, 24, 1385–1393. Kuske, M., Romain, A. -C., & Nicolas, J. (2005). Microbial volatile organic compounds as indicators of fungi. Can an electronic nose detect fungi in indoor environments? Building and Environment, 40, 824–831. Limbo, S., Sinelli, N., Torri, L., & Riva, M. (2009). Freshness decay and shelf life predictive modelling of European sea bass (Dicentrarchus labrax) applying chemical methods and electronic nose. LWT-Food Science and Technology, 42, 977–984. Luts, J., Ojeda, F., Van de Plas, R., De Moor, B., Van Huffel, S., & Suykens, J. A. K. (2010). A tutorial on support vector machine-based methods for classification problems in chemometrics. Analytica Chimica Acta, 665, 129–145. Mahony, F., Green, R. A., Baylis, C., Fernandes, R., & Papkovsky, D. B. (2009). Analysis of total aerobic viable counts in samples of raw meat usingfluorescence-based probe and oxygen consumption assay. Food Control, 20, 129–135. Moy, L., & Collins, M. (1996). Electronic noses and artificial neural networks. American Laboratory, 28, 22. Panigrahi, S., Balasubramanian, S., Gu, H., Logue, C. M., & Marchello, M. (2006). Design and development of a metal oxide based electronic nose for spoilage classification of beef. Sensors and Actuators B: Chemical, 119, 2–14.

D. Wang et al. / Meat Science 90 (2012) 373–377 Parry, A. D., Chadwick, P. R., Simon, D., Oppenheim, B., & McCollum, C. N. (1995). Leg ulcer odour detection identifies beta-haemolytic streptococcal infection. Journal of Wound Care, 4, 404–406. Peris, M., & Escuder-Gilabert, L. (2009). A 21st century technique for food control: Electronic noses. Analytica Chimica Acta, 638, 1–15. Pregla, R., Loke, F., Damm, E., Heizmann, W. R., Boger, Z., & Hetzer, R. (2005). Artificial electronic nose is able to detect and discriminate bacterial species by their smell. The International Journal of Artificial Organs, 28(9), 876. Vapnik, V. (1998). Statistical learning theory. New York: John Wiley and Johns. Xu, L., Wencong, L., Chunrong, P., Qiang, S., & Jin, G. (2009). Two semi-empirical approaches for the prediction of oxide ionic conductivities in ABO3 perovskites. Computational Materials Science, 46, 860–868.

377

Zhang, Z., Tong, J., Chen, D. H., & Lan, Y. B. (2008). Electronic nose with an air sensor matrix for detecting beef freshness. Journal of Bionic Engineering, 5, 67–73. Zhang, S. P., Xie, C. S., Bai, Z. K., Hu, M. L., Li, H. Y., & Zeng, D. W. (2009). Spoiling and formaldehyde-containing detections in octopus with an E-nose. Food Chemistry, 113, 1346–1350. Zhao, J., & Jiang, Q. (2006). Probabilistic PCA for t distributions. Neurocomputing, 69, 2217–2226.