emabalajes que evitan contaminacion (1).pdf

8/10/2019 Emabalajes que evitan contaminacion (1).pdf

1/7

Available online at www.sciencedirect.com

Sensors and Actuators B 128 (2008) 435441

Self-organizing algorithm for classification of packaged fresh vegetablepotentially contaminated with foodborne pathogens

Ubonrat Siripatrawan

Department of Food Technology, Faculty of Science, Chulalongkorn University, Bangkok, Thailand

Received 26 March 2007; received in revised form 25 June 2007; accepted 27 June 2007

Available online 1 July 2007

Abstract

A rapid method for identification of foodborne pathogens contamination in packaged fresh vegetable using electronic sensor array and Kohonenself-organizing map (SOM) algorithm was developed.Escherichia coli was used as thetarget microorganism because its presence in foods indicates

fecal contamination, and the presence of pathogenic microorganisms. E. coli was grown in the packaged fresh vegetable. The electronic sensors

was used to monitor changes in the composition of the package headspace gas phase relating to the biochemical products of E. coli volatile

metabolites. SOM algorithm was then used to classify the data output from the electronic sensor array. The SOM algorithm created a map from a

high dimensional input vector space onto a two-dimensional output lattice. When integrated with SOM algorithm, the electronic sensors proved to

have the ability to classify the packaged fresh vegetable potentially contaminated with pathogens.

2007 Elsevier B.V. All rights reserved.

Keywords: Self-organizing map; Kohonen algorithm; Neural network; Foodborne pathogens; Contamination; Metal oxide sensors

1. Introduction

Numerous outbreaks of foodborne diseases strengthen the

need for rapid and sensitive methods for detection of foodborne

pathogens. Classical methods for the identification and clas-

sification of microorganisms are based on their biochemical,

morphological serological and toxigenic characteristics. These

methods usually require intact viable organisms and a series

of tests requiring the incubation of the microorganisms [1,2].

Early pathogens detection is important to implement disease

control measures[3].Recently, research has focused on devel-

opment of rapid and accurate techniques to identify pathogens

in food products[46].Zhao et al.[7] developed a disposable

electrochemical immunosensor for detection of Vibrio para-

haemolyticus(VP) based on the screen-printed electrode (SPE)

coated with agarose/Nano-Au membrane and horseradish per-

oxidase (HRP) labeled VP antibody (HRP-anti-VP). Wu et al.

[8] applied QCM systemin thedetection of PCR-amplifiedDNA

from real samples ofEscherichia coliO157:H7. The piezoelec-

tric biosensor detected the presence ofE. coli O157:H7 when

Fax: +66 2 2544314.

E-mail address:[email protected].

the DNA strand was complementary to the immobilized probes

with synthetic oligonucleotides.Microorganismscan be characterized by identification of spe-

cific metabolites generated by specific biochemical pathways.

The selection of volatiles for use as incipient disease indicators

has been reviewed in terms of the composite rate of pathogenic

destruction within food products[911].This concept has been

actualized in electronicsensor arrayor electronic nose [12,13].

A considerable number of electronic sensor applications have

been reported, including classification of changes in milk result-

ing from a variety of heat treatments [14], evaluation of the

off-odor in wine[15],quality measurement of smoked salmon

[16]and detection ofSalmonellain nutrient media[17].Elec-

tronic sensor technology is usually based on a hybrid sensor

array system with different selectivity and sensitivity, with the

result being a powerful analytical instrument especially for com-

plex food analyses. However, analysis of volatile compounds

using electronic sensor array often generates large and com-

plicated data set and difficult to interpret if used directly. A

mathematical resolution of complex data is usually performed

in far less time than it takes to conduct physical or chemical

experiments[1719].

Linear methods which have been used to classify the

data include linear discriminant analysis (LDA) and princi-

0925-4005/$ see front matter 2007 Elsevier B.V. All rights reserved.

doi:10.1016/j.snb.2007.06.030
mailto:[email protected]://localhost/var/www/apps/conversion/tmp/scratch_7/dx.doi.org/10.1016/j.snb.2007.06.030http://localhost/var/www/apps/conversion/tmp/scratch_7/dx.doi.org/10.1016/j.snb.2007.06.030mailto:[email protected]


2/7

436 U. Siripatrawan / Sensors and Actuators B 128 (2008) 435441

pal component analysis (PCA). The LDA and PCA is a linear

transformation that is well suited for separating image/signal

data for different objects or class [20]. The main advantage

of linear transforms is that they are easy to design and typi-

cally have closed-form solutions. However, linear transforms

typically extract information from only the second-order corre-

lations in the data (covariance matrix) and ignore higher-order

correlations in the data. Many researchers have suggested that

many signals in the real world are inherently non-symmetric

[21].A number of nonlinear transformation methods for pattern

recognition exist. Artificial neural networks (ANNs) are among

the most commonly used nonlinear techniques.

The most important features of ANN are their learning and

adaptation abilities. According to their learning strategies, ANN

can be classified as supervised and unsupervised networks. In

supervised learning, each time the ANN is exposed to a train-

ing input, the related class information is required as well.

The multilayer perceptrons (MLP) neural network or the feed

forward ANN has been the most popular. The term unsuper-

vised means that the knowledge of environment is not learnedfrom the specific inputoutput examples. Self-organizing map

(SOM) is an unsupervised artificial neural network which is

frequently used for data partitioning and classification. SOM

can be used for grouping of complex sample data without

any strict assumption and without any priori knowledge of

the number of groups present[20]. Lin and Wang [22] com-

pared SOM with various hierachical cluster analysis methods.

The result shows that the performance of SOM in clustering

messy data is better than that of the other hierachical clustering

methods.

The principle of SOM is characterized by the formation of

a topographic map of the input patterns in which the spatiallocations of the neurons in the lattice are indicative of intrin-

sic statistical features contained in the input patterns[23].The

SOM can be considered as a grid with predefined nodes. Prior to

learning, a large unit area that surrounds the winner is selected

as a neighborhood region. During learning, the pattern of filling

the nodes is determined by the degree of similarity between the

data. If an input vector is presented to the SOM network, the

weight vector in the network that is closest to the input vec-

tor is selected as the best-matching (winner) node. The wining

mapping node is defined as that with the smallest Euclidean dis-

tance between the mapping node vector and the input vector

[2426].

Although various rapid methods for detection of microorgan-isms have been developed, no research has used electronic nose

coupled with SOM to classify the contamination of pathogens

directly from the packaged food products. Hence, this research

was aimed to develop a method to identify E. coli contamina-

tion in packaged fresh vegetable using electronic sensor array

coupled with Kohonen neural network. E. coli is a common

member of the normal flora of the large intestine. In this study,

E. coliwas used as the target microorganism in packaged alfalfa

sprouts because its presence in foods indicates fecal contamina-

tion, and the presence of pathogenic microorganisms. Alfalfa

sprouts were chosen as the product component because the

National Advisory Committee on Microbial Criteria for Foods

(NACMCF) [27] identified sprouts as a special problem due

to the potential for pathogen growth during production, while

there is increasing demand for sprouts due to their popularity as

a healthy food[28,29].The electronic nose was used to monitor

the volatiles produced byE. coli. SOM algorithm was used as an

experimental platform (in addition to the instrumental methods)

to identifyE. colicontamination.

2. Materials and methods

2.1. Preparation of inoculated vegetable

The alfalfa seeds (Natural Sprout Company, Springfield,MO)

were soaked in 20,000ppm of calcium hypochlorite prior to ger-

mination as advised by the U.S. Food and Drug Administration

[28]and NACMCF[27].Alfalfa sprouts were grown in a labo-

ratory environment at 20 C and 65% RH with indirect sunlight

and away from any possible contaminations. The sprouts were

harvested after 5 days (fully grown) when length is 3.84cm.

The sprouts were washed and drained several times before use.

The nonpathogenic strainE. coli ATCC 25922 obtained from

the American Type Culture Collection (ATCC, Rockville, MD)

was cultured in tryptic soy broth and incubated at 37 C for

8 h in a gyrotory shaker and centrifuged. Broth was poured

from the culture and the sedimented pellet was resuspended

in sterile Butterfields phosphate buffer which was used as a

dipping suspension. Preliminary experiments were conducted

to determine the population of E. coli necessary in the dip-

ping suspension to result in an initial population of105 CFU/g

on sprouts. Preliminary studies also showed that the electronic

nose was able to detect volatiles produced by E. coli when

the number ofE. coli was higher than 105 CFU/g. The sproutswere then placed in screened baskets, and submerged in the

suspension containingE. colifor 3 min. The uninoculated con-

trol was similarly treated except sterile phosphate buffer was

used in place of the inoculum. Fifty grams of sprouts were

then packed into commercial 1.5-mil, 15 cm 8 cm linear low-

density polyethylene (LDPE) bags and heat-sealed. The total

volume of thesprouts was200 ml which occupied about half of

the total bagvolume. Thesamples were incubated at 10 Cfor1

3 days.

2.2. Microbiological analysis

The microbial cell count was determined on the date of inoc-

ulation and periodically throughout storage at days 13. Serial

dilutions were prepared from the stock suspension, and Petri

plates were inoculated with those dilutions expected to give

countable colonies. Inocula consisting of each of a dilution seri-

als were deposited on prepared plates in duplicate using 3M

Petrifilm Aerobic Count Plates (3M, St. Paul, MN) for deter-

mining aerobic bacteria and 3M PetrifilmE. coli/Coliform Count

Plates (3M, St. Paul, MN) containing Violet Red Bile nutrient

agar as an indicator of glucuronidase activity for E. coli. All

plates were incubated at 37 C for 48 h. Plate countsare recorded

as colony forming units (CFU/g).


3/7

U. Siripatrawan / Sensors and Actuators B 128 (2008) 435441 437

2.3. Electronic nose analysis

A total of 120 samples, including non-inoculated alfalfa

sprouts (control) and alfalfa sprouts inoculated with E. coli at

time zero, were incubated for 13 days prior to analysis.

An electronic nose (Fox 3000, Alpha M.O.S., Hillsborough,

NJ) was used for monitoring changes in volatiles produced by

E. coli growing on the sprouts. The volatile analysis system

combines a measurement chamber for generating the volatile

compounds and a detection system made up of 12 metal oxide

sensors (SYLG, SYG, SYAA, SYGH, SYGCTI, SYGCT, T301,

P101, P102, P401, T702, and PA2). This instrument was linked

to an auto-sampler capable of analyzing a total of 64 samples.

Samples were placed in the HS100 auto-sampler in arbitrary

order. Five millilitre was collected from the headspace of pack-

aged alfalfa sprouts and injected into the electronic sensor. The

temperature of the injection syringe was 40 C. The delay time

between two injections was 300 s. Each injection was repeated,

with separate samples. The electronic signals from the sensors

were digitized and then transferred to the control computer.

2.4. Self-organizing neural network

Data were made up of 120 samples from8 subgroups(sprouts

(SP) and sprouts inoculated with E. coli (EC) in LDPE bags

on the first day of inoculation and incubated at 10 C for 13

days). Each subgroup had 15 replicate samples collected from

several cultivations. Each sample was analyzed using 12 metal

oxide sensors. The sensor responses of all 120 samples were

arrangedina120 12 matrix. Data classificationwas performed

using SOM algorithm. All calculations were carried out using

MATLAB 5.2 routines written by the authors, and making use ofthe toolbox provided by Mathworks (Mathworks, Inc., Natick,

MA).

3. Results and discussion

3.1. Microbial cell counts

The number of aerobic bacteria and E. coli on the alfalfa

seeds was determined. The number of aerobic bacteria was

101102 CFU/g, while noE. coli were found. The alfalfa seeds

were soaked in 20,000 ppm of calcium hypochlorite prior to ger-

mination as advised by NACMCF[27].This treatment has the

potential to substantially reduce microbial contamination whichcan be passedon to the growing sprouts.Gillet al. [30] suggested

that chemical disinfection can reduce the human risk for disease

posed by contaminated seed sprouts. The number of aerobic bac-

teria in alfalfa seeds increased from 101102 to 107 CFU/g

when the alfalfa sprouts were fully-grown. The conditions dur-

ing sprouting (e.g. time, temperature, water activity, pH, and

nutrient level) may have promoted the growth of microflora

[28,30],without affecting the smell, taste or appearance of the

sprouts. Thus, the risk of foodborne disease associated with

sprouts increases during sprouting[27].

The cell counts of aerobic bacteria andE. coli on fully-grown

sprouts with and withoutE. coliinoculation are shown inFig. 1.

Fig. 1. Growth of total aerobic bacteria and E. coliof uninoculated vegetable

(SP) and vegetable inoculated with E. coli(EC).

All samples had a high number of total aerobic bacteria. How-ever,E. coliwas not found in the control samples. The numbers

ofE. coli in the inoculated samples increased from105 CFU/g

on the first day of inoculation to 107 CFU/g after 3 days incu-

bation.

3.2. Electronic sensor array

Each sensor element changes in resistance (max) when

exposed to volatile compounds. The information from electronic

sensor array analysis was extracted from the series of sensor

resistances. In order to produce consistent data for the classifica-

tion, the sensor response was presented with a volatile chemical

relative to the base resistance in air, which is the maximum

change in the sensors electrical resistance divided by the initial

resistance, as follows

Relative resistance change =max 0

0(1)

wheremaxis the maximum change in the sensors electrical

resistance and 0 is the initial baseline resistance of the sen-

sor. The relative resistance change was used for data evaluation

because it gives the most stable result, and is more robust against

sensor baseline variation.

The data matrix comprised 120 samples from 8 subgroups

(SP-D0, SP-D1, SP-D2, SP-D3, EC-D0, EC-D1, EC-D2, andEC-D3) as analyzed using the 12 sensors (SYLG, SYG, SYAA,

SYGH, SYGCTI, SYGCT, T301, P101, P102, P401, T702, and

PA2).Fig. 2shows the average responses of all samples in 8

subgroups to the 12 metal oxide sensors. In Fig. 2, the sensi-

tivities of all samples are compared. These values express the

average sensor responses of each sensor in the range of mea-

surement. Since the sensor outputs from the 12 different sensors

are not homogeneous, a direct comparison of sensitivities is not

adequate to interpret the information from the samples. The sen-

sor responses from an array of nonspecific metal oxide sensors

are generally insufficient to discriminate between a series of

samples.


4/7


Fig. 2. Electronic sensor responses form the headspace of packaged samples

labeled using thefollowing scheme: SP-D0,SP-D1, SP-D2and SP-D3 arepack-

aged vegetable kept at 10 C for 03 days, respectively; EC-D0, EC-D1, EC-D2

and EC-D3 are packaged vegetable inoculated with E. coli on the first day of

inoculation and after stored at 10 C for 13 days, respectively.

In this study, unsupervised SOM was used for analysis of

electronic sensor array data. Although, the supervised MLP neu-

ral network has been the most popular to model the complex

data, it requires prior training pairs (input vectors and corre-

sponding target vectors) to make training possible. Therefore,

the MLP may be unable to provide a real-time response to detect

the contaminated samples. Self-organizing map is an unsuper-

vised neural network which does not need any class information

for learning, but acquires that knowledge by itself during the

training phase through cluster formation. For SOM, the only

input is needed to construct an output. The SOM algorithm cre-ates a mapping from a high dimensional input vector space onto

a two-dimensional output lattice. The SOM network is basically

composed of a single, two-dimensional layer of neurons which

helps provide a visual presentation of data.

3.3. Self-organizing neural network

The unsupervised SOM was used to classify electronic

voltametric response outputs by replacement of the actual

data points using topographic map reference vectors. A two-

dimensional Kohonen output layer was used to help provide a

visual presentation. According to Lee et al. [31],selecting theappropriate number of output nodes is quite difficult and this

is usually experiment-dependent. There is no consensus among

researchers about the subject. To obtain good mapping results,

the number of output nodes in the Kohonen neural network

should be at least 1020% of training vectors. However, using

too few output nodes may cause the congestion of input vectors

over an output node, which may make it difficult to distinguish

the characteristics of the output space.

A Kohonen network consisting of 55 nodes was employed

for classification of the 8 subgroups from the input data matrix

(12120). The predefined neuron number (grid size) in the

Kohonen outputlayerwas chosenbecauseit wassufficient to dis-

tinguish different sample groups. In SOM process, the mapping

nodes are first initialized with random numbers. The SOM is

initialized by assigning small random values to all of the weight

vector elements. The algorithm responsible for the formation

of the SOM proceeds first by initializing the synaptic weight

in the network. Once the network has been properly initialized,

there are three essential processes involved in the application

of the algorithm including sampling, similarity matching, and

updating[25,32,33].

For the sampling process, a sample xfrom the input space

was drawn with a certain probability. Let xdenote an input vec-

tor selected randomly from the input data space and m denotes

the dimension of the input space. The vector x represents the

activation pattern that is applied to the lattice.

x = [x1, x2, . . . , xm]T (2)

In the similarity matching step, the best matching, winning

neuron i(x) attime step twas determined by using the minimum-

distance Euclidean criterion:

i(x) = arg min||x(t) wj(t)|| (3)

wherewjdenotes the synaptic weight vector of neuronj. Ifan m-

dimensional input vector is presented to the SOM network, then

the weight vector in the network that is closest to the input vector

is selected as the best-matching node. The particular neuron i

which is the best matching is called the winning neuron for the

input vectorx.

For the updating process, the synaptic weight vectors of all

neurons were adjusted using the update formula:

wj(t+ 1) = wj(t) + (t)hj,i(x)(xi(t) wj(t)) (4)

hj,i(x) = exp

d2j,i

22

(5)

wherex(t) is the input to node iat timetandwj(t) is the weight

from input node i to output node j at timet.(t) is the learning

rate parameter, hj,i(x)(t) is the neighborhood function centered

around the winning neuroni(x),d2j,iis the distance between the

winning neuron i and the adjacent neuronj, and is the width of

the topological neighborhood [32]. Thewinning node is selected

as the center of a neighborhood, in order to reduce the Euclidean

distance. At timet, the cell learns this input signal. During the

next time t + 1, it hasan information processing ability ofwj(t+

1), which is close to the input signal. The neighboring units thatsurround i(x) also learn the input vector x(t) by following the

same equation[33].

For training, the data vectors are first arranged in random

order and then presented in this order to the neural network for

training. In SOM, the neurons adaptively tend to learn the prop-

erties of the underlying distribution of the space in which they

operate. Additionally, they also tend to learn their places topo-

logically. The training consists of finding the winning neuron,

which is the one whose pattern has the best match andmodifying

the winning node and its closest neighbors in the neuron map

by moving their associated feature vectors closer to the input

vectors[25,32].


5/7


For the training cycle decision, there is no definitive stopping

point.A huristic is to useenough training cycles so that a network

approaches a stable state [20].As the learning progresses, the

mapbecomes more andmore structured. Iterations arecontinued

until all of the weight vectors are stabilized, and hence cluster

regions for each candidate target are properly established on the

SOM output map. At the end of the learning process, the feature

map is spanned over all input values. The network was trained

for 10,000 epochswith patterns selected randomly as opposed to

sequentially. During the training phase the map unfolds to form

a mesh. The neurons are mapped in the correct order at the end of

this phase. The SOM is configured on a two-dimensional feature

space consisting of discrete lattice nodes (the set of discrete

lattice nodes is denoted as the neuron map). Each node in the

lattice hasa feature vector. Thenodesare initialized with random

patterns and the SOM is subsequently trained iteratively.

After training, the algorithm organizes the sensor response

topologically on a two-dimensional grid, in which each node

corresponds to a weight vector. The algorithm constructs these

responsesby nonlinearinterpolation of the responsesin the train-ing set. SOM learns both the distribution and topology of the

input vectors they are trained on. The weight vector most closely

approximates that specific input vector, is chosen as the win-

ning neuron. The SOM feature map usually adjusts its weights

quickly to their inputs. Fig. 3 shows the plot of weight vectors of

winning neurons of each sample class. The SOM visualization

consists of running the same input file against the trained map

and reporting the map grid location that is closest in Euclidean

distance to each input. By labeling each neuron on the map

with the appropriate subgroup terms (Fig. 4),the clustering of

natural groups can be discovered from the electronic response

data. The performance of the SOM was measured as percent ofmisclassification. The SOM gave only 2.5% misclassification.

The SOM pattern provides information on sample classifi-

cation. For instance, node 4 had a weight pattern most similar

Fig. 3. Weight vectors of the trained network labeled using the following

scheme: SP-D0,SP-D1, SP-D2 andSP-D3are packaged vegetable kept at 10 C

for03 days, respectively;EC-D0,EC-D1, EC-D2and EC-D3are packagedveg-

etable inoculated withE. colion the first day of inoculation and after stored at

10

C for 13 days, respectively.

Fig. 4. Self-organizingmap of sensor responses on a 55 rectangular grid with

clusters indicating the sample subgroups labeled using the following scheme:

SP-D0, SP-D1, SP-D2 and SP-D3 are packaged vegetable kept at 10 C for 03

days, respectively; EC-D0, EC-D1, EC-D2 and EC-D3 are packaged vegetable

inoculated with E. coli on the first day of inoculation and after stored at 10 C

for 13 days, respectively.

to the input pattern for EC-D3 (vegetable inoculated with E.

coliand incubated for 3 days), and therefore, this node was rep-

resentative of this subgroup. The output patterns generated by

the Kohonen network for the different volatile fingerprints show

that the self-organizing pattern was able to distinguish different

subgroups with different number ofE. coli.However, samples

from the first day of inoculation (EC-D0) and the control on thefirst day of preparation (SP-D0) overlapped. This suggested that

the capability of the electronic sensors to detect the volatiles

produced by E. coli occurred when the number ofE. coli was

higher than 105 CFU/g. The SOM method can be considered

an alternative way to the classification schemes. Dimensional-

ity reduction from the input space to the network field was also

accomplished using this algorithm.

The volatile metabolites from the headspace of packaged

alfalfa sprouts inoculated withE. coliwere analyzed using elec-

tronic nose. The volatile compounds present in the headspace of

inoculated sprouts and absent in the headspace of uninoculated

sprouts can be used as possible indicators ofE. coli contami-

nation. The electronic nose was used to monitor changes in thecomposition of the gas phase of biochemical products from E.

colivolatile metabolites directly from packaged alfalfa sprouts

(without culturing in standard media). In this research, the elec-

tronic nose has shown potential to detect specificE. colivolatile

metabolites, even though the sprouts contained high aerobic

counts. The primary advantage of the electronic nose in a qual-

ity assurance method is in its speed of analysis, including data

acquisition and interpretation. Rapid, significant data interpreta-

tion is possible using unsupervised SOM neural network. SOM

is able to find relationship between data, grouping and mapping

them topologically. The developed method has potential in the

real-time detection ofE. coli. This method is easy to use and


6/7


enable continuous operation. However, its future lies in attain-

ing selectivities and sensitivities comparable to conventional

methods.

This system is not limited to theaforementionedapplications.

The system can be applied to other packaged food products or

incorporated into HACCP protocols or quality control systems

in the food industries.

4. Conclusion

The electronic nose has shown potential to detect specificE.

coli volatile metabolites, even though the sprouts contained high

aerobic counts. However, the capability of the electronic nose to

detect the volatiles produced byE. colioccurred when the num-

ber ofE. coliwas higher than 105 CFU/g. The algorithm based

on Kohonen self-organizing map was used to organize topolog-

ically sensor response inputs into clusters on a 5 5 rectangular

grid feature map. The SOM algorithm created a map from a high

dimensional input vector space onto a two-dimensional outputlattice. The location of the node in the grid provides information

about the different sample groups. SOM can be used for analyz-

ing multicomponent data, in order to get classification of data

from an analysis of the contribution that each sensor brings to

the whole array. The SOM neural network was shown to provide

patterns which were more easily distinguished than the original

sensor response patterns for similar types of interactions. This

provided the basis of an effective feature extraction method in

that distance values generated by the network allowed better dis-

crimination of a fingerprint. The SOM profiling method can

be considered an alternative way to the classification schemes.

Dimensionality reduction from the input space to the network

field was also accomplished using this algorithm. The sensorarray coupled with SOM has the potential to be a sensitive,

fast, one-step method to identify E. coli contamination in the

packaged samples.

References

[1] A. Gharaibeh, K.J. Voorhees, Characterization of lipid fatty acids in

whole-cell microorganisms using in situ supercritical fluid derivatiza-

tion/extraction and gas chromatography/mass spectrometry, Anal. Chem.

68 (1996) 28052810.

[2] F. Basile, M.B. Beverly, C. Abbas-Hawks, C.D. Mowry, K.J. Voorhees,

Direct mass spectrometric analysis of in situ thermally hydrolyzed andmethylated lipids from whole bacterial cells, Anal. Chem. 70 (1998)

15551562.

[3] P. Skottrup, M. Nicolaisen, A.F. Justesen, Rapid determination of

Phytophthora infestans sporangia using a surface plasmon resonance

immunosensor, J. Microbiol. Methods 68 (2007) 507515.

[4] Council for Agricultural Science and Technology, Foodborne Pathogens:

Review of Recommendations, available from: http://www.cdc.gov/

ncidod/EID/vol9no4/99-0519.html , October 1999.

[5] U. Siripatrawan, B.R. Harte, Solid phase microextraction/gas chromato-

graph/mass spectrometer integrated with chemometrics for detection of

Salmonella typhimurium contaminationin a packagedfreshvegetable, Anal

Chim. Acta. 581 (2007) 6370.

[6] G. Amagliani, C. Giammarini, E. Omiccioli, G. Brandi, M. Magnani,

Detection of Listeria monocytogenes using a commercial PCR kit and

different DNA extraction methods, Food Control 18 (2007) 11371142.

[7] G. Zhao, F. Xing, S. Deng, A disposable amperometric enzyme

immunosensor for rapid detection of Vibrio parahaemolyticus in food

based on agarose/Nano-Au membrane and screen-printed electrode, Elec-

trochem. Commun. 9 (2007) 12631268.

[8] V.C.H. Wu, S. Chen, C.S. Lin, Real-time detection of Escherichia coli

O157:H7 sequences using a circulating-flow system of quartz crystal

microbalance, Biosens. Bioelectron. 22 (2007) 29672975.

[9] O. Canhoto, F. Pinzari, C. Fenelli, N. Magan, Application of electronic

nose technology for the detection of fungal contamination in library paper,Int. Biodeter. Biodegr. 54 (2004) 303309.

[10] A. Jonsson, Electronic nose for microbial quality classification of grains,

Int. J. Food Microbiol. 35 (1997) 187193.

[11] L. Marilley, S. Amupuero, T. Zesiger, M.G. Casey, Screening of aroma-

producing lactic acid bacteria with electronic nose, Int. Dairy J. 14 (2004)

849856.

[12] J.P. Stetter, S. Strathmann, C. McEntegart, M. Decastro, W.R. Penrose,

New sensor arrays and sampling systems for a modular electronic nose,

Sens. Actuators B: Chem. 69 (2000) 410419.

[13] S. Thomas, R. Bro, A new approach for modeling sensor based data, Sens.

Actuators B: Chem. 102 (2005) 719729.

[14] K. Brudzewski,S. Osowski, T. Markiewicz, Classificationof milkby means

of an electronic nose and SVM neural network, Sens. Actuators B: Chem.

98 (2004) 291298.

[15] J.A. Ragazzo-Sanchez, P. Chalier, C. Ghommidh, Coupling gas chro-matography and electronic nose for dehydration and desalcoholization of

alcoholized beverages application to off-flavour detection in wine, Sens.

Actuators B: Chem. 104 (2005) 301308.

[16] J.E. Haugen, K. Rudi, S. Langsrud, S. Bredholt, Application of gas-

sensor array technology for detectionand monitoring of growth of spoilage

bacteria in milk: a model study, Anal. Chim. Acta 565 (2006) 10

16.

[17] U. Siripatrawan, J.E. Linz, B.R. Harte, Electronic sensor array coupled

with artificial neural network for detection of Salmonella typhimurium,

Sens. Actuators B: Chem. 119 (2006) 6469.

[18] V. Devabhaktuni, M.C.E. Yagoub, Y. Fang, J. Xu, Q. Zhang, Neural net-

works for microwave modeling: model development issues and nonlinear

modeling techniques, Int. J. RF Microw. Comput. Aided Eng. 11 (2001)

421.

[19] M.H. Terra, R. Tinos, Fault detection and isolation in robotic manipulators

via neural networks: a comparison among three architectures for residual

analysis, J. Rob. Syst. 18 (2001) 357374.

[20] J.L. Giraudel, S. Lek, A comparison of self-organizing map algorithm and

some conventional statistical methods for ecological community ordina-

tion, Ecol. Model. 146 (2001) 329339.

[21] B. Scholkopf, A. Smola, K.R. Muller, Nonlinear component analysis as a

kernel eigenvalue problem, Neural Comput. 10 (1998) 12991319.

[22] G.F. Lin, C.M. Wang, Performing cluster analysis and discrimination anal-

ysis of hydrological factors in one step, Adv. Water Resour. 29 (2006)

15731585.

[23] T.Kohonen,Self-organizingneuralprojections, Neural Networks19 (2006)

723733.

[24] C. Fernandez, E. Soria, J.D. Martyn, A.J. Serrano, Neural networks for

animal science applications: two case studies, Expert Syst. Appl. 31 (2006)

444450.

[25] M. Maeda, H. Miyajima, S. Murashima, Construction of self-organizing

algorithms for vector quantization, Electron. Eng. Jpn. 127 (1999) 4755.

[26] S. The, C.P. Lim, Monitoring the formation of kernel-based topographic

maps in a hybrid SOM-kMER model, IEEE Trans. Neural Networks 17

(2006) 13361341.

[27] National Advisory Committee on Microbial Criteria for Foods, Micro-

biological Safety Evaluations and Recommendations on Sprouted Seeds,

available from: http://www.cfsan.fda.gov/mow/sprouts2.html, October

26, 1999.

[28] U.S. Food and Drug Administration, Press release, Consumers advised

of risks associated with raw sprouts, P99-13, Guidance for industry

guide to minimize microbial food safety hazards for sprouted seeds,

available from: http://www.fda.gov/bbs/topics/NEWS/NEW00684.html,

October 27, 1999.
http://www.cdc.gov/ncidod/EID/vol9no4/99-0519.htmlhttp://www.cdc.gov/ncidod/EID/vol9no4/99-0519.htmlhttp://www.cfsan.fda.gov/~mow/sprouts2.htmlhttp://www.cfsan.fda.gov/~mow/sprouts2.htmlhttp://www.cfsan.fda.gov/~mow/sprouts2.htmlhttp://www.fda.gov/bbs/topics/NEWS/NEW00684.htmlhttp://www.fda.gov/bbs/topics/NEWS/NEW00684.htmlhttp://www.cfsan.fda.gov/~mow/sprouts2.htmlhttp://www.cdc.gov/ncidod/EID/vol9no4/99-0519.htmlhttp://www.cdc.gov/ncidod/EID/vol9no4/99-0519.html


7/7


[29] Centers for Disease Control and Prevention, Infections Associ-

ated with Eating Seed Sprouts, available from: http://www.cdc.

gov/ecoli/2006/september/, 2006.

[30] J. Gill, W.E. Keene, J.E. Mohle-Boetani, J.A. Farrar, P.L. Waller, C.G.

Hahn, R.P. Cieslak, Alfalfa seed decontamination in Salmonellaoutbreak,

Emerg. Infect. Dis. 9 (2003) 474479.

[31] K. Lee, D. Booth, P. Alam, A comparison of supervised and unsupervised

neural networks in predicting bankruptcy of Korean firms, Expert Syst.

Appl. 29 (2005) 116.[32] S. Haykin, Neural Networks: A Comprehensive Foundation, second ed.,

Prentice Hall, New Jersey, 1999, p. 541.

[33] T. Tanaka, M. Saito, Quantitative properties of Kohonens self-organizing

maps as adaptive vector quantizers, Trans. IEICE 75 (1992) 10851092.

Biography

Ubonrat Siripatrawanobtained her PhD degree in packaging from Michigan

State University. She is an assistant professor in Department of Food Technol-

ogy, Faculty of Science, Chulalongkorn University. Her research interests coverelectronic sensors, GCMS and ANN for rapid detection of microorganisms in

packaged food products, food shelf life simulations and active packaging.
http://www.cdc.gov/ecoli/2006/september/http://www.cdc.gov/ecoli/2006/september/http://www.cdc.gov/ecoli/2006/september/http://www.cdc.gov/ecoli/2006/september/

emabalajes que evitan contaminacion (1).pdf

Documents