
Research Article


Background Interference Elimination in Wound Infection Detection by Electronic Nose Based on Reference Vectorbased Independent Component Analysis


Fengchun Tian,
Jia Yan,
Shan Xu,
Jingwei Feng,
Qinghua He,
Yue Shen
and
Pengfei Jia


ABSTRACT

Background interference is serious and widespread problem in wound infection detection by electronic nose (ENose). When mice are used as experimental subjects, the background interference, i.e., the odor of the mice themselves, is very strong and useful information is often buried in it. A new method of eliminating the background interference and detecting wound infection, based on an ENose in cooperation with reference vectorbased Independent Component Analysis (ICA) denoising algorithm is proposed. It employs ICA to decompose each signal of the sensor array and extract the independent components and then discriminates the useful sources and Background interference through the Correlation with the reference Vector. The independent components of which the background interference had been eliminated are used as the inputs of Radial Basis Function (RBF) network for discrimination. The result shows that this method is effective and practical for background interference elimination in the detection of wound infection by ENose.




How
to cite this article:
Fengchun Tian, Jia Yan, Shan Xu, Jingwei Feng, Qinghua He, Yue Shen and Pengfei Jia, 2012. Background Interference Elimination in Wound Infection Detection by Electronic Nose Based on Reference Vectorbased Independent Component Analysis. Information Technology Journal, 11: 850858. DOI: 10.3923/itj.2012.850.858 URL: https://scialert.net/abstract/?doi=itj.2012.850.858



Received:
December 10, 2011; Accepted: March 28, 2012;
Published: May 29, 2012 

INTRODUCTION
An electronic nose (ENose), which is composed of an array of gas sensors as
well as the corresponding pattern recognition algorithm, is able to imitate
the olfaction system of humans and mammals and is used for the recognition of
gas and odor (Hines et al., 1999; Bicego
et al., 2002; Ciosek and Wroblewski, 2006).
It plays a constantly growing role in the identification and quantification
of odor (Abd ElAziz, 2011) and has become a powerful
tool to detect of vapor chemicals in disease diagnostics (Turner
and Magan, 2004; Gardner et al., 2000) such
as for lung diseases (Anh et al., 2005; Di
Natale et al., 2003; Phillips et al.,
2003) diabetes (Yu et al., 2005; Ryabtsev
et al., 1999), urinary tract disease (Lin et
al., 2001; Pavlou et al., 2002).
Wounds are a big public hazard in the world and the incidence of two severe
complications of wounds, infection and Multiple Organ Failure (MOF), is a serious
concern for doctors. Rapid and timely monitoring of traumatic inflammation,
especially identification of type of wound infection and infection levels of
bacteria, is conducive to guiding the doctor's diagnosis and treatment. It takes
long time for the traditional bacteriological diagnosis on wound infection and
therefore the treatment is often delayed. The defense reaction of posttraumatic
bodies contains the inflammatory response resulted from injured and broken cells,
which is a sterile inflammatory process and inflammatory response caused by
bacteria when secondary infection appears. Although, the two inflammatory responses
are different in nature, it is difficult to distinguish clinically in early
stage. It is dependent on the experience of physicians to determine whether
the wound is infected before obtaining the evidence of bacteriology and immunology
detection. And the two most important ways for physicians to diagnose are to
observe the characters of excreta of the wound and to smell the wound odor.
The bacterial infected wounds have a variety of special smell before the obvious
excreta appear. If we can discriminate the special smell of different bacterial
wounds, it will be conducive for rapid and timely diagnosis of wound infection.
However, the smell ability of human is not sensitive enough to discriminate
the odor. The ENose has much better sensitivity compared to that of human being.
So the use of ENose technology to achieve a faster and easier diagnosis of wound
infection is feasible. Previous works have demonstrated that it is feasible
to use ENose to detect bacteria including investigation of bacterial Volatile
Organic Compounds (VOCs) from cultures and also from swabs taken from wound
infected patients (Gibson et al., 1997; Dutta
et al., 2002; Setkus et al., 2006;
Persaud et al., 2008; Thomas
et al., 2010; Byun et al., 2010).
Independent Component Analysis (ICA) is a multivariate statistical method which is used to extract hidden components of signals, only given observed signals that are assumed to be linear mixtures of some unknown sources. The purpose of ICA is to decompose original signals into several mutually independent components, using statistics of order greater than two from the probability densities of the signals. In this process, we don’t know other priori knowledge except that the sources are statistically independent. ICA developed accompanied by Blind Source Separation (BSS) problem and it is a way to solve the BSS.
In a complex background, the received signals are often mixtures of different
information sources. Usually, it may appear that, we cannot obtain effective
discrimination among various measurement samples by directly putting the original
signals to the pattern classifier. So, in order for good prediction or classification.
It is necessary to make suitable data preprocessing to obtain useful information
from the data. ICA is based on the statistics independence between the information
sources. Compared with the traditional filtering and cumulative average methods,
it hardly damages the useful details of the signals in removing the noise; its
denoising performance is usually better than the traditional filtering method.
Compared with the traditional method of signal separation based on feature analysis,
such as Principal Component Analysis (PCA) (Khan et al.,
2004) and Singular Value Decomposition (SVD), ICA is an analysis method
based on the higher order statistical properties and in many applications analysis
of higher order statistical properties more conforms to reality. Recently, there
are many successful applications of ICA, including applications to biomedical
signal processing (Ikeda and Toyama, 2000; Wubbeler
et al., 2000), image processing (Hyvarinen and
Kosko, 2001; Luo and Boutell, 2005; Soltanizadeh
and Shokouhi, 2008; Wen and Miao, 2012) and financial
data analysis (Oja et al., 2000). ICA approach
has also successfully been used in electronic nose data (Di
Natale et al., 2002; Kermit and Tomic, 2003;
Balasubramanian et al., 2008; He
et al., 2008). However, they did not solve the problem that how to
discriminate the useful independent source and the noise. Especially in the
application of wound infection detection by ENose, ICA has not been used to
eliminate the strong background interference.
INDEPENDENT COMPONENT ANALYSIS
A key point in multivariate data analysis is to find suitable representations
of data. A suitable representation means a certain desirable feature of the
data becomes more accessible in the further analysis using a certain transform
of the original data.
Basic linear ICA model can be described as (Comon, 1994;
Hyvarinen and Oja, 2000):
where, s = [s_{1},s_{2}, .....,s_{N}]^{T} is
Ndimensional unknown source signa S_{i} (i = 1, ....., N) are mutually
independent; x = [x_{1},x_{2}, .....,x_{M}]^{T}
is the Mdimensional observed signal; A is an unknown MxN mixing matrix containing
the coefficients of the mixing system. The observed signal x is linear combinations
of the unknown source signal s. This statistical model is called a basic ICA
model (Hyvarinen et al., 2001). ICA will estimate
the matrix A and source signals s in the same time, only knowing the observed
signals x. It means we will find a linear transform or unmixing system B, which
satisfies:
where, y = [y_{1},y_{2}, .....,y_{N}]^{T} and y_{i} (i = 1, ....., N) are as independent as possible and B is an estimate of A^{1}. By doing this, approximations of the original source signals before the mixing are recovered. Because mixing matrix A and source signals are unknown, the solution process of ICA algorithm is not a process to find the inverse matrix of A and the separated signal y is only the approximation of the source signal. So, there are two uncertainties: (1) there is a certain proportional relationship between the separated signals and the source signals in amplitude and (2) We cannot determine the order of the separated signals. However, in many applications, most information of signals is represented in the signal waveform rather than the signal amplitude and the order. In addition, although we do not know much about the source signals, in some cases, we can discriminate them according to actual situation after separating the independent source signals. So, it is acceptable for the two uncertainties in ICA.
As mentioned before, ICA can be applied only if the source components are independent
of each other. Mathematically, statistical independence is defined by the joint
probability density. Provided that two random variables m_{1} and m_{2}
are independent if and only if their joint probability function density can
be decomposed following formula (Hyvarinen and Oja, 2000;
Hyvarinen et al., 2001):
where, p_{1}(m_{1}) and p_{2}(m_{2}) are the
marginal probability density functions of m_{1} and (m_{2}),
respectively. This definition can be extended to the case of n random variables,
in which the joint probability density is a product of marginal probability
densities of n random variables. Unlike PCA which only uses the secondorder
statistics of the signals and describes the data in the orthogonal constraint,
ICA finds out the independent hidden information, removes higher order redundant
information among the components and extracts the independent sources through
analyzing the higher order statistical correlation of multidimensional observed
data.
In electronic nose signal analysis, a received signal from a sensor is usually a result of a combination of different gases in varying proportions. It means that the response of a sensor depends on the joint influence of a mixture of gases. We can think that the measured signal of a sensor is usually a weighted superposition of several independent components. So, the ICA is an effective decomposition method to analyze the ENose signals. In addition, the studies of physiology of olfaction suggest that the preprocessing of cognitive and perceptive information for human has a feature of removing the redundant. And ICA also shows similar characteristics in this aspect because the mutual information between individual components is the least. Therefore, if using ICA instead of PCA to ENose systems should be able to obtain better treatment results. Here we introduced the wound infection detection ENose denoising algorithm based ICA. DENOISING ALGORITHM BASED ICA For wound infection detection ENose, the experiment's environment and mouse body odors produced a strong background interference of sensor signals and affected the system's recognition accuracy seriously. This paper proposed a reference vectorbased ICA denoising algorithm to eliminate the strong background interference in experiments.
ICA model with noise: The following formula can be considered as ICA
model with noise (Hyvarinen et al., 2001):
where, s = [s_{1},s_{2}, .....,s_{N}]^{T} is Ndimensional unknown source signal s_{i} = (i = 1, ....., N) are mutually independent; x = [x_{1},x_{2}, .....,x_{M}]^{T} is the Mdimensional observed signal; n = [n_{1},n_{2}, .....,n_{N}]^{T} is the Ndimensional additive noise; A is an unknown MxN mixing matrix containing the coefficients of the mixing system. The observed signal x is linear combinations of the unknown source signals. In the basic ICA model, we often assume that there are no interfering signals. But this is an ideal situation and in fact, interference signals are present and often cannot be ignored. ICA model with noise considers the noise as an independent component of source signals and one independent component is the noise in the matrix after the ICA transformation. We can separate the mutually independent target signal and noise by performing ICA of the observed signals. However, due to the uncertainty of the ICA, the order and magnitude of the independent components after the ICA transform are uncertain. So the desired signals and noise cannot be directly distinguished in the output the ICA and it is necessary to discriminate the transformed components. So, we proposed reference vectorbased ICA denoising algorithm. Reference vectorbased ICA denoising algorithm: First, we obtain independent components by performing ICA transform of the observed signals. Then we define a reference vector and distinguish the useful signals and noise by comparing the correlations between each independent component and the reference vector. We propose two methods to define reference vector. The first one is to take the average of each observed signals as the reference vector. Because the observed signals consist mainly of the useful signals rather than noise, so that the independent component which has smallest correlation with the reference vector is regarded as the noise. Therefore, we calculate correlation coefficients between each independent component and the reference vector, respectively and distinguish the useful signals and noise by comparing the correlation coefficients. Then, we remove the noise component and use the useful signals for pattern recognition.
For our ENose data, there are fifteendimensional observed signals x_{1},x_{2},
.....,x_{15} and we can obtain threedimensional independent components
y_{1},y_{2}, .....,y_{15} by ICA transform. The specific
algorithm is as follows: (1) Generate the reference vector x_{0} = [x_{1}+x_{2}+.....+x_{15}]/15;
(2) Calculate the correlation coefficients between y_{1},y_{2},
.....,y_{15} and the reference vector x_{0}, respectively and
obtain CORR (y_{1}, x_{0}),..., CORR (y_{2}, x_{0}),...,CORR
(y_{15}, x_{0}); (3) Compare the absolute values of the fifteen
correlation coefficients and the independent components with greater absolute
value of correlation coefficients with reference vector x_{0} are identified
as the responses of sensors on useful source signals, while the independent
component with smallest absolute value of correlation coefficient with reference
vector x_{0} is considered as the response on the background interference;
here, suppose that absolute value of CORR (y_{15}, x_{0}) is
the smallest and y_{15} is the response on the background interference;
(4) Remove y_{15} and obtained a fourteendimensional data y_{1},
y_{2},.....,y_{14} to used as inputs of classifier which are
the responses of sensors on useful source signals and so, we eliminate the effect
of background interference; (5) Finally, in order to make the response data
be in the same order of magnitude, we normalize the remaining independent components
y_{1}, y_{2},.....,y_{14} using the formula:
Here, y_{max} and y_{min} are the maximum and minimum components among y_{1}, y_{2};.....,y_{14}, y_{i}’ is the normalized data of y_{i} = (I = 1, ..., 14) and put into the neural network classifier. The second method to define reference vector is to directly take the measured background interference as a reference vector. Considering that the strong background interference during experiment mainly comes from the mouse body odors, we can take the sensor response of mice without wound as the reference vector. Separated independent component which is the most relevant to the reference vector is considered as the interference. The algorithm steps are similar to the method described previously and the step (4) is changed to remove the independent component which has the largest absolute value of correlation coefficient with reference vector. For our study, there are 80 samples and after extracting the maximum of sensor response as feature we obtain an 80*15 raw data matrix. Each line represents one sample and each column represents the output of one sensor. After ICA transform is carried out for the 80*15 raw data matrix, we obtain 80*15 independent component matrix. Each column represents one independent component. Remove one dimension background interference according to reference vectorbased denoising algorithm and receive 80*14 data matrix. Then we use cross validation method that divide 80 samples into k parts and k1 parts are used for training the RBF network classifier and the rest one part is used for test. Neural network classifier: After eliminating background interference by reference vectorbased ICA denoising algorithm, we put the sensor responses into RBF neural network classifier to distinguish the wound infection types. RBF neural network is a feedforward back propagation neural network, consists of three layers: input layer, hidden layer and output layer. The network topology structure is shown in Fig. 1.
The hidden layer consists of a set of radial basis functions as the network
activation function. The radial basis function is a Gaussiantype function.

Fig. 1: 
Topology structure of RBF network 
The hidden layer node calculates the Euclidean distance between the center
of radial basis functions and the network input vector and then uses the result
as the input of the radial basis functions. The output of radial basis functions
is:
where, p is the rdimensional input vector, C_{i} is the center vector of the i the hidden node, b is the bias of radial basis layer, σ is a real constant known as spread factor of radial basis function which expresses the kernel size, .^{2} is the 2norm of vectors. The outputs of the n nodes of output layer are linear combinations of the outputs of hidden layer nodes as: where, b_{j} is the bias of output layer, W2_{i,j} is the weight between the ith hidden node and the jth output node.
RBF method is an interpolation technique in highdimensional space, which constitutes
a hidden layer space using a radial basis function as the basis function. The
hidden layer transforms lowdimensional input data into highdimensional space,
making the linearly inseparable lowdimensional problem be linearly separable
in highdimensional space. For each training sample, it only corrects a small
number of weights and bias. The RBF network is quite suitable for implementing
multiclass and highdimensional classification problems with advantages of
convergence rate, probability of reaching global points, local sensitivity,
etc., (Qasem and Shamsuddin, 2010; Mahi
and Izabatene, 2011).
Gas sensor array: In this study, an array of fourteen metal oxide sensors
and one electrochemical sensor was used according to the metabolites of the
infected pathogens The sensors and there prime sensitive characteristics are
shown in Table 1.
Table 1: 
Pathogens in wound infection and their metabolites 

Table 2: 
Sensitive characteristics of gas sensors 

To enhance the ability of restraining environmental interference, a temperature
sensor (LM35DZ), a humidity sensor (HIH4000) and a pressure sensor (SMI5552)
are added into the sensor array; these sensors provide readings of the ambient
information. The response signals of the sensor array obtained from the odor
of the wound are first conditioned through a conditioning circuit and then sampled
and saved in a computer via a 14bit data acquisition card.
Sample preparation and measurement: There are five kinds of mice, wounded
and uninfected, infected with Pseudomonas aeruginosa, Escherichia
coli, Staphylococcus aureus, respectively, as well as no wounds and
infection (used as background). The metabolites of the three pathogens are shown
in Table 2. Each mouse has one wound in its hind leg. The
mice for experiment are provided by Animal Experiment Center of Third Military
Medical University. Each mouse is put into a big glass bottle with a rubber
stopper. Two holes are made in the rubber stopper with two thin glass tubes
inserted, respectively. One glass tube is fixed above the wound as close as
possible. The VOCs of the mouse wound flow out of the bottle through the glass
tube and clean air flow into the bottle through another glass tube. The dynamic
headspace method is adopted during all the experiments and the process is as
follows. The first stage is the baseline stage, in which the sensors are exposed
to clean air for three minutes. The second stage is the response stage, which
consists of a five minutes exposure of sensor to the headspace air of the wound
conveyed into the sensor chamber by a pump.
The third stage is the recovery stage; the sensors are exposed to clean air
again for fifteen minutes. At the end of each experiment, prior to the next
experiment, a ten minutes purging of the sensor chamber using highpurity nitrogen
is performed. The gas flow is controlled by a gas flow rate control system,
which contains a rotor flowmeter, a pressure retaining valve, a steady flow
valve and a needle valve. The flow rate is kept at 50 mL min^{1}. Eighty
experiments for all four kinds of mice in the same conditions are made and so
80 samples are collected. Figure 23 show
the QS01 sensor response on one Staphylococcus aureus infected mouse
and background, respectively. The body odor of mouse itself is so strong that
the useful wound information is buried in this background.

Fig. 3: 
Response of QS01 on background 
From Fig. 23, we can see that these two
signals are quite similar, the background is not a simple additive noise which
is added to the signal of wounds, it also contains signal of some source which
is different from the wounds odor.
RESULTS AND DISCUSSION
In order to evaluate the generalization performance, a standard statistical
generalization error estimation methodcross validation, largely used for the
purposes of classification, were used (Bishop, 1995).
All analysis of the data was implemented in MATLAB R2008b. Due to very fast
convergence speed, the fast fixedpoint iterative algorithm based on approximate
negentropy, which was called Fast ICA algorithm, was used to conduct the ICA
and the Fast ICA package was provided by Hyvarinen (1999).
The Fast ICA algorithm was run using the symmetric approach, where in all the
independent components were separated simultaneously. The nonlinearity function
g (u) = tanh (u) was used. Other nonlinearity function was also tested but
did not improve ICA performance significantly.
To demonstrate the validity of the proposed methods, the PCA, ICA and adaptive noise canceller with Normalized LeastMeanSquare (NLMS) algorithm were used as contrasts. So eight different methods are compared:
• 
The maximum response values of 15 sensors are used as features 
• 
Extracting 15 principal components for the maximum response values of
15 sensors by PCA and the 15 principal components are used as features 
• 
Extracting 3 principal components for the maximum response values of 15
sensors by PCA with cumulative variance contribution above 95% and the 3
principal components are used as features 
• 
Extracting 5 principal components for the maximum response values of 15
sensors by PCA with cumulative variance contribution above 99% and the 5
principal components are used as features 
• 
The maximum response values of 15 sensors which are eliminated background
using adaptive noise canceller with NLMS algorithm are used as features
the number of coefficients, step size and leakage factor in NLMS is 11,
0.02 and 0.6, respectively 
• 
Extracting 15 independent components for the maximum response values of
15 sensors by ICA and the 15 independent components are used as features 
• 
Eliminating background interference and dimension reduction for the maximum
response values of 15 sensors by reference vectorbased ICA denoising algorithm,
wherein reference vector is the average of each observed signals 
• 
Eliminating background interference and dimension reduction for the maximum
response values of 15 sensors by reference vectorbased ICA denoising algorithm,
wherein reference vector is the response of background interference 
Table 3: 
Comparison among four different methods 

The plot of the classification rate with the different k value in kfold cross
validation is shown in Fig. 4. It is obvious that the performance
changes with the varying k value. For the same k value, we also can obtain different
results because of dividing data into different k parts. For a particular division,
we can obtain a special classification rate. A leaveoneout (LOO) procedure
test, which can be viewed as an extreme form of kfold cross validation in which
k is equal to the number of examples, was performed to validate model robustness.
In our case, k is equal to 80. In LOO cross validation, all sample data of each
class are used for training except one, which is left for testing.

Fig. 4: 
Classification rate with different k value 
The classification results using LOO and the parameters values of RBF neural
network classifier of each method are given in Table 3.
From Table 3 it can be noticed that, decomposition of the raw data by PCA can remove the correlation between variables and improve the classifying performance. After dimension reduction, the classification rates reduced greatly, though the cumulative variance contribution is above 95 and 99%, respectively. It means that the principal components with high variance contribution do not necessarily benefit to classification. Eliminating background using adaptive noise canceller with NLMS algorithm can receive better classification result than the original response method and obtain 87.5% classification rate. It means that adaptive noise canceller can eliminate background interference to a certain extent. Extracting independent components of the raw data without dimension reduction operation by ICA has higher classification results than the former methods. It shows that, for wound infection detection ENose data, the ICA can decompose and DeNoise the original data better, extract the hidden information more beneficial to classification in the original data. However, only using ICA decomposition, we cannot discriminate independent useful source signals and background interference. So, we introduced two reference vectorbased ICA denoising algorithms and the two proposed methods can achieve 95 and 96.25% correct recognition rate, respectively, which are much better than the other denoising methods. Preprocessing the data of gas sensor array by ICA can realize redundancy eliminating and decorrelation of samples effectively and find independent useful source signals and noise. Using the two proposed methods, especially directly taking the measured background noise as a reference vector, we can discriminate the source signals and noise, eliminate the background interference and improve pattern recognition accuracy greatly. CONCLUSION In this study, a new method of detecting wound infection, based on an electronic nose (ENose) and reference vectorbased ICA denoising algorithm, is proposed. The two proposed reference vectorbased ICA denoising algorithms achieve 95 and 96.25% classification accuracy for mouse wound infection detection, respectively. They are better than that of PCA, PCA with cumulative variance contribution above 95%, PCA with cumulative variance contribution above 99%, adaptive noise canceller and the traditional ICA, which obtain 92.5, 62.5, 65, 87.5 and 93.75% classification accuracy, respectively. The results show that reference vectorbased ICA denoising algorithm, especially directly taking the measured background interference as a reference vector, can discriminate the source signals and the interference effectively, eliminate the background interference and improve pattern recognition accuracy greatly for wound infection detection based ENose. Traditional PCA and adaptive noise canceller methods are not effective to process the complex ENose signals of wound infection detection. In summary, this method is a useful tool for classification for the case of strong background interference and an increasingly accessible technology for real time, accurate and fast detection of wound infection. ACKNOWLEDGMENTS This study is supported by China Postdoctoral Science Foundation funded project ((Project No: 20090461445), the Natural Science Foundation Project of CQ CSTC, 2009, Peoples Republic of China (CSTC, 2009BA2021), the Clinical Research Fund of Third Military Medical University (2007XG077), Innovative Talent Training Project, the Third Stage of “211 Project”, Chongqing University (Project No: S09102) and the Fundamental Research Funds for the Central Universities (Project No. CDJXS10160001).
Fengchun Tian and Jia Yan developed the concept and all the work is under their
direction. Jia Yan, Jingwei Feng and Pengfei Jia developed and designed experiments,
accomplished hardware design. Lianqing Fu and Shan Xu accomplished software
design. Qinghua He and Yue Shen developed the concept, provided experiment samples
and medical guidance.

REFERENCES 
Abd ElAziz, S.E., 2011. Control strategies of stored product pests. J. Entomol., 8: 101122. CrossRef  Direct Link 
Anh, D.T.V., W. Olthuis and P. Bergveld, 2005. A hydrogen peroxide sensor for exhaled breath measurement. Sens. Actuators B: Chem., 111: 494499. Direct Link 
Balasubramanian, S., S. Panigrahi, C.M. Logue, C. Doetkott, M. Marchello and J.S. Sherwood, 2008. Independent component analysisprocessed electronic nose data for predicting Salmonella typhimurium populations in contaminated beef. Food Control, 19: 236246. CrossRef 
Bicego, M., G. Tessari, G. Tecchilli and M. Bettinelli, 2002. A comparative analysis of basic pattern recognition techniques for the development of small size electronic nose. Sens. Actuators B: Chem., 85: 137144. Direct Link 
Bishop, C.M., 1995. Neural Networks for Pattern Recognition. Oxford University Press, Oxford, UK., ISBN: 9780198538646.
Byun, H.G., K.C. Persaud and A.M. Pisanelli, 2010. Woundstate monitoring for burn patients using Enose/SPME system. ETRI J., 32: 440446. CrossRef  Direct Link 
Ciosek, P. and W. Wroblewski, 2006. The analysis of sensor array data with various pattern recognition techniques. Sens. Actuators B: Chem., 114: 8593. Direct Link 
Comon, P., 1994. Independent component analysis: A new concept? Signal Process., 36: 287314. CrossRef  Direct Link 
Di Natale, C., A. Macagnano, E. Martinelli, R. Paolesse and G. D`Arcangelo et al., 2003. Lung cancer identification by the analysis of breath by means of an array of nonselective gas sensors. Biosens. Bioelectron., 18: 12091218. Direct Link 
Di Natale, C., E. Martinelli and A. D'Amico, 2002. Counteraction of environmental disturbances of electronic nose data by independent component analysis. Sens. Actuators B, 82: 158165. CrossRef 
Dutta, R., E.L. Hines, J.W. Gardner and P. Boilot, 2002. Bacteria classification using Cyranose 320 electronic nose. Biomed. Eng., Vol. 1. 10.1186/1475925X14
Gardner, J.W., H.W. Shin and E.L. Hines, 2000. An electronic nose system to diagnose illness. Sensors Actuators B, 70: 1924. CrossRef  Direct Link 
Gibson, T.D., O. Prosser, J.N. Hulbert, R.W. Marshall, P. Corcoran et al., 1997. Detection and simultaneous identification of microorganisms from headspace samples using an electronic nose. Sens. Actuators B, 44: 413422. CrossRef  Direct Link 
He, X., S. Wei and R. Wang, 2008. Independent component analysis and neural network applied on electronic nose system. Proceedings of the International conference on Bioinformatics and Biomedical Engineering, May 1618, 2008, Shanghai, China, pp: 490493.
Hines, E.L., E. Llobet and J.W. Gardner, 1999. Electronic noses: A review of signal processing techniques. Proc. IEE Circuits Devices Syst., 146: 297310. Direct Link 
Hyvarinen, A. and B. Kosko, 2001. Image Denoising by Sparse Code Shrinkage. WileyIEEE Press, New York, ISBN: 9780470544976, pp: 554568.
Hyvarinen, A. and E. Oja, 2000. Independent component analysis: Algorithms and applications. Neural Networks, 13: 411430. CrossRef 
Hyvarinen, A., 1999. Fast and robust fixedpoint algorithms for independent component analysis. IEEE Trans. Neural Networks, 10: 626634. CrossRef  Direct Link 
Hyvarinen, A., J. Karhunen and E. Oja, 2001. Independent Component Analysis. Wiley, New York.
Ikeda, S. and K. Toyama, 2000. Independent component analysis for noisy dataMEG data analysis. Neural Networks, 13: 10631074. CrossRef 
Kermit, M. and O. Tomic, 2003. Independent component analysis applied on gas sensor array measurement data. IEEE Sens. J., 3: 218228. CrossRef 
Khan, M.A.U., M.K. Khan, M.A. Khan, M.T. Ibrahim, M.K. Ahmed and J.A. Baig, 2004. Principal component analysis of directional images for face recognition. Inform. Technol. J., 3: 290295. CrossRef  Direct Link 
Lin, Y.J., H.R. Guo, Y.H. Chang, M.T. Kao, H.H. Wang and R.I. Hong, 2001. Application of the electronic nose for uremia diagnosis. Sens. Actuators B: Chem., 76: 177180. Direct Link 
Luo, J. and M. Boutell, 2005. Natural scene classification using overcomplete ICA. Pattern Recognit., 38: 15071519. CrossRef 
Mahi, H. and H.F. Izabatene, 2011. Segmentation of satellite imagery using RBF neural network and genetic algorithm. Asian J. Applied Sci., 4: 186194. CrossRef  Direct Link 
Oja, E., K. Kiviluoto and S. Malaroiu, 2000. Independent component analysis for financial time series. Proceedings of the IEEE Adaptive Systems for Signal Processing, Communications and Control Symposium ASSPCC., October 14, 2000, Lake Louise, AB Canada, pp: 111116.
Pavlou, A.K., N. Magan, C. McNulty, J. Jones, D. Sharp, J. Brown and A.P. Turner, 2002. Use of an electronic nose system for diagnoses of urinary tract infections. Biosens. Bioelectron., 17: 893899. PubMed 
Persaud, K.C., A.M. Pisanelli, A. Bailey, K. Dunn and M. Falasconi et al., 2008. Identification of wound infection by limited set of volatile products. Proceedings of the IEEE International Conference Sensors, October 2629, 2008, Lecce Italy, pp: 13751377.
Phillips, M., R.N. Cataneo, A.R.C. Cummin, A.J. Gagliardi and K. Gleeson et al., 2003. Detection of lung cancer with volatile markers in the breath. Chest, 123: 21152123. PubMed 
Qasem, S.N. and S.M. Shamsuddin, 2010. Generalization improvement of radial basis function network based on multiobjective particle swarm optimization. J. Artif. Intell., 3: 116. CrossRef  Direct Link 
Ryabtsev, S.V., A.V. Shaposhnick, A.N. Lukin and E.P. Domashevskaya, 1999. Application of semiconductor gas sensors for medical diagnostics. Sens. Actuators B: Chem., 59: 2629. CrossRef 
Setkus, A., A.J. Galdikas, Z.A. Kancleris, A. Olekas and D. Senuliene et al., 2006. Featuring of bacterial contamination of wounds by dynamic response of SnO2 gas sensor array. Sens. Actuators B, 115: 412420. CrossRef  Direct Link 
Soltanizadeh, H. and S.B. Shokouhi, 2008. Feature extraction and classification of objects in the rosette pattern using component analysis and neural network. J. Applied Sci., 8: 40884096. CrossRef  Direct Link 
Thomas, A.N., S. Riazanskaia, W. Cheung, Y. Xu and R. Goodacre et al., 2010. Novel noninvasive identification of biomarkers by analytical profiling of chronic wounds using volatile organic compounds. Wound Rep. Reg., 18: 391400. CrossRef  PubMed 
Turner, A.P.F. and N. Magan, 2004. Electronic noses and disease diagnostics. Nat. Rev. Microbiol., 2: 161166. CrossRef 
Wen, X.L. and M.X. Miao, 2012. Research and simulation on blind separation algorithm in mixing image based on ICA. Inform. Technol. J., 11: 456460. CrossRef 
Wubbeler, G., A. Ziehe, B. Mackert, M. Muller, K.R.L. Trahms and C. Curio, 2000. Independent component analysis of noninvasively recorded cortical magnetic DCfields in humans. IEEE Trans. Biol. Med. Eng., 47: 564599. CrossRef  Direct Link 
Yu, J.B., H.G. Byun, M.S. So and J.S. Huh, 2005. Analysis of diabetic patient`s breath with conducting polymer sensor array. Sens. Actuators B: Chem., 108: 305308. CrossRef 



