Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
4Activity
0 of .
Results for:
No results containing your search query
P. 1
ijcta2011020120 Application of Bioelectrical Impedance Sensing Techniques for Dengue Infection with Non-linear Autoregressive model

ijcta2011020120 Application of Bioelectrical Impedance Sensing Techniques for Dengue Infection with Non-linear Autoregressive model

Ratings: (0)|Views: 47|Likes:
Published by editor2533

More info:

Categories:Types, Reviews, Book
Published by: editor2533 on Feb 10, 2011
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

03/02/2013

pdf

text

original

 
1
Application of Bioelectrical Impedance Sensing Techniques forDengue Infection with Non-linear Autoregressive model
H. Abdul Rahim
1
 
1
Department of Control andInstrumentationEngineering, Faculty of Electrical Engineering,Universiti TeknologiMalaysia,81310 UTM Skudai, Johor,Malaysia.herlina@fke.utm.my F. Ibrahim
2
 
2
Department of BiomedicalEngineering, Faculty of Engineering, Universityof Malaya,50603 Kuala Lumpur,Malaysia.M.N. Taib
3
 
3
Faculty of ElectricalEngineering, UniversitiTeknologi Mara, 40450Shah Alam, Selangor,Malaysia.
Abstract
This paper discussed novel system identification for bioelectrical impedance measurement  parameter for monitoring dengue infections byusing nonlinear AR (NAR) based on Artificial Neural Network (ANN). Bioelectrical impedancemeasurement indicate the volume of Hb of thesubjects and NAR model with regularized approach yields better accuracy by 80.60% for bioelectrical impedance sensing method. Inbuilding the model, three parameter wereconsidered; the final prediction error (FPE),
 Akaike’s Information Criteria (AIC), and Lipschitz 
number.
 Keywords:
dengue fever, NAR model, AIC, Lipschitz, FPE, ROC and AUC.
1. Introduction
Many infectious diseases such as dengue,malaria, typhoid and hepatitis producecharacteristic variations in the composition of blood [1] . These variations can be a characteristicchange in number, size or shape of certain bloodcells. For example, in anaemia, the red blood cell(RBC) count is reduced [1]. Other diseases maycause changes in the chemical composition of theblood serum and other body fluid, like thecharacteristically elevated in size and shape, or achemical analysis of the blood serum can,therefore, provide important information for thediagnosis of such diseases [1]. Similarly, otherbody fluids, smears, and small samples of livetissue, obtained by biopsy, are studied through thetechnique of bacteriology, serology and histologyto obtain clues for the diagnosis of diseases.However, these techniques are invasive becausethe bacteriology, serology and histology diagnosisrequires t
he sample of human‟s smear from the
throat, blood and tissue respectively. The latestcommercial technique takes two hours to detectdengue fever by serological confirmation usingsamples of serum, plasma or heparinized wholeblood [1]. This test is still invasive and expensiveand can only be performed by trained medicalpersonnel.Dengue fever (DF) ranks highly among the newlyemerging infectious diseases in public healthsignificance. Hence it is considered to be the mostimportant of the arthropod-borne viral diseases. InMalaysia, the disease is endemic but majoroutbreaks seem to occur at least once in every fouryears [1]. Dengue fever was first reported inMalaysia after an epidemic in Penang in 1902 [2,3]. Since the early 1970s, the World HealthOrganization (WHO) has been actively involved indeveloping and promoting strategies for treatmentand control of dengue. In 1997, WHO published asecond guide to the diagnosis, treatment andcontrol of dengue haemorrhagic fever [4]. Denguewere reported throughout the year and started toincrease from 1997 to 1998. In 1998, 27,373dengue cases with 58 deaths were reported ascompared to 19,544 cases with 50 deaths in 1997.This has shown an increase of 7,829 cases or40.1% over the number of cases in 1997 [5].
H.Adbul Rahim,F.Ibrahim,M.N.Taib, Int. J. Comp. Tech. Appl., Vol 2 (1), 207-215207
ISSN: 2229-6093
 
2
Therefore, accurate classification of dengueinfection a very useful tool for doctors indiagnosing diseases early.Fatimah et. al. [6] describe a noninvasiveprediction system for predicting the day of defervescence of fever in dengue patients usingANN. The developed system bases its predictionsolely on clinical symptoms and signs and theresults show that around 90% prediction accuracy.This paper describes a noninvasiveclassification system for dengue infections usingNAR models. The rest of the paper is structuredas follows: Section 2 gives a brief review of system identification. In Section 3 nonlinearautoregressive (NAR) model is presented.Methods are provided in Section 4. In Section 5shows the results. Finally, concluding remarksand discussions are presented in Section 6.
2. System identification
This section describes the general methodologyconducted for classifying the dengue infectionsdisease as shown in a flowchart in Figure 1 whichconsists of five stages.The first stage began with the data collectionon the dengue disease. Then a suitable structurewill be selected to match the dynamics of the data.The Levenberg-Marquardt algorithm is used totrain the ANN. At the fourth stage, ROC analysiswas applied to illustrate the sensitivity, thespecificity and the AUC percentage of theappropriate model.It is a common practice in various scientificand engineering disciplines to represent observeddiscrete time random processes by autoregressive(AR) models.A fundamental problem in systemidentification is the choice of the nature of themodel which should be used for the system understudy. Some of the problems in systemidentification are:i)
 
determining the order of the modelii)
 
selection of a suitable criterion fordetermining the accuracy of the modeliii)
 
designing an input signal which willmaximize the accuracy of the estimates of the parameter of the model.Usually, the parameter estimation is carried outby least squares or maximum likehood procedures[7]. An important part of statistical inference dealswith model order selection. Several criteria havebeen proposed as subjective bases for selecting of the system identification model order. Atechnique based on the prediction error variance,the Final prediction error (FPE) was developed byAkaike. Akaike also proposed another well-
known criterion, Akaike‟s Information Criterion
(AIC), that is derived from information theoreticconcepts. To determine the model order, theapproach, ranges from the classical based on theestimated residuals of the fitting model [8] to thosefounded on information [9] and coding theory [10]of Bayesian analysis [11, 12]. The determinationof the order and estimation of the parameters of the models is the primary concern here. Theproblem of determination of suitable structure andorder of the system from its input/output data hasseveral workable solutions, especially linearsystem.The choice of linear system and modelstructures are the most critical task in anyidentification methodology. An alternative to themodeling techniques of linear systems is theimplementation of computational intelligent,which is mostly based on Artificial NeuralNetworks (ANN). The vast development of ANNin the last few years, led to the use of ANN insystem identification. Narenda and Parthasarathy[13] proposed that ANN should be used inconjunction with system theory in thedevelopment of realizable models.
3. Nonlinear Autoregressive Model
The NAR model consists of an autoregressivewhich is represented as past output data and thenonlinear function was selected as hyperbolictangent.
)](),(),...,1([)(
ˆ
un y y y
 y
(1)where
is a nonlinear part,
 y(t)
and
u(t)
representthe output and input, respectively.
n
 y
is theassociated maximum lags. Block diagram forNAR model is as shown in Figure 2.Figure 3 shows how a feedforward ANN can beconfigured as a NAR model. The function
inEquation 1 is realized by a feedforward network,the predictor will have a feedback when theregressors are selected as in an AR model. Theequation for the predictor has the form:
 y
un y y
)](),()...1([)(
 
(2)
H.Adbul Rahim,F.Ibrahim,M.N.Taib, Int. J. Comp. Tech. Appl., Vol 2 (1), 207-215208
 
3
3.1 Model Order Selection
Model order selection is dependent upon thequality of the model since the model order isvaried and the cost function is monitored. Auseful measure to aid this procedure is to measurethe significance of each additional model.Assessing the significance of each model is notonly necessary for model order selection but alsofor further analysis of the estimated model and canaid the design and analysis of medicalapplications. In this research we used Lipschitznumber.He and Asada [14] introduced Lipschitznumber method for the estimation of the modelorder of nonlinear input-output models. Theapproach is based on the continuity property of nonlinear functions, which represents input-outputmodels of continuous dynamic systems. Byevaluating the modification of an index, which isdefined as Lipschitz number with successivemodification of the model orders, the appropriatemodel orders can be determined more simply andreliably.
3.2 Model Estimation
The third step is model estimation, whichinvolves determining the numerical values of thestructural parameters, which minimise the errorbetween the system to be identified, and its model.In this research we used LM algorithm.The MLP may be taught to perform a specificfunction through a process called training. Thetraining process is performed using a set of guidedweight update rules. In this thesis, the LMalgorithm was chosen as the training algorithm.This section presents the LM algorithm in detail.
3.3 Regularization
If the network has been trained to a very smallvalue of criterion, the model needs not beparticularly good. A good performance on thetraining set does not automatically imply that themodel generalizes well to new inputs. Inparticular, it was shown that if the model structurewas too large (contained many weights) it led tooverfitting [15], that is, the noise in the training setwas also modelled. The average generalizationerror was introduced as a quantity assessing agiven model structure. One way of controlling theaverage generalization error was to extend thecriterion with a term called regularization bysimple weight decay [15]. The weight decayreduced the variance error at the expense of ahigher bias error.
3.4 Model Validation
Receiver operating characteristic (ROC) curvesare commonly used in medicine and healthcare[16], where they are used to quantify the accuracyof diagnostic tests [17, 18]. The performance of 
an “expert” human or machine, can be represented
objectively by ROC curves [19]. Such curvesshow, for example, the trade-off between adiagnostic test correctly identifying diseasedpatients as diseased, rather than healthy, versuscorrectly identifying healthy patients as healthy,rather than diseased. Terms commonly used inROC curves are sensitivity, specificity anddiagnostic accuracy, to show the accuracy of thedesigned system.ROC curves display the relationship betweensensitivity (true positive rate) and 1-specificity(false positive rate) across all possible thresholdvalues that define the positivity of a disease. Theyshow the full picture trade-off between truepositive rate and false positive rate at differentlevels of positivity.The ANN must be trained before the ROCcurve can be generated. The resulting network is
referred to as a “basic trained network”. This
initial instance of the ANN provides one operatingpoint. The result is a set of instances of thenetwork chosen to represent a point on the ROCcurve. The goodness of this set of network instances are then evaluated using separate testdata.Table 1 shows a diagnostic accuracy resultsafter training the ANN. The decision variable canproduce two sets of values, which represents twocategory types dengue infection. The true dengueinfection is denoted as D+, whereas the falsedengue infection is indicated as D-.In general, four possible decisions and twotypes of errors are made when comparing a testresult with a diagnosis, as shown in Table 1. If both diagnosis and test are positive, it is called atrue positive (TP). The probability of the TP tooccur is estimated by counting the true positives inthe sample and dividing by the sample size. If thediagnosis is positive and the test is negative it iscalled a false negative (FN). False positive (FP)and true negative (TN) are defined similarly. Thetwo sets of values produced in the threshold arethe total positive and negative indicated as T+ andT-.
H.Adbul Rahim,F.Ibrahim,M.N.Taib, Int. J. Comp. Tech. Appl., Vol 2 (1), 207-215209

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->