A New Intelligence Based Approach for Computer-Aided Diagnosis of Dengue Fever

0.75 MB·English

by  Vadrevu Sree Hari Rao

(cid:13)c 20XXIEEE.PERSONALUSEOFTHISMATERIALISPERMITTED.PUBLISHEDINIEEE.DOI:10.1109/TITB.2011.2171978 1
A New Intelligence Based Approach for
Computer-Aided Diagnosis of Dengue Fever
Vadrevu Sree Hari Rao, Senior Member, IEEE, and Mallenahalli Naresh Kumar
Abstract—Identification of the influential clinical symptoms as enzyme-linked immunosorbent assays (ELISA) and real-
and laboratory features that help in the diagnosis of dengue time polymerase-chain reaction (RT-PCR) which are based
feverinearlyphaseoftheillnesswouldaidindesigningeffective
on nucleic and acid hybridization [3]. A recent study [4] on
publichealthmanagementandvirologicalsurveillancestrategies.
the behavior of C-type lectin domain family 5, member A
Keepingthisasourmainobjectivewedevelopinthispaper,anew
5 computational intelligence based methodology that predicts the (CLEC5A) gene may result in a strategy for reducing tissue
1
diagnosis in real time, minimizing the number of false positives damage which would help improve the odds of survival of
0
and false negatives. Our methodology consists of three major the patients suffering from DHF and dengue shock syndrome
2
components (i) a novel missing value imputation procedure that
(DSS). A multivariate model was developed in [5] for pre-
n canbeappliedonanydatasetconsistingofcategorical(nominal)
dicting hemoglobin (Hb) using predictors such as reactance
a and/or numeric (real or integer) (ii) a wrapper based features
J selection method with genetic search for extracting a subset of obtained from a single frequency bioelectrical impedance
1 most influential symptoms that can diagnose the illness and (iii) analysis, sex, nausea/vomiting sensation and weight. These
3 an alternating decision tree method that employs boosting for strategies can be employed only after 2 − 12 days from
generating highly accurate decision rules. The predictive models the onset of the illness and require state-of-the-art laboratory
] developedusingourmethodologyarefoundtobemoreaccurate facilities.
L than the state-of-the-art methodologies used in the diagnosis of
M the dengue fever. The World Health Organization (WHO) has arrived at a
classification scheme for identifying the infected individuals
Index Terms—dengue fever, classification, clinical diagnosis,
.
t prediction, imputation, features selection, genetic search, alter- based on clinical symptoms and laboratory features. The de-
a
nating decision trees velopment of predictive models for diagnosis of dengue fever
t
s
based on these schemes is affected by missing or incomplete
[
data records in the clinical databases [6] which may arise due
1 I. INTRODUCTION to any or all of the following reasons (i) value being lost
v DENGUE fever (DF) is a mosquito-borne infectious dis- (erased or deleted) (ii) not recorded (iii) incorrect measure-
2
ease caused by the viruses of the genus Togaviridae, ments(iv)equipmenterrorsand(v)anexpertnotattachingany
6
0 subgenusFlavirus.Thetransmissionofthisdiseaseisthrough importance to a particular clinical procedure. Usually data is
0 the bites of vectors (aedes aegypti, aedes albopictus) carrying notcollectedfromanorganizedresearchpointofview[7].The
0 thevirusesbelongingtoFlavigenus[1].Fromitsfirstappear- presenceoflargenumberofclinicalsymptomsandlaboratory
2. anceinthePhilippinesin1953,thediseasehasbeenidentified features requires one to search large sub spaces for optimal
0 as one of the most important arthropod-borne viral disease feature subsets. These issues unless addressed appropriately
5 in humans [2]. Dengue virus infection has been reported in wouldhinderthedevelopmentofaccurateandcomputationally
1 morethan100countries,with2.5billionpeoplelivinginareas effective diagnostic system.
:
v where dengue is endemic. The annual occurrence is estimated
In view of the above challenges, we present the following
i to be around 100 million cases of DF and 250,000 cases of
X novel features of our work:
dengue hemorrhagic fever (DHF).
r
a The diagnosis of dengue fever presents great challenges as • to identify the missing values (MV) in the data set
the symptoms overlap with other febrile illnesses. Accurate and impute them by using a newly developed novel
diagnosisispossibleonlyafterconductingdefinitivetestssuch imputation procedure;
• to identify a set of clinical symptoms that would enable
ManuscriptreceivedMay13,2011:revisedAugust24,2011andSeptember early detection of suspected dengue in children and
30,2011;acceptedOctober7,2011.
adults, which reduces the risk of transmission of the
Vadrevu Sree Hari Rao is with the Department of Mathematics, Jawahar-
lal Nehru Technological University, Hyderabad, Andhra Pradesh, 500 085, dengue fever in the community;
India. Also, he is an advisor for International Centre for Interdisciplinary • to identify the laboratory features and clinical symptoms
Research and Innovation, VNRVJIET Campus, Hyderabad, India. e-mail:
that would enable better diagnosis and understanding of
[email protected]
Mallenahalli Naresh Kumar is with the Software and Database Systems the disease in suspected dengue individuals. This renders
Group,NationalRemoteSensingCenter(ISRO),Hyderabad,AndhraPradesh, optimal utilization of the laboratory resources required
500625,India.e-mail:nareshkumar [email protected]
for confirmed diagnosis;
(cid:13)c 20xxIEEE.Personaluseofthismaterialispermitted.Permissionfrom
IEEE must be obtained for all other uses, in any current or future media, • tobuildapredictivemodelthathasacapabilityofrender-
includingreprinting/republishingthismaterialforadvertisingorpromotional ing effective diagnosis in realtime. Further we compare
purposes,creatingnewcollectiveworks,forresaleorredistributiontoservers
its performance with other state-of-the-art methods used
orlists,orreuseofanycopyrightedcomponentofthisworkinotherworks.
DOI:10.1109/TITB.2011.2171978 in the diagnosis of dengue fever.
(cid:13)c 20xxIEEE.Personaluseofthismaterialispermitted.PermissionfromIEEEmustbeobtainedforallotheruses,inanycurrentorfuturemedia,including
reprinting/republishingthismaterialforadvertisingorpromotionalpurposes,creatingnewcollectiveworks,forresaleorredistributiontoserversorlists,orreuseofany
copyrightedcomponentofthisworkinotherworks.DOI:10.1109/TITB.2011.2171978
The present paper is organized as follows: A survey of the III. ANEWMETHODOLOGYFORCOMPUTER-AIDED
state-of-the-art techniquesfor the diagnosisof dengue feveris DIAGNOSISOFDENGUEFEVER
presented in Section II, while in Section III we describe our Motivated by the above issues we propose a new method-
novel methodology for computer-aided clinical diagnosis of ology comprising of a novel non parametric missing value
dengue. The performance evaluation of the methodologies is imputation method that can be applied on data sets consisting
described in Section IV. The description of the data sets and of attributes that are of the type categorical (nominal) and/or
theexperimentalresultsarepresentedinSectionV.Wepresent numeric (integer or real). The methodology proposed in [15]
a comparison of our new imputation methodology with other ignores missing values while generating the decision tree,
imputation methods in Section VI. In Section VII we discuss whichrenderslowerpredictionaccuracies.Wehaveembedded
thecomputationalcomplexityofournewmethod.Comparison the new imputation strategy (Section III-B) before generating
of our new methodology with other state-of-the-art methods the alternating decision tree which results in the improved
forms the subject of Section VIII. Conclusions and discussion performance of the classifier on data sets having missing
are deferred to Section IX. values. Also, we develop an effective wrapper based features
selection algorithm in order to identify the most influential
features subset. The present methodology comprises in uti-
lizing the new imputation embedded alternating decision tree
II. SURVEYOFTHESTATE-OF-THE-ARTTECHNIQUESFOR
and the wrapper based features subset selection algorithm.
DIAGNOSISOFDENGUEFEVER
This methodology can predict the diagnosis of dengue in real
time.Infactthemachineknowledgeacquiredbyutilizingthis
Logistic regression method was employed to identify clini-
novelmethodologywillbeusefultodiagnoseotherindividuals
cal symptoms and laboratory features in 381 individuals, out
based on clinical symptoms and laboratory features where
of which 148 were confirmed dengue [8]. The data records
the clinical decision is unavailable. We designate this novel
with missing values (MV) are ignored and are deleted from
methodology as NM throughout this work.
thedataset.In[9],thestudywasconductedonclinicalrecords
comprisingof341childrenand597adultsoutofwhich38and
A. Data representation
107 respectively were laboratory-confirmed positive dengue
A clinical data set can be represented as a set S
cases. In this study the data fields that are incomplete or
having row vectors (R ,R ,...,R ) and column vec-
inaccurate for all suspected dengue cases were replaced with 1 2 m
tors (C ,C ,...,C ). Each record can be represented as
the known values corresponding to the information in the 1 2 n
an ordered n-tuple of clinical and laboratory attributes
medical charts. A C4.5 decision tree which has an in built
(A ,A ,...,A ,A ) for each i = 1,2,...,m where
mechanismofhandlingMVwasemployedin[10]todevelopa i1 i2 i(n−1) in
the last attribute (A ) for each i, represents the physician’s
diagnostic algorithm to differentiate dengue from non-dengue in
illnessonadatasetcomprisingof1200patientsofwhich173 diagnosistowhichtherecord(Ai1,Ai2,...,Ai(n−1))belongs
had DF, 171 had DHF and 20 had DSS. A support vector and without loss of generality we assume that there are no
missing elements in this set. Each attribute of an element in
machine (SVM) based methodology was employed in [11]
to analyze the expression pattern of 12 genes of 28 dengue S that is Aij for i = 1,2,...,m and j = 1,2,...,n−1 can
patients of which 13 were DHF and 15 were DF cases. A either be a categorical (nominal) or numeric (real or integer)
type. Clearly all the sets considered are finite sets.
setofseveninfluentialgeneswereidentifiedthroughselective
removal of expression data of these twelve genes.
B. A new non-parametric imputation strategy
In the above studies the MV were either removed [8], or
filled with approximate values based on medical charts [9]. Thefirststepinanyimputationalgorithmistocomputethe
These approaches would lead to biased estimates and may proximity measure in the feature space between the clinical
eitherreduceorexaggeratethestatisticalpower.Methodssuch recordstoidentifythenearestneighborsfromwherethevalues
as logistic regression, maximum likelihood and expectation can be imputed. The most popular metric for quantifying the
maximization have been employed for imputation of MV, but similarity between any two records is the Euclidean distance.
they can be applied only on data sets that are either nominal Eventhoughthismetricissimplertocompute,itissensitiveto
or numeric. There are other imputation methods such as k- thescalesofthefeaturesinvolved.Furtheritdoesnotaccount
nearest neighbor imputation (KNNI) [12]; k-means clustering for correlation between the features. Also, the categorical
imputation (KMI) [13]; weighted k-nearest neighbor imputa- variables can only be quantified by counting measures which
tion (WKNNI) [14] and fuzzy k-means clustering imputation callsforthedevelopmentofeffectivestrategiesforcomputing
(FKMI) [13] that have been applied on other data sets but not the similarity [16]. Considering these factors we first propose
ondenguefeverdatasets.However,theauthorsin[8],[9],[11] a new indexing measure ICl(Ri,Rk) between two typical
haveemployedmethodssuchasoddsratio(OR)andselective elements Ri, Rk for i,k = 1,2,...,m, l = 1,2,...,n−1
inclusion or exclusion of attributes for obtaining features sub belonging to the column Cl of S which can be applied on any
sets of data sets of dengue fever. But these methods do not type of data, be it categorical (nominal) and/or numeric (real
yeild effective diagnosis as all interactions or correlations or integer). We consider the following cases:
between the features and the diagnosis are not considered in Case I: A =A
in kn
these studies. Let A denote the collection of all members of S that
(cid:13)c 20xxIEEE.Personaluseofthismaterialispermitted.PermissionfromIEEEmustbeobtainedforallotheruses,inanycurrentorfuturemedia,including
reprinting/republishingthismaterialforadvertisingorpromotionalpurposes,creatingnewcollectiveworks,forresaleorredistributiontoserversorlists,orreuseofany
copyrightedcomponentofthisworkinotherworks.DOI:10.1109/TITB.2011.2171978
belong to the same decision class to which R and R indicate the cardinalities of the respective subsets. We
i k
belong and does not have MV. Based on the type of the define the indexing measure between the two records
attribute to which the column C belongs, the following R and R as
l i k
situations arise:
(cid:26) max{βr, δs}, for i(cid:54)=k;
ICl(Ri,Rk)= 0, δs βr otherwise.
(i) Elements of the column C of S are of categorical
l
(nominal) type: where β represents the cardinality of the subset P
r βr
We now express A as a disjoint union of non- all of whose elements have first co-ordinates A in
il
empty subsets of A, say Bγp1,Bγp2,...,Bγps ob- set P and δs represents the cardinality of that subset
tained in such a manner that every element of A Q , all of whose elements have first co-ordinates A
δs kl
belongs to one of these subsets and no element in set Q.
of A is a member of more than one subset of A.
(cid:83) (cid:83) (cid:83)
That is A = Bγp1 Bγp2 ,..., Bγps, in which (ii) Elements of the column Cl of S are of numeric type:
γp1,γp2,...,γps denotethecardinalitiesoftherespec- If the type of the attribute is integer we follow the
tive subsets Bγp1,Bγp2,...,Bγps formed out of the procedure discussed in Case II item (i). For fractional
set A, with the property that each member of the same numbers we define the index I (R ,R ) between the
Cl i k
subset has the same first co-ordinate and members of two records R andR as
i k
notwodifferentsubsetshavethesamefirstco-ordinate.
We define an index I (R ,R )=(cid:26) max{AΛil,AΛkl}, for i(cid:54)=k;
(cid:40) min{γpi,γqk}, for i(cid:54)=k; Cl i k 0, otherwise.
ICl(Ri,Rk)= 0, γqk γpi otherwise. IntheabovedefinitionΛ=min{P#,Q#}whereP#,
and Q# denote the average of the first column entries
where γpi represents the cardinality of the subset ofalltheelementsofthesets PandQexcludingthose
Bγpi, all of whose elements have first co-ordinates with MV in the lth column.
A and γ represents the cardinality of that subset
il qk The proximity or distance scores between the clinical
B , all of whose elements have first co-ordinates
γqk records in the data set S can be represented as D =
A .
kl {{0,d ,...,d };{d ,0,...,d };...;{d ,d ,...,0}}
12 1m 21 2m m1 m2
(cid:113)
where d = (cid:80)n−1I2 (R ,R ). For each of the missing
(ii) Elements of the column Cl of S are of numeric type: ik l=1 Cl i k
value instances in a record R our imputation procedure
Numeric types can be classified further as integers i
(awttrhioblueteniusmofbeinrste)goerrtryepaelt(hferancwtioenfaolllonwumthbeerpsr)o.cIefduthree first computes the score z(dij) = (cid:113)m1−1(d(cid:80)ijmi−=1d()dij−d) where
j = 1,2...,m and d denotes the mean distance. We then
discussedinCaseIitem(i).Forfractionalnumberswe
pick up only those records (nearest neighbors) which satisfy
construct the index I (R ,R ), based on the ratio of
Cl i k the condition z(d ) ≤ 0 where {d ,d ,...,d } denote
the values of the elements A ,A of lth column to ij i1 i2 im
il kl the distances of the current record R to all other records in
the mean of the set of elements belonging to A that do i
the data set S. If the type of attribute is categorical or integer,
not have MV and is given by
then the data value that has the highest frequency (mode)
(cid:26) min{Ail,Akl}, for i(cid:54)=k; of occurrence in the corresponding columns of the nearest
I (R ,R )= A# A#
Cl i k 0, otherwise. recordsisimputed.Forthedatavaluesoftyperealweimpute
the mean of data values in the corresponding columns of the
In the above definition A# denotes the average of the
nearest records.
lth column entries of all the elements of the set A
Illustrativeexample:Thefollowingexampleillustratesthe
excluding those with MV in the lth column.
spirit of the new imputation algorithm. Consider a data set
Case II: Ain (cid:54)=Akn represented by the matrix S consisting of rows R1=(?, 12.0,
Clearly Ri and Rk belong to two different decision positive), R2=( yes, 10.5, positive), R3=( no, 14.0, positive)
classes. Consider the subsets Pi and Qk consisting of and R4=(no, 13.0, negative). The missing value instance (’?’)
members of S that share the same decision with Ri in this data set is present in record R1 and column C1. These
and Rk respectively and does not have MV. Clearly rowscorrespondtothedatarecordsoffourindividuals.Clearly
(cid:84)
Pi Qk =∅. Basedon the typeof theattribute to which the Case I item (i) of the imputation algorithm applies to
the column Cl belongs, the following situations arise: this data set for determining the missing value. The matrix of
the indexing measure I has the following rows: (0,0.86) and
(i) Elements of the column C of S are of nominal or (0,0.99) in which γ =0, γ =1 and A# =12.17. The rela-
l p q
categorical type: tivedistancesbetweenR andtheotherrecordsarecomputed
1
Following the procedure discussed in Case I item (i) as {0.93,0,0} and the corresponding z-scores are obtained as
we write P and Q as a disjoint union of non-empty {−0.57,−0.57,1.154}.Sincez≤0forthedistancesbetween
subsets of P ,P ,...,P and Q ,Q ,...,Q R and R and also R and R , we conclude that the records
β1 β2 βr δ1 δ2 δs 1 2 1 3
respectively in which β ,β ,...,β and δ ,δ ,...,δ R and R are nearer to R and hence the highest frequency
1 2 r 1 2 s 2 3 1
(cid:13)c 20xxIEEE.Personaluseofthismaterialispermitted.PermissionfromIEEEmustbeobtainedforallotheruses,inanycurrentorfuturemedia,including
reprinting/republishingthismaterialforadvertisingorpromotionalpurposes,creatingnewcollectiveworks,forresaleorredistributiontoserversorlists,orreuseofany
copyrightedcomponentofthisworkinotherworks.DOI:10.1109/TITB.2011.2171978
(mode) of the data value in column C is ’yes’. Accordingly Algorithm 1 The NM Methodology
1
this value is a suitable candidate for imputation. Input:(a) DatasetsforthepurposeofdecisionmakingS(m,n)wheremandnare
numberofrecordsandattributesrespectivelyandthemembersofSmayhaveMV
inanyoftheattributesexceptinthedecisionattribute,whichisthelastattribute
intherecord.
C. Identification of influential features (b) ThetypeofattributeCofthecolumnsinthedataset.
Output:(a) ClassificationaccuracyforagivendatasetS.
In situations presented by real world processes, influential
(b) PerformancemetricsAUC,SE,SP.
features are often unknown a priori, hence features that are Algorithm
redundant or those that are weakly participating in decision (1) IdentifyandcollectallrecordsinadatasetS
(2) ImputetheMVinthedatasetSusingtheprocedurediscussedinSectionIII-B.
making must be identified and appropriately handled. The (3) Extracttheinfluentialfeaturesusingawrapperbasedapproachwithgeneticsearch
features selection procedures can be categorized as random foridentifyingfeaturessubsetsandalternatingdecisiontreeforitsevaluationas
discussedinSectionIII-C.
or sequential. The sequential methods such as forward selec- (4) Split the dataset in to training and testing sets using a stratified k fold cross
tion, backward elimination and bidirectional selection employ validation procedure. Denote each training and testing data set by Tk and Rk
respectively.
greedy methods and hence may not often be successful in (5) Foreachkcomputethefollowing
findingtheoptimalfeaturessubsets.Incontrasttothisstochas- (i) BuildtheADTusingtherecordsobtainedfromTk.
(ii) Computethepredictedprobabilities(scores)forbothpositiveandnegative
tic optimization methods such as genetic algorithms (GAs)
diagnosisofdenguefromtheADTbuiltinStep(5)-(i)usingthetestdataset
performglobalsearchandarecapableofeffectivelyexploring Rk.DesignatethesetconsistingofallthesescoresbyP.
large search spaces [17]. In our approach we adopt a wrapper (iii) IdentifyandcollecttheactualdiagnosisfromthetestdatasetRk intoset
denotedbyL.
subset based feature evaluation model [18] where the method (6) RepeattheSteps(5)-(i)toStep(5)-(iii)foreachfold.
ofclassificationitselfisusedtomeasuretheimportanceofthe (7) ObtaintheperformancemetricsAUC,SEandSPutilizingthesetsLandP.
(8) RETURNAUC,SE,SP.
features sub set identified by the GA. (9) END.
D. Predictive modeling using decision trees
probabilities of all the k folds. For each threshold value the
An alternating decision tree (ADT) consists of decision
measuresSEandSParecomputed.Thefalsepositiverateand
nodes(splitternode)andpredictionnodeswhichcaneitherbe
truepositiveratevaluesoftheROCistakenas(1-SP)andSE
aninteriornodeoraleafnode.Thetreegeneratesaprediction
respectively. The AUC is computed by applying a trapezoidal
node at the root and then alternates between decision nodes
rule on the data points of the ROC curve. The optimal cut
and further prediction nodes. Decision nodes specify a pred-
off or operating point is the threshold that is closest point to
icate condition and prediction nodes contain a single number
(0,1) on the ROC curve which gives the equal error rate. The
denotingthepredictivevalue.Aninstancecanbeclassifiedby
optimal values of AUC, SE, SP are computed for this cut off
following all paths for which all decision nodes are true and
point.
summing the relevant prediction nodes that are traversed. A
positivesumimpliesmembershipofoneclassandthenegative
V. EXPERIMENTSANDRESULTS
sum indicates the membership of the opposite class.
In our methodology we have employed a stratified ten-fold
cross validation ( k = 10) procedure. We applied a standard
IV. PERFORMANCEEVALUATIONMETHODS
implementationofSVMwithradialbasisfunctionkernel[11]
The standard definitions of the performance measures such
using LibSVM package [19]. The GA algorithm for features
as the specificity (SP), sensitivity (SE), receiver operator
selection has been performed using the parameter values:
characteristics (ROC) and area under ROC (AUC) based on
cross over probability=1.0 and mutation probability=0.001.
number of true positives, true negatives, false positives and
The standard implementation of C4.5, LOR algorithms in
false negatives are utilized in our experimental analysis. We Weka(cid:13)c [20] are considered for evaluating the performance
employed a stratified k-fold cross validation for estimating
of our algorithm. We have implemented the NM algorithm
the test error on classification algorithms. We have randomly and the performance evaluation methods in Matlab(cid:13)c. A non-
divided the given data set into k disjoint subsets. Each subset
parametric statistical test proposed by Wilcoxon [21] is used
is roughly of equal size and has the same class proportions
to compare the performances of the algorithms. We compared
as in the original data set. The classification model has been
the NM with the state-of-the-art methodologies employed
built by setting aside one of the subsets as test data set and
in diagnosis of dengue fever using different performance
train the classifier using the other nine subsets. The trained
measures discussed in Section IV.
model is then employed in classifying the test data set. The
experiment is repeated by setting aside each of the k subsets
A. Data sets
astestdatasetsoneatatime.TocomputeROCfork foldswe
firsttrainaclassifierusingthetrainingdatasetofak foldand We have obtained four surveillance data sets from case-
then obtain the scores in terms of the predicted probabilities patients admitted into hospitals located in central and western
forpositivesandnegativesfromthetrainedclassifierusingthe States of India. Standard procedures were adopted in collect-
testdatasetcorrespondingtothesamefoldasthetrainingdata. ing the clinical and demographic attributes of the patients.
Once all the probabilities and corresponding actual decisions The probable cases of the dengue fever are arrived through
are collected, the ROC is obtained by first computing the definitivelaboratorytestssuchasELISA.Thepatientsrecords
thresholds using the quartiles of the cumulative predictive include clinical symptoms: fever, fever duration, headache,
(cid:13)c 20xxIEEE.Personaluseofthismaterialispermitted.PermissionfromIEEEmustbeobtainedforallotheruses,inanycurrentorfuturemedia,including
reprinting/republishingthismaterialforadvertisingorpromotionalpurposes,creatingnewcollectiveworks,forresaleorredistributiontoserversorlists,orreuseofany
copyrightedcomponentofthisworkinotherworks.DOI:10.1109/TITB.2011.2171978
TABLE I Performance comparison of the NM with other TABLE III Influential features subsets identified by NM
methodologies (C4.5, SVM and LOR) on the data sets used Data #Orignal #influential Accuracy features
in the present study set features
features (%) identified
Dataset Method Accuracy SE SP AUC DS1 16 5 100.00 retro-orbital pain , arthralgia,
(%) feverduration,platelet,fever
NM 100.00 100.00 100.00 1.00 DS2 9 6 86.53 vomiting or nausea, myalgia,
C4.5 96.44 95.90 97.27 1.00 rash,bleedingsites,abdominal
LOR 91.02 89.49 93.36 0.96 pain,arthralgia
DS1
SVM 96.75 97.18 96.09 0.97 DS3 16 2 100.00 Hb,fever
NM 86.53 88.97 82.81 0.93 DS4 9 2 95.48 retro-orbitalpain,arthralgia
C4.5 82.35 87.18 75.00 0.84
LOR 72.91 74.36 70.70 0.78
DS2
SVM 78.17 89.49 60.94 0.75
NM 100.00 100.00 100.00 1.00 TABLEIIWilcoxonmatched-pairsranksumtestforcompar-
C4.5 94.97 95.41 93.55 0.99
ing the performance of NM with other methodologies used in
LOR 92.71 92.79 92.47 0.96
DS3
SVM 98.99 98.69 100.00 0.99 diagnosis of dengue fever
NM 95.48 98.03 87.10 0.95
C4.5 90.20 91.48 86.02 0.91 Dataset Method Ranksum(+,-) p-value
DS4 LOR 88.44 89.84 83.87 0.90 C4.5 55.0,0.0 0.002
SVM 92.71 98.03 75.27 0.87 LOR 55.0,0.0 0.002
DS1 SVM 45.0,0.0 0.004
C4.5 55.0,0.0 0.002
LOR 55.0,0.0 0.002
DS2 SVM 55.0,0.0 0.002
C4.5 36.0,0.0 0.008
LOR 36.0,0.0 0.008
retro-orbital pain (eye pain), myalgia (body pain), arthralgia
DS3 SVM 10.0,0.0 0.125
(jointpain),nauseaorvomiting,bleedinggums,rash,bleeding C4.5 38.5,6.5 0.074
LOR 37.0,8.0 0.098
sites, restlessness and abdominal pain and laboratory features:
DS4 SVM 27.0,9.0 0.25
haemoglobin (Hb), white blood cell count (WBC), packed
cell volume (PCV) and platelets. The last attribute in data
The above comparisons and statistical tests clearly demon-
set is the decision attribute. The clinical records are then re-
strate the significance of our methodology in identifying the
groupedintofourdatasets.Thefirstdataset(DS1)comprises
suspected dengue both in children and adults. The imputation
of 646 adults (age≥ 16 years) with clinical symptoms and
strategy employed in our methodology has improved the
laboratory features out of which 256 were dengue positive
classificationaccuracieswhencomparedwithC4.5whichuses
and 390 are dengue negative. The second data set (DS2) is
a modified information gain measure to generate the decision
a part of DS1 consisting of only clinical symptoms (ignoring
tree in presence of MV. The mean imputation strategies
the laboratory features) and has the same number of records
adopted in SVM and LOR could not render classification
as in DS1. The third data set (DS3) consists of 398 children
accuracies higher than NM.
(age between 5−15 years) [9] with clinical symptoms and
The features subsets identified by the NM is shown in
laboratory features, out of which 93 were dengue positive and
TableIII.Theapplicationoffeaturesselectionmethodreduced
305 were dengue negative. The fourth data set (DS4) is a part
the number of attributes by 75% in DS1 and 87.5% in DS3
of DS3 with only clinical symptoms and has same number of
data sets. Our methodology identified some of the clinical
records as DS3.
symptoms and laboratory features in adults (vomiting and
abdominal pain) different from those in children which are
in concurrence with earlier studies [22], [23]. The clinical
attribute rash was identified as an important feature in adults
but not in children. This may be explained by the relative
frequencyofthesecondaryinfectionsinadults[24].Arthralgia
was found to be influencing the final diagnosis of dengue
B. Results both in children and adults. The ROC curves comparing the
performance of NM with other methodologies are shown in
Figs. 1a-1d. The operating point or cut off point (p<0.001)
TheperformanceoftheNMiscomparedwithothermethod- is shown as a pentagon on each of the ROC curves. The ROC
ologies (C4.5, SVM and LOR) on the data sets used in the curves clearly demonstrate the superior performance of NM
present study and the classification accuracies are presented over other methods used in the diagnosis of dengue.
in Table I. A hundred percent accuracy is reported by NM
both in data sets DS1 and DS3. The Wilcoxon matched-pairs VI. PERFORMANCECOMPARISONOFNEWIMPUTATION
rank sum test results comparing the accuracies of NM with ALGORITHMWITHBENCHMARKINGDATASETS
other methodologies are shown in Table II. For example, the Since no specific studies on imputation of missing values
positiveranksumof55.0andnegativeranksumof0.0witha in dengue data sets we have utilized some bench marking
p-value<0.01forC4.5usingdatasetDS1(firstrowTableII) data sets obtained from Keel and University of California
indicates the superior performance of the new methodology Irvin (UCI) machine learning data repositories [25], [26] to
over C4.5 and also in respect of other methods as well. test the performance of the new imputation algorithm. The
(cid:13)c 20xxIEEE.Personaluseofthismaterialispermitted.PermissionfromIEEEmustbeobtainedforallotheruses,inanycurrentorfuturemedia,including
reprinting/republishingthismaterialforadvertisingorpromotionalpurposes,creatingnewcollectiveworks,forresaleorredistributiontoserversorlists,orreuseofany
copyrightedcomponentofthisworkinotherworks.DOI:10.1109/TITB.2011.2171978
(a)DS1 (b)DS2 (c)DS3 (d)DS4
Fig.1. ROCcurves
Wilcoxon statistics in Table IV is computed based on the
accuracies obtained by the new imputation algorithm with the
accuracies of those obtained by other imputation algorithms
using a C4.5 decision tree. The results in Table IV clearly
demonstrate the fact that our algorithm is superior to other
imputation algorithms as the positive rank sums are higher
than the negative rank sums (p<0.05) in all the cases.
TABLE IV Wilcoxon sign rank statistics for matched pairs
comparingthenewimputationalgorithmwithotherimputation
methods using C4.5 decision tree
Method RankSums Test Critical p-value
(+,-) Statistics Value Fig.2. ComputationalcomplexityoftheNM
FKMI 78.5,12.5 12.5 18 0.021
KMI 85.0,6.0 6 18 0.003
KNNI 76.0,15.0 15 18 0.032
presence of the linear trend between the time taken and the
WKNNI 83.0,8.0 8 18 0.006
varyingdatabasesizesensuresthenumericalscalabilityofthe
performance of NM in terms of asymptotic linearity.
VII. COMPUTATIONALCOMPLEXITY
VIII. COMPARISONOFRELATEDMETHODOLOGIESON
The computational complexity is a measure of the perfor-
DENGUESTUDIES
mance of the algorithm. For each data set having n attributes
and m records, we select only those subset of records m ≤ In this section we compare the results (Table V) obtained
1
m, in which missing values are present. The distances are in [8]–[10] with the results of our new methodology on the
computed for all attributes n excluding the decision attribute. current data set of 1044 individuals including children and
So, the time complexity for computing the distance would be adults. As compared to [9] where children with rash were
O(m ∗(n−1)).Thetimecomplexityforselectingthenearest having SE of 41.2% and SP of 95.5% our methodology when
1
records is of order O(m ). For computing the frequency of appliedonthedatasetDS2resultedinanaccuracyof86.53%,
1
occurrences for nominal attributes and average for numeric SE of 88.97% and SP of 82.81% which is considered to be
attributes the time taken would be of the order O(m ). a good classification model as both SE and SP are higher
1
Therefore, for a given data set with k-fold cross validation than 80%. In [10] both clinical and laboratory features were
having n attributes and m records, the time complexity of utilized to develop decision rules using C4.5 decision tree
our new imputation algorithm would be k ∗ (O(m ∗ (n − and they have reported a SE of 87.8% and SP of 75.7%. In
1
1) ∗ m) + 2 ∗ O(m )) which is asymptotically linear. Our comparisonto[10]ourmethodologywhenappliedonDS1and
1
experiments were conducted on a personal computer having a DS3hadresultedinSEof100%andSPof100%.Fromthese
Intel(R) core (TM) 2 Duo, CPU @2.93 GHZ processor with comparisonsweconcludethatthenewmethodologypresented
4 GB RAM. For each data set the computational time for inthisstudyifappliedonthedatasetsusedin[8]–[10]would
imputation and features selection is measured in terms of the yield more accurate results.
numberofCPUclockcycleselapsedinseconds.Basedonthe
results, we obtain a scatter plot (red line in Fig. 2) between IX. CONCLUSIONSANDDISCUSSION
the varying database sizes and the time taken by NM. Also, A new methodology (NM) with built in features for im-
we employed a linear regression on our results and obtained putation of missing values and identification of influential
the relation between the time taken (T) and the data size (D) attributes is discussed. The NM has out performed the state-
as T = 0.96D+5.54, α = 0.05, p < 0.05, r2 = 0.98. The of-the-art methodologies in diagnosis of dengue fever on all
(cid:13)c 20xxIEEE.Personaluseofthismaterialispermitted.PermissionfromIEEEmustbeobtainedforallotheruses,inanycurrentorfuturemedia,including
reprinting/republishingthismaterialforadvertisingorpromotionalpurposes,creatingnewcollectiveworks,forresaleorredistributiontoserversorlists,orreuseofany
copyrightedcomponentofthisworkinotherworks.DOI:10.1109/TITB.2011.2171978
TABLEVEvaluationofNMwithotherrelatedmethodologies
[3] S.DePaulaandB.Fonseca,“Dengue:areviewofthelaboratorytestsa
on dengue studies clinicianmustknowtoachieveacorrectdiagnosis,”BrazJInfectDis.,
vol.8(6),pp.390–398,2004.
State-of-the- #Patients Records Methods Accuracy SE SP
[4] S.-T.Chen,Y.-L.Lin,M.-T.Huang,M.-F.Wu,S.-C.Cheng,H.-Y.Lei,
art (DF) with (%) (%) (%)
MV C.-K.Lee,T.-W.Chiou,C.-H.Wong,andS.-L.Hsieh,“Clec5aiscritical
for dengue-virus-induced lethal disease,” Nature, vol. 453(7195), pp.
Chadwick 381 deleted LOR, 84.5 84 85
672–676,2008.
et al., [8] (148) OR
(clinical) [5] F. Ibrahim, N. Ismail, M. Taib, and A. W. Wan, “Modeling of
Chadwick 381 deleted -do- 76.5 74 79 hemoglobin in dengue fever and dengue hemorrhagic fever using bio-
et al., [8] (148) electricalimpedance.”PhysiolMeas,vol.25(3),pp.607–15,2004.
(laboratory) [6] M. N. Colleen, A. G. William, L. K. Merril, C. D. Naylor, and
Ramos et 938 manual -do- 68.95 41.2 95.5 S. L.Duncan, “Dealingwith missing datain observationalhealth care
al., [9] (38) update outcome analyses,” Journal of Clinical Epidemiology, vol. 53(4), pp.
(clinical, 377–383,2000.
children)
[7] K. J. Cios and W. Mooree, “Uniqueness of medical data mining,”
Tanner et 1200 deleted C4.5 81.75 87.8 75.7
ArtificialIntelligenceinMedicine,vol.26,pp.1–24,2002.
al., [10] (173)
[8] D.Chadwick,B.Arch,A.Wilder-Smith,andN.Paton,“Distinguishing
(laboratory)
Gomes et 20(15) - SVM 85 - - dengue fever from other infections on the basis of simple clinical and
al., [11] (gene laboratory features: application of logistic regression analysis,” J Clin
database) Virolology,vol.35(2),pp.147–153,2006.
NM (DS1) 1044 imputed ADT, 100 100 100 [9] M. M. Ramos, K. M. Tomashek, D. F. Arguello, C. Luxemburger,
(adults,clinical (256) (new GA L. Quiones, J. Lang, and J. L. Muoz-Jordan, “Early clinical features
&laboratory) algo- of dengue infection in puerto rico,” Transactions of the Royal Society
rithm) ofTropicalMedicineandHygiene,vol.103(9),pp.878–884,2009.
NM (DS2) 1044 -do- -do- 86.53 88.97 82.81
[10] L.Tanner,M.Schreiber,J.Low,A.Ong,andT.Tolfvenstam,“Decision
(adults, (256)
treealgorithmspredictthediagnosisandoutcomeofdenguefeverinthe
clinical)
NM (DS3) 1044 -do- -do- 100 100 100 earlyphaseofillness,”PLoSNeglTropDis,vol.2(3),pp.1–9,2008.
(children, (93) [11] A.L.V.Gomes,L.J.K.Wee,A.M.Khan,L.H.V.G.Gil,E.T.A.
clinical & Marques, Jr, C. E. Calzavara-Silva, and T. W. Tan, “Classification of
laboratory) denguefeverpatientsbasedongeneexpressiondatausingsupportvector
NM (DS4) 1044 -do- -do- 95.48 98.03 87.10 machines,”PLoSONE,vol.5(6),pp.1–7,2010.
(children, (305) [12] G.BatistaandM.Monard,“Ananalysisoffourmissingdatatreatment
clinical) methodsforsupervisedlearning,”AppliedArtificialIntelligence,vol.17,
no.5,pp.519–533,2003.
[13] J. Deogun, W. Spaulding, B. Shuart, and D. Li, “Towards missing
data imputation: A study of fuzzy k-means clustering method,” in
the four data sets considered in our experiments. The NM
4th International Conference of Rough Sets and Current Trends in
has generated a decision tree with an accuracy of 100.0% in Computing(RSCTC’04), ser. Lecture Notes on Computer Science, vol.
childrenandadultsusingbothclinicalandlaboratoryfeatures. 3066. LectureNotesInComputerScience,2004,pp.573–579.
[14] O. Troyanskaya, M. Cantor, G. Sherlock, P. Brown, T. Hastie, R. Tib-
Based on the performance measures we conclude that the use
shirani,D.Botstein,andR.Altman,“Missingvalueestimationmethods
of the new imputation strategy and features selection methods fordnamicroarrays,”Bioinformatics,vol.17,pp.520–525,2001.
withwrapperbasedsubsetevaluationusinggeneticsearchhas [15] Y. Freund and L. Mason, “The alternating decision tree learning al-
gorithm,” in Proceeding of the Sixteenth International Conference on
improved the accuracies of the predictions. Though the new
MachineLearningBled,Slovenia. ACM,1999.
methodologydiscussedinthispapermaybetakenasauniver- [16] U. Tadashi, M. Yoshihide, K. Daichi, S. Masami, and K. Kenji, “Fast
sal tool for the effective diagnosis of this disease, it remains multidimensional nearest neighbor search algorithm based on ellipsoid
distance,”InternationalJournalofAdvancedIntelligence,vol.1(1),pp.
to be seen whether or not this methodology is geographically
89–107,2009.
independent. However, we are willing to share our predictive [17] D.E.Goldberg,Geneticalgorithmsinsearch,optimizationandmachine
methodologies and strategies with the researchers working on learning. Addison-Wesley,1989.
[18] K. Ron and H. J. George, “Wrappers for feature subset selection,”
dengue fever all over the globe. We hold the view that more
ArtificialIntelligence,vol.97,pp.273–324,1997.
intensive and introspective studies of this kind will pave way [19] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector
for better clinical management and virological surveillance of machines,” ACM Transactions on Intelligent Systems and Technology,
vol.2(3),pp.1–27,2001.
dengue fever.
[20] I.WittenandE.Frank,DataMining:Practicalmachinelearningtools
andtechniques. MorganKaufmann,SanFrancisco.,2005.
[21] F.Wilcoxon,“Individualcomparisonsbyrankingmethods,”Biometrics
ACKNOWLEDGMENTS Bulletin,vol.1(6),pp.80–83,1945.
[22] J.G.-R.EnidandG.R.-P.Jos,“Dengueseverityintheelderlyinpuerto
WethanktheAssociateEditorandtheanonymousreviewers
rico,”PanAmJPublicHealth,vol.13(6),pp.362–368,2003.
for their constructive suggestions on our paper. This research [23] W. Ole, H. Suchat, B. Chureeratana, C. Kesinee, S. Yoawalark, and
is supported by the Foundation for Scientific Research and P. Sasithon, “Risk factors and clinical features associated with severe
dengue infection in adults and children during the 2001 epidemic in
Technological Innovation (FSRTI)- A Constituent Division
chonburi, thailand,” Tropical Medicine and International Health, vol.
of Sri Vadrevu Seshagiri Rao Memorial Charitable Trust, 9(9),pp.1022–1029,2004.
Hyderabad - 500 035, India. [24] C. Cobra, J. G. Rigau-Prez, G. Kuno, and V. Vomdam, “Symptoms
of dengue fever in relation to host immunologic response and virus
serotype, puerto rico, 19901991,” American Journal of Epidemiology,
REFERENCES vol.142(11),pp.1204–1211,1995.
[25] A. Alcal-Fdez, A. Fernandez, Luengo, J. Derrac, S. G. J., L. Snchez,
[1] D.Gubler,“Dengueanddenguehemorrhagicfever,”ClinicalMicrobi- and F. Herrera, “Keel data-mining software tool: Data set repository,
ologyReviews,vol.11,pp.480–96,1998. integrationofalgorithmsandexperimentalanalysisframework,”Journal
[2] T.P.Monath,“Dengue:Therisktodevelopedanddevelopingcountries,” ofMultiple-ValuedLogicandSoftComputing,2010.
ProceedingsoftheNationalAcademyofSciencesoftheUnitedStates [26] A. Frank and A. Asuncion, “UCI machine learning repository,” 2010.
ofAmerica,vol.91(7),pp.2395–2400,1994. [Online].Available:http://archive.ics.uci.edu/ml