Computing Conference 2017

18-20 July 2017 | London, UK

A Belief Rule based Expert System to Diagnose

Dengue Fever under Uncertainty
Mohammad Shahadat Hossain and Israt Binteh Habib Karl Andersson
Department of Computer Science and Engineering Pervasive and Mobile Computing Laboratory
University of Chittagong Luleå University of Technology
University-4331, Bangladesh SE-931 87 Skellefteå, Sweden
[email protected], [email protected] [email protected]

Abstract—Dengue Fever is a debilitating mosquito-borne fever (usually between 39.5-41°C, persisting for 1-7 days
disease, causing sudden fever, leading to fatality in many cases. A only), severe headache, pain behind eyes, congested
Dengue patient is diagnosed by the physicians by looking at the conjunctiva, muscle and joint pains, nausea or vomiting and
various signs, symptoms and risk factors of this disease. rash [7]. While DHF is a less common but it is a severe form
However, these signs, symptoms and the risk factors cannot be of dengue where the patient shows bleeding. DSS is considered
measured with 100% certainty since various types of as the most risky, responsible for most of the deaths, especially
uncertainties such as imprecision, vagueness, ambiguity, and in children. It is important to note that all of the signs and
ignorance are associated with them. Hence, it is difficult for the symptoms may not appear concurrently, making the diagnosis
physicians to diagnose the dengue patient accurately since they
of dengue difficult. In addition to the signs and symptoms,
don’t consider the uncertainties as mentioned. Therefore, this
various risk factors such as mosquito breeding place, mosquito
paper presents the design, development and applications of an
expert system by incorporating belief rule base as the knowledge bite and population under risk of dengue fever should be taking
representation schema as well as the evidential reasoning as the into account to perform the task of diagnosis of the disease.
inference mechanism with the capability of handling various However, the accurate measurement of these signs,
types of uncertainties to diagnose dengue fever. The results symptoms and risk factors is challenging because there exists
generated from the expert system are more reliable than from different types of uncertainties with each of them. Eventually,
fuzzy rule based system or from human expert. it is difficult to obtain the accurate early diagnostics result,
which is crucial to detect this disease, and to safe human life,
Keywords—Belief Rule Base; Dengue Fever; Evidential
Reasoning; Expert System; Uncertainty especially the children as mentioned. Therefore, the research
presented in this paper carried out a comprehensive survey in
I. INTRODUCTION consultation with the expert physicians, located in the various
hospitals, medical colleges in the City of Chittagong,
Dengue is a mosquito-borne viral disease; causing flu-like
Bangladesh to identify the various types of uncertainties
illness which can be turned into lethal complication, called
associated with each sign, symptom and risk factor. It is
severe dengue. The global occurrence of this disease has
important to note that dengue frequently occurs in this part of
increased dramatically in the recent years, and hence; half of
Bangladesh. Table 1 illustrates the results of the survey, for
the world’s population is now at risk [1]. Approximately 975
example, the types of uncertainties associated with “fever” as
million of these populations live in urban areas in tropical and
identified, consisting of randomness and ignorance. The reason
sub-tropical countries of South-East Asia, the Pacific and the
for this is that during dengue fever, temperature may go up and
Americas [2]. Ninety percent of the dengue patients are
down, which may cause uncertainty due to randomness and
children under the age of 15 years and it is endemic in 112
ignorance, resulting wrong diagnosis of the disease.
countries [3].
The current gold standard technique in monitoring risk in
The dengue virus usually spreads through a cycle of
dengue is to take the patient’s blood sample for the
transmission, comprising human-to-mosquito-to-human [4][5].
measurement of total haematocrit (HCT), haemoglobin (Hb)
Typically, a high level of dengue virus in the blood of a person
concentration, platelet count, and liver function status [2][8].
is noticed four days after being bite by an infected Aedes
However, such approach is considered as invasive and hence,
aegypti mosquito. While, five days after being bite by the
regular blood sampling will cause further injury to the
mosquito, the various signs and symptoms of dengue fever,
subcutaneous tissue and potential risk to the dengue patients
consisting of fever, headache, rash, congested conjunctive and
[9]. On the contrary, non-invasive approach to diagnose
relative bradicardia are noticed [6]. Although there is no
dengue fever is in practice, carried out by looking at the signs,
specific treatment for dengue or severe dengue, an early
symptoms and risk factors. However, the effectiveness and the
diagnosis by looking at the signs and symptoms could play an
accuracy of the non-invasive approach depend on handling the
important role to reduce the rate of mortality [1].
various types of uncertainties. Therefore, the traditional way of
There are three different forms of dengue fever. They are diagnosis of dengue fever by using non-invasive approach,
Dengue Fever (DF), Dengue Haemorrhagic Fever (DHF) and usually carried out by the physicians, is incapable of producing
Dengue Shock Syndrome (DSS). DF is characterized by high accurate result. This can be addressed by developing an expert

179 | P a g e
978-1-5090-5443-5/17/$31.00 ©2017 IEEE
 Computing Conference 2017
18-20 July 2017 | London, UK
system with the capability of handling uncertainty in prediction. A rule based expert system to classify risk in
knowledge representation as well as in inference mechanism. dengue infections by using bioelectrical impedance analysis
Therefore, this paper presents the development of an expert developed [13]. An expert system for diagnosing dengue fever
system (ES), based on a methodology, known as the belief based on basic information gathered from literature study and
rule-base inference methodology using the evidential reasoning expert interview developed [14]. Although fuzzy logic based
(RIMER) [10]. expert system can handle uncertainty due to vagueness,
ambiguity and fuzziness, it is incapable of handling uncertainty
TABLE I. TYPES OF UNCERTAINTY IN DENGUE DISEASE due to ignorance, incompleteness and randomness. From Table
Dengue 1 it can be observed that the uncertainties associated with the
Uncertainty Types Discussion signs and symptoms of the dengue fever are associated with
Factors
There are many similarities among incompleteness, ignorance and randomness in addition to the
dengue, chikungunya and flu virus vagueness, fuzziness and ambiguity. ANN based approach is a
Ignorance,
Fever [10]. In dengue disease temperature
randomness
goes ups and down, so random data type of black-box based system, which is incapable of handling
may give wrong result. various types of uncertainties associated with the signs and
Patients sometimes may be unable symptoms of dengue fever. Neither [13] nor [14] used the rule
Imprecision, to express their pain precisely. base in the knowledge representation schema or in the
Headache
incompleteness Without precise expression expert
assessment is incomplete.
inference mechanism are capable of addressing the uncertainty
Patients may have some rash in their issues that have been elaborated in related to dengue fever.
Rash Ignorance body but they may ignore this
symptom. However, Belief Rule Based Expert Systems (BRBES) are
It is possible that patients may have widely used in the clinical domain where there exists plethora
Congested
allergic problem. That’s why of uncertain clinical information [15][16]. In the BRBES,
Inconsistency conjunctivitis occurs. In this way belief rule base (BRB) is used to develop the knowledge base
Conjunctiva
inconsistent data may be mixed with
dengue assessment. by addressing various types of uncertainties such as
As relative bradicardia is measured incompleteness, ambiguity, vagueness and imprecision. While
Relative Complexity,
by digital blood pressure machine, it inference mechanism consists of number procedures namely
contains quantitative data. So input transformation, rule activation weight calculation, belief
bradicardia inconsistency
complex information may lead to
misdiagnosis.
update and rule aggregation by using evidential reasoning
Expert avoids measuring this issue. algorithm. The strength of BRBES’s inference procedure is
So ignorance occurs. Sometimes that it has the capability to handle various types of uncertainty
mosquito bite is low but patients and hence, facilitates a robust decision making process. In the
Vagueness,
Number of may have dengue fever, sometimes
mosquito bite
ignorance,
mosquito bite is high but patients
following section an overview of BRBES methodology will be
randomness presented.
may don’t have dengue fever
consequently. So random data may
lead to misdiagnosis. III. OVERVIEW OF BRBESS METHODOLOGY
Mosquito
breeding place
Incompleteness Expert assessment is incomplete. The Belief Rule Based Expert Systems (BRBESs)
Experts may not have the correct methodology consists of two parts, namely knowledge
Population Ignorance information about the population representation schema by using belief rule base and the
density of the area. inference mechanism, which consists of a number of
The remaining of the paper is structured as follows. Section procedures. This section elaborates each of them.
two presents the related works. Section three provides an A. Domain Knowledge Representation using BRB
overview of RIMER methodology. Section four presents the
architecture, design and implementation of the proposed Belief Rule Base (BRB) is the extension of traditional IF-
BRBES. Experimental results and discussion are then THEN rule base where belief degrees are embedded with the
presented. A conclusion is included to summaries the referential values of the consequent attribute while antecedent
contribution of the research. part contains the referential value of the antecedent attribute. A
belief rule base (BRB) is used to capture the nonlinear causal
II. LITERURE REVIEW relationships under uncertainty, which is not the case with the
traditional IF-THEN rule base. Equation (1) represents the
Fuzzy logic based expert system to diagnose seven tropical example of a belief rule.
infectious diseases developed [11], where dengue fever
considered. This system used fuzzy logic based knowledge ( )( ) (
­ IF X 1 is A1k Λ X 2 is A2k Λ.......Λ X Tk is ATkk
Rk : ®
) (1)
representation schema to handle uncertainty due to vagueness
while certainty factor used to handle the inability of an expert ¯THEN {(D1isB1k ), (D2isB2 k ), ......, (DN is BNk )}
N
in defining the relationship between the symptoms of the
disease with certainty. Artificial Neural Network (ANN) based ¦β
j =1
jk ≤1
system proposed to facilitate the analysis of dengue fever [12].
It describes a non-invasive prediction system, enabling the ‫ߠݐ݄݃݅݁ݓ݈݁ݑݎ݄ܽݐ݅ݓ‬௞ ሺ݇ ൌ ͳǡ ǥ ǡ ‫ܮ‬ሻ
prediction of the day of the defervescence of the dengue ܽ݊݀ܽ‫ߜݏݐ݄݃݅݁ݓݏ݁ݐݑܾ݅ݎݐݐ‬ଵ ǡ ߜଶ ǡ ǥ ǡ ߜ்ೖ
patient. This system used multilayer feed-forward neural Here, ܺ௜ೖ ሺ݅ ൌ ͳǡ ǥ ǡ ܶ௞ ሻ is referred to as ith antecedent
networks and used clinical signs and symptoms to carry out the attribute while ‫ܣ‬௞்ೖ ( ݅ ൌ ͳǡ ʹǡ ǥ ǡ ܶ௞ ) represents its referential

180 | P a g e
978-1-5090-5443-5/17/$31.00 ©2017 IEEE
 Computing Conference 2017
18-20 July 2017 | London, UK
value.ܶ௞ is the number of antecedent attributes used in the kth remaining matching degrees will be assigned to the
rule. Di (i=1, 2, …, N) is the referential value of the consequent corresponding referential values of an antecedent attribute by
attribute with degree of belief ȕik (i=1, 2, …, N). The kth rule is taking account of the fact that they are related to different
said to be complete if σ௡௝ୀଵ ߚ௝௞ ൌ ͳ otherwise, it is belief rules.
incomplete. L is the number of belief rules in the rule base.
Equation (2) represents the more precise example of a belief TABLE II. INPUT TRANSFORMATION
rule from the domain of dengue fever. Serial Antecedent Antecedent
High Medium Low
No. Name Value
­ IF (Population is High) ∩ (Mosquito Breeding Place is (2)
° 1 Rash Low,10% 0.0 0.2 0.8
° Medium) ∩ (Mosquito Bite is Medium) Congested Medium,
R1 : ® 2 0.0 1 0
° THEN Dengue Risk Assessment is Conjunctiva 50%
¯° {(High(0.4), Medium(0.6), Low(0.2))} 3 Relative Bradicardia High, 80% 0.6 0.4 0
Here, high, medium and low are the referential values of 2) Rule Activation Weight calculation
the consequent attribute “Dengue Risk Assessment” of the This consists of calculating the combined matching degree
belief rule. The above rule states that the risk of dengue of a Rule followed by the calculation of its activation weight. Eq.
person in his present living condition is high with 40% degree (1) illustrates that the rule consists of more than one antecedent
of belief, while medium with 60% degree of belief and low attributes. Therefore, the combined matching degree of a rule
with 0% degree of belief. This belief rule is complete since the considering all the antecedent attributes is necessary to find
summation of degree of belief associated with each of the out. By applying simple weighted multiplicative functions as
referential value of the consequent attribute is 1 shown in Eq. (6) this can be achieved. The use of this function
[(0.4+0.6+0+0)=1]. Hence, it can be argued that the rule is allows the demonstration of the relationship among the
complete and in this way uncertainty due to incompleteness antecedent attributes of a rule.
can be addressed in the belief rule base.
α ki = ∏i =1 (α ik )
Tk δ ki' (5)
B. Inference Procedure in BRBESs
The inference procedure of BRBEs consists of input Here, Įi represents the individual matching degree of a
transformation, rule activation weight calculation, belief update referential value of an antecedent attribute as described earlier.
and rule aggregation as will be discussed below. įkt is the relative weight of an antecedent attribute, which is
used in the kth rule and is calculated by dividing the weight of
1) Input Transformation the attribute by the maximum weight of all antecedent
The input transformation consists of finding the matching attributes of the kth rule to normalize the value of įkt which
degrees of the input value of an antecedent attribute to its means that its value should range between 0 and 1.
various referential values by using equations (3), (4) and (5).
Hence, this is equivalent to transforming an input value into a When the kth rule is activated, the weight of activation of
distribution over the referential values of an antecedent the kth rule, ߱݇, is calculated by using the following formula
attribute. The input value of an antecedent attribute (for [17][18][19][20].
example headache) is usually collected from the patients or
θ k ∏ i =1 (α ik ) (6)
Tk δ ki'
from the physicians in terms of linguistic terms such as high, θ kα k δ ki
ωk = = , δ ki' =
medium, low and normal. It can also be collected as numerical ¦
L
θ j ª∏ i =1 (α i ) º max i =1,...., Tk {δ ki }
¦
'
θ jα j L Tk j δ ji
j =1 j =1 «¬ »¼
value for example in case of fever.
3) Belief Update
IF ( H value≥ Input value≥ M value) THEN (3) There could be some situations when it would not be
H value − Input value
Medium =
H value − M value
, High = 1− Medium, Low = 0.0 possible to collect the input values of all the antecedent
attributes of a rule and hence, the uncertainty due to ignorance
occurs. Such situation of uncertainty can be handled by
IF (Mvalue ≥ Inputvalue ≥ Lvalue )THEN (4)
Mvalue − Inputvalue updating the belief degrees assigned to the referential values of
Medium = , Medium = 1 − Low High = 0.0
M value − Lvalue the consequent attribute during the development of the initial
Table 2 shows the matching degree of the antecedent belief rule base by applying Eq. (6).
attribute value into its different referential values. For example,
the input value of the antecedent attribute “Rash” is identified
βik = βik
¦
Tk

t=1 (τ (t, k) ¦ α ) Jt

j=1 tj
(7)
as “low”, which is weighted as 10% by the expert and its
¦ τ ( t, k )
Tk

corresponding matching degrees associated with the referential t=1

values (in this case they are “High”, “Medium” and “Low”) where
obtained by using equations (3), (4) and (5). A rule is said to be
­ 1,if P isused in defining R t = 1,...,T
k( k)
activated when the matching degrees (ĮH ,ĮM, ĮL) are assigned °
to one of the referential values of the antecedent attribute. This τ ( t, k ) = ® i

phenomenon is considered as the packet antecedent of a rule. °̄ 0, otherwise

The reason for assigning matching degree to only one
referential value of the antecedent attribute of a rule is that only Here, ߚҧ ik is the original belief degree while ȕik is the
one referential value is considered as shown in Eq. (1). A belief updated belief degree. The original belief degree updated if any
rule base consists of a number of rules. However, the

181 | P a g e
978-1-5090-5443-5/17/$31.00 ©2017 IEEE
 Computing Conference 2017
18-20 July 2017 | London, UK
ignorance noticed. For example, if the antecedent “headache” The explanation of this architecture is given below:
is ignored, then the initial belief degrees are updated as shown
in Table 3. 1) Interface Layer
This layer enables the interaction between the users and the
TABLE III. BELIEF DEGREE UPDATE BRBES. Users can input the value of the antecedent attribute
as well as can see the system outputs. A combination of tools
Rule High Medium Low Activation used to develop the BRBES interface, consisting of PHP,
Id D1 D2 D3 Weight
Initial 1 0 0
HTML, Javascript, and JQuery. The reason for using these
Matching tools is that they allow the creating of user-friendly web-based
1 0 0 interface by using simple code.
1 Degree 0.096
Update 0.096 0 0
2) Application processing layer
Initial 0.6 0.3 0.1 This layer supports the inference procedures of the BRBES
Matching consisting of input transformation, rule activation weight
2 0.7 0.2 0.1 0.064
Degree calculation, belief update and the rule aggregation. PHP has
Update 0.0448 0.0128 0.0064 been used in this layer since due to its simplicity and shorter
4) Rule Aggregation development cycles this language is largely used. Moreover, it
In order to diagnose the dengue fever based on the input allows the system to be accessible through online.
values of the signs, symptoms and risk factors, the activated
rules need to be aggregated. The aggregation can be achieved 3) Data Management Layer
by applying the Eq. (9), which encapsulates the evidential This layer stores and manages the knowledge base.
reasoning algorithm [4][11]. MySQL, which is a flexible and user-friendly relational
database, is used at the back-end to store and manipulate the
βj =
μ× [∏ (ω β
L
k =1 k jk ) L
(
+ 1 − ω k ¦ j =1 β jk −∏ k =1 1 − ω k ¦ j =1 β jk
N N
)] (8) initial BRB. MySQL also allows the faster access of data and
1− μ × [∏ L
k =1
1 − ωk ] ensures the security as needed by the BRBES.
(9) There exist some individuals such as domain experts,
( ) − (N − 1)º»
−1
ª N

«¦∏ k =1 ω k β jk + 1 − ω k ¦ j =1 β jk
L N
knowledge engineers and users of the system, who play
μ = « j =1 » important roles to acquire knowledge and to develop the
«× L 1 − ω
¬« ∏k =1
(
k ¦ j =1 β jk
N
) »
¼» knowledge base of the expert system. Domain experts who
When the diagnosis of the dengue fever achieved [using have the knowledge of dengue fever and in our case they are
Eq. (8)] with certain belief degrees associated with each of the the physicians engaged in giving treatment to the patients. On
referential values (High, Medium, Low), this can be converted the contrary, knowledge engineers encode the expert's
into crisp value by using some preference values [4] by knowledge in a declarative form that can be used by the expert
applying Eq. (9). system in the form of initial belief rule-base. The BRBES
architecture as shown in Fig. 1 incorporates the above
y m = ¦ n =1 D n × Bn (m ) (10)
N
viewpoint.
B. Knowledge Base Construction in BRB
IV. BELIEF RULE BASED EXPERT SYSTEM (BRBES) TO In consultation with the domain experts a BRB framework
DIAGNOSE DENGUE FEVER to diagnose dengue fever has been developed (Fig. 2) and this
This section presents the design, implementation, is considered as the starting point to construct the knowledge
knowledge base construction and an overview of BRBES base. The BRB framework as depicted in Fig. 2 represents a
interface. hierarchical multilevel as well as nonlinear structure necessary
to diagnose the dengue fever. The tree consists of three
A. Architecture, Design and Implementation of the BRBES different types of nodes namely leaf nodes, mid-level nodes
The architecture of a system can be defined as the and the root node. The data related to the leaf nodes attributes
organization of its components, consisting of inputs, process, need to be collected to obtain the root node value and this can
and outputs. In addition, the pattern of the system organization, be obtained by applying various inference algorithms as
known as architectural style, is also considered. The proposed mentioned before.
BRBES architectural framework adopts three-layer
For example, the level of the critical signs (A11) of the
architectural style, consisting of interface, application
dengue fever of a patient can be assessed by getting input
processing and data management layers as depicted Fig. 2.
values of the three child nodes (A4, A5, and A6).

182 | P a g e
978-1-5090-5443-5/17/$31.00 ©2017 IEEE
 Computing Conference 2017
18-20 July 2017 | London, UK

Fig. 2. The BRB Framework to Diagnose Dengue Fever

TABLE IV. INITIAL BRB FOR A11 SUB RULE BASE

Rule Rule IF Antecedent THEN Consequent
Id weight A4‫ޔ‬A5‫ޔ‬A6 is A11 is
1 1 H‫ޔ‬H‫ޔ‬H {(H,1.00), (M,0.00), (L,0.00)}
2 1 H‫ޔ‬H‫ޔ‬M {(H,1.00), (M,0.00), (L,0.00)}
3 1 H‫ޔ‬H‫ޔ‬L {(H,0.9),(M,0.1), (L,0.00)}
Fig. 1. The BRBES architecture 4 1 H‫ޔ‬M‫ޔ‬H {(H,0.4), (M,0.5), (L,0.1)}
5 1 H‫ޔ‬M‫ޔ‬M {(H,0.3), (M,0.4), (L,0.3)}
The BRB framework to diagnose dengue fever consist of 6 1 H‫ޔ‬M‫ޔ‬L {(H,0.00),(M,0.5), (L,0.5)}
four sub-rule bases namely A13, A12, A11 and A10. The total ... .... ... ....
number of rules in a sub-rule base can be calculated by the 25 1 L‫ޔ‬L‫ޔ‬H {(H,0.00), (M,0.0), (L,1.00)}
26 1 L‫ޔ‬L‫ޔ‬M {(H,0.00), (M,0.00), (L,1.00)}
following Eq. (11).
27 1 L‫ޔ‬L‫ޔ‬L {(H,0.00), (M,0.00), (L,1.00)}
(11)
L = ∏ i =1 J i
T
C. BRB Interface
where An interface is a media through which the users interact
with the system. Fig. 3 illustrates the interface of the BRBES to
L = Total number of rules in a sub rule base diagnose dengue fever. The interface allows the inputting of
‫ܬ‬௜ = Referential values of ith antecedent attributes the leaf nodes data of the BRB framework (Fig. 3), which are
For example, A11 sub rule base consists of three attributes collected from the patients or the physicians. This interface
(A4, A5, and A6) and each of them with three referential facilitates the displaying of the results not only at the top level
values. Hence, J4=3, J5=3, and J6=3, and therefore but also at the mid-level in terms of fuzzy values or crisp
L=(3*3*3)=27 after applying Eq. (11). Thus, this sub rule base values (Fig. 3). These fuzzy values can be converted into crisp
consists of 27 rules as illustrated in Table 4. A BRB can be value by using Eq. 10. For example, the fuzzy values of the
established in four different ways: (1) by using domain expert “common sign”, which is a mid-level node, obtained for its
knowledge; (2) by examining historical data; (3) by using different referential values consists of (High, 75.09%),
previous rules if available; and (4) by creating rules in a (Medium, 22.688%) and (Low, 2.222%). The belief degree
random way without any prior knowledge. In this research, attached with the referential values of the “common sign”
initial BRB has been constructed by extracting heuristic belief calculated by using Eq. (9). Using Eq. (10) the crisp value for
rules from the domain experts. It is assumed that all the rules these fuzzy values obtained which is 86.434. In this way, the
and the attributes have equal weight. Initial BRB for the A11 final crisp value for the top-level node or the root node can be
sub rule base is listed in Table 4. An example of a belief rule obtained which is 71.433% (Fig 3).
taken from Table 4 is elaborated below. V. RESULTS AND DISCUSSION
R6: IF Rash is High AND Congested Conjunctivas is To test the reliability and the validity of the BRBES to
Medium AND Relative Bradicardia is Low THEN Critical Sign diagnose dengue fever the leaf nodes data of the BRB tree (Fig.
is {(High,0), (Medium, 0.5), (Low,0.5)}. 2) from the patients of the two hospitals, located in the city of

183 | P a g e
978-1-5090-5443-5/17/$31.00 ©2017 IEEE
 Computing Conference 2017
18-20 July 2017 | London, UK
Chittagong, Bangladesh, collected. The collected data were fed with AUC of 0.814 (95% confidence intervals 0.730 – 0.898),
into BRBES to obtain the suspicion level of dengue fever for which is less than AUC of BRBES (0.833).
each patient as shown in Table V. In addition, the expert
opinion or the opinion of the concerned physicians on the
suspicion level of dengue fever of each the patient was also
collected as shown in Table V.

Fig. 4. ROC curves comparing BRBES’s result and Expert’s opinion

TABLE V. DENGUE ASSESSMENT RESULT BY BRBES AND FUZZY LOGIC

Fig. 3. BRBES Interface for Dengue Fever BASED EXPERT (FRLES) SYSTEM

It is interesting to note that the physicians were not aware Case Expert
BRBES FLBES Output
Study Opinion
of the uncertainty phenomenon associated with the signs and
1 65.4440 65 63.9 1
symptoms of the dengue fever. Eventually, there could be high 2 84.6215 85 76.8 1
risk of inaccuracy with the physicians’ suspicion level of 3 37.9265 39 42.9 0
dengue fever. Laboratory test data of the patients considered as 4 81.9800 83 75.3 1
the benchmark data. For example, if the lab test confirms a 5 59.3295 53 59.4 0
patient with dengue fever then it is considered as “1” otherwise 6 71.4330 70 68 1
it is considered as “0”. 100 dataset have been collected from 7 75.5745 70 74.5 1
the patients to test the BRBES. For simplicity, Table V only 8 44.8850 42 37 0
9 29.3610 25 29 0
presents data collected from ten patients.
10 51.2925 55 49.6 0
The Receiver Operating Characteristic (ROC) curves are
widely used to analyze the effectiveness of diagnostic tests
having ordinal or continuous results [21][22]. Therefore, it has
been considered to measure the reliability of BRBES. The
accuracy or the performance of the BRBES in assessing the
suspicion level of dengue fever can be measured by calculating
the Area Under Curve (AUC). If the AUC is one then the
results generated by the BRBES can be considered as accurate.
The ROC curve plotted by the blue line in Fig. 4 is associated
with the results generated by the BRBES with AUC of 0.833
(95% confidence intervals 0.755 - 0.911). The ROC curve
plotted by the green line in Fig. 4 is obtained against the
physician’s opinion, and its AUC is 0.810 (95% confidence
intervals 0.726 - 0.894).
Table VI illustrates the results generated by using Fuzzy
Logic Based Expert System (FLBES) along with BRBES and
expert opinion for the same set of data as shown in Table V.
Fig. 5 illustrates the ROC curves for the BRBES, FLBES
and expert opinion. The ROC curve plotted by the green line in
Fig. 5. ROC curves for comparing BRBES and Fuzzy Rule Based Expert
Fig. 4 is associated with the results generated by the FLBES System (FRBES)

184 | P a g e
978-1-5090-5443-5/17/$31.00 ©2017 IEEE
 Computing Conference 2017
18-20 July 2017 | London, UK

TABLE VI. DATASET FOR SYSTEM TESTING

Sl Expert opinion
A1 A2 A3 A4 A5 A6 A7 A8 A9 System Result (%) Benchmark
No (%)
1 105´ High High Medium Medium High High High High 65.4440 65 1
2 105´ High High High High High High High High 84.6215 85 1
3 98.6´ Low Low Low Low Low High High Medium 37.9265 39 0
4 105´ High High High High High High High High 81.9800 83 1
5 100´ High High Medium High High High High Medium 59.3295 53 0
6 105´ High High High High High High High High 71.4330 70 1
7 105´ High High High High High High High High 75.5745 70 1
8 104´ High High Low Low Low High Medium Medium 44.8850 42 0
9 98.6´ Low Low Low Low Low Low Low Low 29.3610 25 0
10 105´ Low Low High Low Low High High Medium 51.2925 55 0
[5] G. Kuno, "Factors Influencing the Transmission of Dengue Viruses, "In
TABLE VII. RELIABILITY COMPARISON AMONG THREE EXPERT SYSTEMS Dengue and Dengue Hemorrhagic Fever,” eds. D. J. Gubler & G. Kuno
(Cambridge:CABI, 2001):61–88.
Asymptotic 95% [6] F. Rodhain and L. Rosen, "Mosquito Vectors and Dengue Virus-Vector
Expert System AUC
confidence interval Relationships." In Dengue and Dengue Hemorrhagic Fever, eds. D. J.
BRBES 0.833 0.755-0.911 Gubler & G. Kuno (Cambridge: CABI, 2001):45–60.
FLBES 0.814 0.730-0.898 [7] World Health Organization, “Dengue: Guidelines for Diagnosis,
Expert’s opinion 0.810 0.726-0.894 Treatment, Prevention and Control,” Geneva: World Health
Organization and the Special Programme for Research and Training in
VI. CONCLUSION Tropical Diseases, 2009.
In recent years, the fatality rate of dengue fever increases [8] C.H. Kuo, D.I. Tai, C.S. Chang-Chein, C.K. Lan, S.S. Chiou, and Y.F.
Liaw, "Liver Biochemical test and Dengue Fever", The American
substantially all over the world. Therefore, an effective system Journal of Tropical Medicine and Hygiene, 47(3):265-270, 1992.
to diagnose dengue disease is essential for the people who are
[9] F. Ibrahim, M.N. Taib., W.A.B.W. Abas, C.C. Guan, and S. Sulaiman,
living in dengue risk zones of the world. This research “A novel approach to classify risk in dengue hemorrhagic fever (DHF)
demonstrates the development and applications of a belief rule using bioelectrical impedance analysis,” IEEE Transactions on
based expert system (BRBES) with the ability to handle Instrumentation and Measurement 54(1):237–244, 2005.
various types of uncertainties associated with the signs, [10] D.E. Carey, ”Chikungunya and dengue: A case of mistaken identity?” J.
symptoms and risk factors of dengue fever. It has been Hist. Med. Allied Sci. 26(3):243-262, 1971.
demonstrated that the BRBES generated results are better than [11] P.M. Prihatini and K.G.D. Putra, “Fuzzy Knowledge-based System with
that of expert opinion or the fuzzy rule based expert system. In Uncertainty for Tropical Infectious Disease Diagnosis,” IJCSI
International Journal of Computer Science Issues, Vol. 9, Issue 4, No 3,
addition, different types of scenarios or the simulated results pp. 157-163, 2012.
can be generated by using the BRBES, for example at the mid- [12] F. Ibrahim, M.N. Taib, W.A. Abas, C.C. Guan, and S. Sulaiman, “A
level factors of the BRBES tree such as “Common Sign”, novel dengue fever (DF) and dengue haemorrhagic fever (DHF) analysis
“Critical sign” and “risk factors”. This type of capability will using arti¿cial neural network (ANN),” Computer Methods Programs
help to analysis the disease from various perspectives to Biomed. 79(3):273-281, 2005.
develop an appropriate strategic plan to reduce the risk of [13] F. Ibrahim, M.I. Mohamad, S.N. Makhtar and J. Ibrahim, “Classification
dengue fever in the context of a country. of Risk in Dengue Fever and dengue Haemorrhagic Fever using Rule
Based Expert System,” 3rd Kuala Lumpur International Conference on
ACKNOWLEDGMENT Biomedical Engineering, pp. 50-53, 2006.
[14] S. Karim, H. Suryaningsih, and A. Lause, "Expert System for
We would like to extend our gratitude to the Swedish Diagnosing Dengue Fever", Seminar Nasional Aplikasi Teknologi
Research Council (under grant 2014-4251) for the funding and Informasi (SNATI 2007)
the provision of resources to conduct this research. [15] M.S. Hossain, K. Andersson, and S. Naznin, “A Belief Rule Based
Expert System to Diagnose Measles under Uncertainty,” 2015
REFERENCES International Conference on Health Informatics and Medical Systems
[1] World Health Organization, “Dengue and dengue haemorrhagic fever,” (HIMS'15), Las Vegas, NV, USA, 2015.
Factsheet No 117, revised May 2008. Geneva, World Health [16] R. Karim, K. Andersson, M.S. Hossain, M.J. Uddin, and M.P. Meah, “A
Organization, 2008 (http: //www. who. int /mediacentre/ Belief Rule based Expert System to Assess Clinical Bronchopneumonia
factsheets/fs117/en/). Suspicion,” Future Technologies Conference (FTC 2016), San
[2] F.P. Pinheiro and S.J. Corber, “Global situation of dengue and dengue Francisco, CA, USA, 2016.
haemorrhagic fever and its emergence in the Americas,” World Health [17] J.B. Yang, J. Liu, J.Wang,, H.S. Sii and H.W. Wang, “Belief rule-base
Stat Q 1997;50:161–169. inference methodology using the evidential reasoning approach–
[3] World Health Organisation, “Prevention and control of dengue and RIMER”, IEEE Transactions on Systems Man and Cybernetics Part A-
dengue haemorrhagic fever: comprehensive guidelines,” WHO Regional Systems and Humans 36(2):266-285, 2006.
publication, SEARO, No 29, 1999. [18] J.B. Yang, “Rule and utility based evidential reasoning approach for
[4] Alameda County Mosquito Abatement District, "Biological Notes on multi-attribute decision analysis under uncertainties”, European Journal
Mosquitoes," Life cycle of the mosquito (2011). of Operational Research 131(1):31–61, 2007.

185 | P a g e
978-1-5090-5443-5/17/$31.00 ©2017 IEEE
 Computing Conference 2017
18-20 July 2017 | London, UK
[19] M.S. Hossain, P.O. Zander, S. Kamal, and L. Chowdhury, ”Belief Rule [21] C.E. Metz. “Basic principles of ROC analysis,” Semin. Nucl. Med.
Based Expert Systems to Evaluate E-Government: A Case Study,” 8(4):283–298, 1978.
Expert Systems 32(5):563-577, 2015. [22] J.A. Hanley. “The Robustness of the “Binormal,” Assumptions Used in
[20] Y.M. Wang, J.B. Yang, and D.L. Xu, ”Environmental impact Fitting ROC Curves,” Med. Decis. Making 8(3):197–203, 1988.
assessment using the evidential reasoning approach,” European Journal
of Operational Research 174(3):1885 –1913, 2006.

186 | P a g e
978-1-5090-5443-5/17/$31.00 ©2017 IEEE