Content uploaded by Paweł Pławiak
Author content
All content in this area was uploaded by Paweł Pławiak on Jun 12, 2019
Content may be subject to copyright.
Content uploaded by Paweł Pławiak
Author content
All content in this area was uploaded by Paweł Pławiak on May 11, 2019
Content may be subject to copyright.
Content uploaded by Paweł Pławiak
Author content
All content in this area was uploaded by Paweł Pławiak on May 11, 2019
Content may be subject to copyright.
SYSTEMS-LEVEL QUALITY IMPROVEMENT
IAPSO-AIRS: A novel improved machine learning-based system
for wart disease treatment
Moloud Abdar
1
&Vivi Nur Wijayaningrum
2
&Sadiq Hussain
3
&Roohallah Alizadehsani
4
&Pawel Plawiak
5
&
U. Rajendra Acharya
6,7,8
&Vladimir Makarenkov
1
Received: 21 March 2019 /Accepted: 13 May 2019
#Springer Science+Business Media, LLC, part of Springer Nature 2019
Abstract
Wart disease (WD) is a skin illness on the human body which is caused by the human papillomavirus (HPV). This study mainly
concentrates on common and plantar warts. There are various treatment methods for this disease, including the popular immu-
notherapy and cryotherapy methods. Manual evaluation of the WD treatment response is challenging. Furthermore, traditional
machine learning methods are not robust enough in WD classification as they cannot deal effectively with small number of
attributes. This study proposes a new evolutionary-based computer-aided diagnosis (CAD) system using machine learning to
classify the WD treatment response. The main architecture of our CAD system is based on the combination of improved adaptive
particle swarm optimization (IAPSO) algorithm and artificial immune recognition system (AIRS). The cross-validation protocol
was applied to test our machine learning-based classification system, including five different partition protocols (K2, K3, K4, K5
and K10). Our database consisted of 180 records taken from immunotherapy and cryotherapy databases. The best results were
obtained using the K10 protocol that provided the precision, recall, F-measure and accuracy values of 0.8908, 0.8943, 0.8916 and
90%, respectively. Our IAPSO system showed the reliability of 98.68%. It was implemented in Java, while integrated devel-
opment environment (IDE) was implemented using NetBeans. Our encouraging results suggest that the proposed IAPSO-AIRS
system can be employed for the WD management in clinical environment.
Keywords Wart disease .Data mining .Machine learning .Computer-aided diagnosis system .Artificial immune recognition
system .Improved adaptive particle swarm optimization
This article is part of the Topical Collection on Systems-Level Quality
Improvement
*Moloud Abdar
m.abdar1987@gmail.com
Vivi Nur Wijayaningrum
vivinurw@gmail.com
Sadiq Hussain
sadiq@dibru.ac.in
Roohallah Alizadehsani
ralizadehsani@deakin.edu.au
Pawel Plawiak
plawiak@pk.edu.pl
U. Rajendra Acharya
aru@np.edu.sg
Vladimir Makarenkov
makarenkov.vladimir@uqam.ca
1
Département d’Informatique, Université du Québec à Montréal,
Montréal, QC, Canada
2
Department of Information Technology, Politeknik Negeri Malang,
Malang, East Java, Indonesia
3
Dibrugarh University, Dibrugarh, Assam, India
4
Institute for Intelligent Systems Research and Innovations (IISRI),
Deakin University, Geelong, Victoria, Australia
5
Institute of Telecomputing, Faculty of Physics, Mathematics and
Computer Science, Cracow University of Technology,
Krakow, Poland
6
Department of Electronics and Computer Engineering, Ngee Ann
Polytechnic, Singapore, Singapore
7
Department of Biomedical Engineering, School of Science and
Technology, Singapore University of Social Sciences,
Singapore, Singapore
8
Department of Biomedical Engineering, Faculty of Engineering,
University of Malaya, Kuala Lumpur, Malaysia
Journal of Medical Systems (2019) 43:220
https://doi.org/10.1007/s10916-019-1343-0
Introduction
Wart disease (WD) is a skin illness on the human body caused
by the human papillomavirus (HPV). Three main groups of
warts are: epidermodysplasia verruciform is (EV), cutaneous
and mucosal [1]. The cutaneous warts are very general ail-
ments in both adults and children which have significant prev-
alence with 3%to 13% among general population and approx-
imately 33% among primary school children [2,3]. There are
four different types of warts such as common, flat, filiform,
and plantar warts. Usually, common warts can be observed on
the hands, flat warts are frequently observed on the legs and
on the back of the hands [4], while plantar warts are usually
seen on the soles of the feet [5]. This study mainly concen-
trates on common and plantar warts. There are various treat-
ment methods for these types of wart diseases. However, this
study considers two widely popular treatment methods: im-
munotherapy and cryotherapy. Since patients respond differ-
ently to different treatment methods, the physicians are inter-
ested to find the customized treatments. Therefore, it is impor-
tant to characterize the wart disease, and thereby evaluate the
therapy or treatment response on the patients having this dis-
ease. This study tries to provide a novel methodology based
on machine learning (ML) techniques to classify the WD
treatment response. There have been a few studies in the lit-
erature that have investigated the WD treatment response
problem using different ML methods [1,6,7]. These studies
suffer from the following drawbacks: first, the current ML-
methods applied for WD evaluation have low accuracies and
not robust enough to be reliable. Second, the current methods
are not fully automated and as a result, the manual evaluation
of the WD treatment response suffers from subjectivity and
laboriousness. Furthermore, the current methods are biased,
since the sample size is small. Hence, no extensive research
has been done to explore all facets of ML-based strategies to
investigate WD treatment response.
Different types of ML-based methods can be used in
computer-aided diagnosis (CAD) system which has shown to
be effective in providing second opinion to the physicians in
many fields of healthcare [8] such as neurology [9,10], cardi-
ology [11,12], endocrinology [13,14], dermatology [15]and
diabetic retinopathy [16]. This semi-automated detection ap-
proach not only save the lives of the patients but is also more
economical and represents a faster diagnosis process.
It may be noted that due to the importance of ML tech-
niques, it has been successfully applied on various diseases
[16–31]. Akben [1] proposed a new methodology using the
fuzzy informative images and decision trees for the detection
of WD. Author has applied ML methods on both cryotherapy
and immunotherapy data sets and obtained accuracies of
94.40% and 90% for cryotherapy and immunotherapy data
sets, respectively. In another study by Uzun et al. [6]haveused
k-Nearest Neighbors (KNN) algorithm on cryotherapy and
immunotherapy data sets. They have obtained the highest ac-
curacy of 80% for both data sets for K= 7. Khozeimeh et al. [7]
proposed an expert system for selection of wart treatment
methods. In their study, they treated 90 patients with immuno-
therapy and 90 patients with cryotherapy and liquid nitrogen.
They were able to predict the results of the treatment method
with the accuracy of 80% using a fuzzy rule-based system.
Recently, particle swarm optimization (PSO) is one of the
more prominent evolutionary computation methods intro-
duced which offers several advantages such as [32]: (1) it is
an intelligence-based paradigm, (2) does not have mutation
calculation and overlapping issues, and (3) offers simplicity
in calculation. The only challenge is that the PSO was initially
non-adaptive in nature.
Adaptive approaches are well-known techniques used to
improve the performance of conventional methods. These ap-
proaches have been effectively combined with
different methods such as PSO [33] and genetic algorithms
(GAs) [34]. The PSO method includes a collection of several
particles that move in the search space in which each particle
has two characteristics: position and velocity. Generally, the
PSO method updates the position of particles using a personal
best position (pbest) and a global best (gbest). The adaptive
approaches allow a better features search capability over clas-
sical PSO [33]. The adaptive PSO further suffers from
updating the velocity and position of particles, which lowers
the performance of the CAD system. We have removed this
weakness by setting the maximum and minimum velocity
limit of particles and dynamically adjusting the control param-
eters c
1
and c
2
during the solution search process. We call this
as improved adaptive particle swarm optimization (IAPSO)
method.
The PSO algorithm has been extensively used for the diag-
nosis of heart disease [35,36], in health care [37], for
the Parkinson disease detection [38], for the detection of hem-
orrhage [39], in neurological application [40], for the breast
cancer detection [41,42] and for Alzheimer’s disease detection
[43,44]. The adaptive PSO has also been used to attain better
accuracy during the breast cancer diagnosis [45]. The effective-
ness of APSO, which is an optimization method, has been
shown in a variety of applications such as sonar image se-
quences [46], humanoid robots [47] and tandem blade optimi-
zation [48]. The velocity and position of particles have signif-
icant effect on the performance of APSO. The reason for that is
that the position and velocity of each particle is updated during
the iteration so that each particle moves closer to a better solu-
tion. In this study, we update the position and velocity of each
particle to improve the performance of APSO. The proposed
method is called IAPSO (improved APSO).
Our database consisted of 180 records taken from the original
immunotherapy and cryotherapy data equally (90 records for
each of the original data types). To avoid the effect of variation
in the IAPSO method, we first normalized the data set as part of
220 Page 2 of 23 J Med Syst (2019) 43:220
the data preparation and preprocessing step. Cross-validation
(CV) protocol was employed to test our machine learning para-
digm consisting of five partition protocols (K2, K3, K4, K5, and
K10). The CV protocol was evaluated using several performance
metrics that includes: precision, recall, F-measure, and accuracy.
We benchmarked our algorithm against the previously published
method called artificial immune recognition system (AIRS). We
further evaluated our CAD system by swapping the core classi-
fier of IAPSO (with AIRS) by six different classifiers (shown as a
general block in Fig. 1) namely simple AIRS, Bayes network
(BN), Multilayer Perceptron (MLP), J48, Random forest (RF),
and hierarchical LVQ (H-LVQ) to understand the generalization
behavior of the ML system. As part of the performance evalua-
tion of the ML system, we further computed the reliability index
(RI). The system was implemented in the Java, while integrated
development environment (IDE) was implemented using
NetBeans.
The velocity and position of particles in APSO was up-
dated in two major steps. We first set the maximum and min-
imum velocity limits of particles. Then, the values of c
1
and c
2
were adjusted. Therefore, the main contributions of this study
can be listed as follows: (1) it presents a new standard WD
database, (2) it applies different partition protocols from each
method to check the performance of the algorithms, (3) it
improves the performance of classical PSO (IAPSO), and
(4) it describes a novel evolutionary-based approach to im-
prove the performance of AIRS using IAPSO for WD treat-
ment response. Our results demonstrate that IAPSO can im-
prove the performance of the AIRS method on a WD data-
base. Moreover, the proposed methodology outperformed the
five classical methods used for comparison.
The rest of the work is structured into seven main sections
which are as follows. Section BClinical data^describes the clin-
ical data sets used in our research followed by the proposed
methodology in Section BThe proposed methodology^.
Section BIAPSO architecture^provides our proposed methodol-
ogy for classification of the WD treatment response. The
experimental protocols are presented in Section BExperimental
protocols^, followed by the results description in
Section BResults^. The performance evaluation and discussion
are provided in Section BPerformance evaluation^,followedby
Section BDiscussion^. Finally, we have presented the conclu-
sions in Section BConclusion^.
Clinical data
The main data used in this paper comprised of two data sets:
plantar and common warts data. Both of these data sets were
gathered in the dermatology clinic of Ghaem Hospital in
Mashhad, IRAN between the period of January 2013 and
February 2015. These data sets can be downloaded from the
UCI machine learning repository [7,49–51]. The information
from 180 patients with plantar and common warts was record-
ed. Each data set includes 90 records. We believed that the
numbers of records in these data sets were not enough for
optimal performance of the new algorithm. To tackle this chal-
lenge, we introduced a new standard WD database which was
designed by combining both data sets and therefore consisted
of 8 features selected for 180 records. Thus, the new data set:
(1) includes more records; and (2) has more features, com-
pared to the old data sets. The important point about the com-
bined data set is that it has no missing values, which is very
critical for ML techniques. More information about the attri-
butes and the statistical summary of the original and new data
sets are briefly discussed below.
The cryotherapy and immunotherapy data sets
The first data set includes seven features from 90 patients,
collected when the cryotherapy approach was used. The sec-
ond data set includes eight factors gathered from 90 patients
when the immunotherapy approach was used. More informa-
tion about these two data sets can be found in [7,49–51].
Fig. 1 Global architecture of
application of machine learning
algorithms for Wart disease
prediction
J Med Syst (2019) 43:220 Page 3 of 23 220
The combined wart data (CWD)
The original data sets used two different treatment methods.
Thus, we combined both data sets to make a new data set. In
addition, we would argue that these treatment methods can be
considered as a prediction feature for WD. Since the first data
set includes seven factors and the second data set includes eight
factors, we tried to use all common factors available in both
data sets. Hence, the induration diameter of initial test feature
existing in the second data set was removed. Furthermore, we
added a new factor named BProcedure of treatment^as an input
variable with other common factors (see Table 1):
The proposed methodology
This study applied a new machine learning (ML) system on
the WD data set. The proposed ML-based system (named
IAPSO for AIRS) is presented in Fig. 2. The preprocessing
approach was used as an initial step in our proposed method-
ology. In the first step of the study, the min-max normalization
approach was applied. Then, the classical AIRS system was
applied on the data set. In the second step, the performance of
AIRS was optimized using the improved version of the APSO
algorithm (IAPSO). The K-fold cross validation was
employed for training-testing of the system. The general view
of the proposed approach is given in Fig. 2.
Figure 2indicates that, we first applied the data preprocessing
approach using the min_max normalization technique. Different
fold cross validation techniques (K2, K3, K4, K5, and K10) were
used on the training and testing data. The parameters of the
AIRS algorithm were optimized using the IAPSO method.
Figure 3shows the application of PSO and AIRS, in combina-
tion, on the WD data set. We have used an improved version of
PSO (named IAPSO) as one of the widely employed techniques
for optimizing parameters in the AIRS method.
The procedure of IAPSO for AIRS describes the stage of
hybridization between IAPSO and AIRS. This stage starts
with the use of IAPSO by initializing the position and velocity
of particles. Then, to find out the fitness value of each particle,
the AIRS algorithm is executed. The position of particles that
have been initialized in the IAPSO process consists of 7
values which will be used as parameters in the AIRS algo-
rithm. The AIRS algorithm is run until the stop criteria are
met, and then the accuracy value is obtained. The accuracy
value is used as a fitness value of a particle and the IAPSO
process is resumed until the maximum number of iterations is
reached. The whole process of the proposed IAPSO method
for AIRS is shown in Algorithm 1, in Appendix 2.
IAPSO architecture
This study introduces a new diagnosis model for the detection
of WD. Our major hypothesis was validated using one classi-
fier namely, AIRS, one evolutionary algorithm (PSO), one
WD data set (the new one), min-max normalization approach
and five different cross-validation (K= 2, 3, 4, 5, and 10)
protocols. We briefly explained the AIRS, PSO, IAPSO, and
min-max normalization approaches in the following sections.
In this work, we have used 5 well-known ML techniques,
namely NB, MLP, J48, RF, and H-LVQ.
Particle swarm optimization
The Particle Swarm Optimization (PSO) algorithm is a stochas-
tic optimization strategy which cannot make any assumptions
regarding the gradient of the objective function introduced by
Eberhart and Kennedy [52,53]. This method was inspired from
the natural social behaviors and dynamic movements with com-
munications of various birds, insects and fishes. Indeed, the
PSO updates the position of particles using a personal best
position (pbest)andaglobalbest(gbest)[54]. More informa-
tion about PSO is presented in [55–57] and Appendix 3.
The PSO as one of the optimization methods expected to
perform well to improve the quality of solutions. This paper
uses the application of PSO to improve the quality of solutions
provided by AIRS to identify diseases accurately by adding
adaptive mechanisms. In a previous study, an improvised
PSO was carried out by converting particle coordinates when
updating the velocity, which is called Rotated Particle Swarm
(RPS) [58]. The calculation process uses a matrix and hence
Table 1 The combined wart data
set (CWD) No. Feature Range Type
1. Response to treatment Yes or No Output (target)
2. Gender 88 Men and 92 Women Input
3. Age (years) 15–67 Input
4. Time elapsed before treatment (month) 0–12 Input
5. The number of warts 1–19 Input
6. Types of wart (Count) 1–Common (101), Plantar (31), Both (48) Input
7. Surface area of the warts
c
(mm
2
)4–900 Input
8. Procedure of treatment 1 = Immunotherapy and 2 = Cryotherapy Input
220 Page 4 of 23 J Med Syst (2019) 43:220
requires a long computing time. In another study [59], a particle
transfer mechanism was proposed. The transfer mechanism
removes the fitness personal best value obtained previously if
the average value is less than the threshold. In our opinion, this
Wart Data (WD)Protocol Type
Normalized WD Normalized WD
Offline Classifier
Optimization
(IAPSO)
Online
Classifier
Ground Truth
Labels
Offline Training
Parameters
Predicted
Class
Offline Wart Classification System Online Wart Classification System
Classifier
Type
Classifier
Type
Normalization Normalization
Training WD Testing WD
Fig. 2 The general view of the
proposed system
Start
Initialize Velocity and Position of
Particles
For eac h Particle call AIRS() to
Calculate the Fitness Value
Assign the P aramete r Values
Calculate Affinity between Memory
Cell and Training Antigen
Clone the Memory Cell based on its
Affinity and Put them into ARB Pool
Calculate Total Stimulation of ARBs
in each Class
Store the best ARB into Memory Cell
Calculate Distance between Memory
Cell and Test Antigen
Calculate Accuracy
Set Personal Best (Pbest) and
Global Best (Gbest)
Update Velocity and Position of
Particles
Set the Minimum
and Maximum
Veloci ty Limit of
Particles
Make Gbest as a Sol ution
End
AIRS()
Change the Values
of c
1
and c
2
Adaptively
Fig. 3 Illustration of the parameter optimization procedure in the proposed IAPSO-AIRS system: the bolded boxes show how the position and velocity
of each particle is updated
J Med Syst (2019) 43:220 Page 5 of 23 220
approach is good because its solutions can avoid local optimum
solutions and use of mutation processes in genetic algorithms.
However, if the particle transfer is too far away, then the particle
might come out of the search area because of the selection of an
incorrect new solution point. Therefore, our study uses mini-
mum and maximum particle velocity settings to avoid particles
coming out of the search area. In addition, the control param-
eter settings are also added to adjust particle movements adap-
tively according to fitness values so that particles can find better
solutions in each iteration.
Improved PSO
The process of updating the velocity and position of particles
is carried out using Eq. D.1and Eq. D.2. In this study, we
initialized PSO by setting the minimum and maximum veloc-
ity limits for each particle in the search space. The velocity
limits are influenced by the minimum and maximum values of
each parameter used. Calculation of the minimum and maxi-
mum velocity limit of particles is shown in eqs. (1)and(2).
vmaxi¼kxmaxi−xmini
ðÞ
2ð1Þ
vmini¼−vmaxið2Þ
where, kis a random number with a range of values [0, 1]. In
this study, the value of kequal to 0.6 was used, vmax
i
is the
maximum velocity limit of a particle in the search space, v
min
i
is the minimum velocity limit of a particle in the search
space, x max
i
is the maximum value limit ofa parameter, and x
min
i
is the minimum value limit of a parameter.
The improvement of PSO was also carried out by applying
random injection mechanism each iteration in multiples of 3.
The mechanism was performed by generating random parti-
cles which then were inserted into the population. The purpose
of the random injection mechanism is to prevent search results
from being trapped at the local optimum solution.
Adaptive PSO
The performance of PSO is strongly influenced by the balance
of social and cognitive parameters. Determination of the pa-
rameter values c
1
and c
2
is an important factor in determining
the balance parameters to produce optimal solutions.
Parameters that are not well determined will affect the perfor-
mance of PSO and lead to a premature convergence.
PSO as a population-based search algorithm which is often
trapped in a local optimum solution. This is because there are
many particles that move freely in the search space to find a
solution. The poor settings of the PSO parameters used, such
as particle starting points, particle velocity, and control param-
eters that regulate the movement of particles, causes particles
to move inefficiently.
The use of adaptive mechanism in the PSO is expected
to regulate particle movements during the solution search
process. The control parameters in this case are the fitness
values that regulate particle movements and can be
changed adaptively at certain iterations by adjusting to
the quality of solutions. If most of the particles in the
search space have a fitness value greater than the fitness
value obtained in the previous iteration, then the value of
the cognitive parameters must be increased for these par-
ticles to move faster in order to get closer to the personal
best of each particle. Conversely, if the particles in the
search space tend to produce a fitness value that is worse
than the fitness value in the previous iteration, it means
that the control parameter is not able to direct the particle
to a better solution so that the cognitive parameter value
Table 3 Initial results obtained using the classical AIRS algorithm
applied on the new WD dataset with 180 records
Class PRE REC FM ACC (%)
K2 Success 0.840 0.882 0.861 –
Failure 0.745 0.672 0.707 –
Weighted Average 0.808 0.811 0.809 81.11
K3 Success 0.824 0.824 0.824 –
Failure 0.656 0.656 0.656 –
Weighted Average 0.767 0.767 0.767 76.66
K4Success 0.882 0.815 0.847 –
Failure 0.686 0.787 0.733 –
Weighted Average 0.815 0.806 0.808 80.55
K5 Success 0.840 0.882 0.861 –
Failure 0.745 0.672 0.707 –
Weighted Average 0.808 0.811 0.809 81.11
K10 Success 0.825 0.874 0.849 –
Failure 0.722 0.639 0.678 –
Weighted Average 0.790 0.794 0.791 79.44
Bold values indicate the highest classification metrics
Table 2 The summary of training for the AIRS algorithm
Parameter Value
Threshold of Affinity 0.482
Total instances for training 180
Total replacements of memory cell 69
Mean ARB clones for each refinement iteration 94.716
Mean total resources for each refinement iteration 144.319
Mean pool size for each refinement iteration 112.225
Mean memory cell clones for each antigen 16.8
Mean ARB refinement iterations for each antigen 2.094
Mean ARB prunings for each refinement iteration 103.17
220 Page 6 of 23 J Med Syst (2019) 43:220
must be decreased. In addition, if the best particle in the
current iteration has a fitness value greater than the fitness
value obtained in the previous iteration, then the value of
the social parameter needs to be added so that the particle
can move faster in order to get closer to a better solution.
Conversely, if the fitness value obtained at this time is
smaller than the fitness value in the previous iteration,
then the particle will be directed to do a search around
the point of the solution that has been obtained previous-
ly. Thus, the particles in the search space will make dif-
ferent adjustments to the movement so that the solution
obtained is not a local optimum solution [60].
This study used an adaptive PSO mechanism to dy-
namically adjust the control parameters c
1
and c
2
during
the search process. The adaptive PSO mechanism is based
on the fitness value of each particle in each iteration, so
that the values of c
1
and c
2
change adaptively. Eq. (3)is
used to update c
1
and c
2
adaptively:
c0
1¼c1þ0:05 and c00
1¼c1−0:05
c0
2¼c2þ0:05 and c00
2¼c2−0:05
ð3Þ
where c0
1is used to update the value of c
1
when more
than 20% of particles in the population have a new Pbest
value that is better than the Pbest value in the previous
iteration. If the number of particles that have a new Pbest
value is better than the Pbest value in the previous itera-
tion (should be less than 20%), then the value of c
1
will
be reduced using c00
1.Thec0
2is used to update the value of
c
2
when the Gbest value in the current iteration is better
than the Gbest value in the previous iteration. Conversely,
if the Gbest value in the current iteration does not in-
crease, then the value of c
2
will be reduced using c00
2.
Min-max normalization
The min-max normalization approach executes a linear
transformation on the principal data [61]. This technique
is a way to map a value dof Pto d′in the range
[new_min(p), new_max(p)]. Basically, the min-max nor-
malization is computed using d′,wheremin(p) is the min-
imum value of the feature and max(p) is the maximum
value of the feature. It should be expressed that the min-
max normalization approach maintains the relation among
the principal data values.
Artificial immune recognition system (AIRS)
Individual bodies have a special kind of defense system
called the Bimmune system^[62]. This system protects
our body from various diseases caused from pathogens,
germs, and other toxic substances. The artificial immune
system (AIS) algorithm is a computationally intelligent
and rule-based machine learning method that has been
widely applied to solve various complex computational or
engineering problems (e. g., pattern recognition) [63–67].
The initial AIS was developed from the theoretical immu-
nology field in mid-1980. In other words, AIS is a learning
algorithm that was inspired from the human being immune
system [68]. This study uses the AIRS algorithm, which is a
kind of supervised learning of AIS [69]. Bai [32] showed
that there is a mapping between the immune system and
AIRS, which is given in [32,62–69].
Classical classification techniques
As we discussed earlier, this work compares the perfor-
mances of the proposed method with five well-known ML
85.56
85.83
86.94
87.41
89.11 89.07 89.17 89.44 89.17 89.44
84.50
85.50
86.50
87.50
88.50
89.50
90.50
5 101520253035404550
eulaV ssentiF egarevA
The Number of Particles
Test Result for the Number of Particles
Fig. 4 Test results for the number
of particles
J Med Syst (2019) 43:220 Page 7 of 23 220
algorithms. The algorithms used are as follows: Bayes
network (BN), multilayer perceptron (MLP), J48, random
forest (RF), and hierarchical LVQ (H-LVQ). More infor-
mation regarding each algorithm can be found in [70–79].
Experimental protocols
In the present work, we have applied data pre-processing
(min-max normalization), one classical algorithm (AIRS),
and one evolutionary technique (IAPSO). We have carried
out three experimental protocols, including (i) applying tradi-
tional AIRS on the wart data, (ii) optimization of parameters in
AIRS using IAPSO, and (iii) discussion and comparison of
the proposed method with other classifiers.
Experimental protocol 1: AIRS
The major objective of this section is to assess the per-
formance of the classical AIRS method on WD and
CWD data sets. In the first experiment, the AIRS algo-
rithm was applied. The summary of training is shown in
Tab le 2:
Due to the randomization of the data using the K-
fold cross validation, the second experiment was repeat-
ed 20 times (with T = 20 trials). It should be noted that,
in this step, we ran the method 20 times and then all
metrics were averaged. For this step, Weka 3.8 was
used as an open source data mining tool which was
carried out on a PC computer equipped with Windows
7, Pentium Dual Core CPU E5300 @2.60 GHz and
2GBRAM.
87.78
88.33
88.61
89.63
89.17
87.50
87.00
87.50
88.00
88.50
89.00
89.50
90.00
0.4 0.5 0.6 0.7 0.8 0.9
e
u
laV ss
e
nt
iF
egarevA
The Value of Inertia Weight
Test Result for Inertia Weight
Fig. 6 Test results for the value of
inertia weight
85.93
86.39
87.64
88.89 89.07 89.17 89.44 89.17 89.17 89.44
84.50
85.50
86.50
87.50
88.50
89.50
90.50
5 101520253035404550
eulaV sentiF egarev
A
The Number of Iterations
Test Result for the Number of Iterations
Fig. 5 Test results for the optimal
number of iterations
220 Page 8 of 23 J Med Syst (2019) 43:220
Experimental protocol 2: IAPSO for AIRS
The second experiment (optimization) was implemented in
Java Version: 1.8.0, when NetBeans 8.0.2 was used as IDE.
Here, we used a Windows 10 operating system with Intel i5
2.4GHz CPU and 8.00 GB RAM. Several tests were per-
formed to determine the best parameters to be used in the
IAPSO for AIRS algorithm. The testing phase consisted of
several testing sets and different parameter vcombinations,
including the number of particles, the number of iterations,
the value of inertia weight, and the values of c
1
and c
2
.
Testing the number of particles was used to determine the
number of particles necessary to produce the optimal solution.
The number of particles used in this test scenario ranged from
5 to 50. The number of iteration used was 10, the weight of
inertia 0.5, and the values of c
1
and c
2
2 for both of them.
Experimental protocol 3: IAPSO for AIRS on different
partition protocols
The evaluation of the performance of the proposed methodol-
ogy for different partition protocols was the main objective of
this section. Here, the proposed methodology was applied on
the CWD data set with different partition protocols (K2, K3,
K4, K5, and K10) and then several performance metrics such
as precision (PRE), recall (REC), F-measure (FM), and accu-
racy (ACC), were evaluated.
Results
Results of experimental protocol 1: AIRS
In this work, we first applied the classical AISR method
on the CWD and the original immunotherapy and
cryotherapy data (90 records for each of the original data
types), separately. It should be noted that for each parti-
tion protocol (K=2,3,4,5,and10),AIRSwasrepeated
20 times and then the averages of various measures were
presented. The performance of AIRS was investigated
using PRE, REC, FM, and ACC for two different classes
Table 4 Results of the classification (IAPSO for AIRS) using optimal
parameters and K10 applied on the new data set
Trial Number PRE REC FM ACC (%)
1 0.8859 0.8762 0.8807 89.44
2 0.8826 0.8833 0.8814 88.33
3 0.8782 0.8720 0.8750 88.88
40.8908 0.8844 0.8875 90.00
5 0.8898 0.8735 0.8807 89.44
6 0.8839 0.8865 0.8852 89.44
7 0.8762 0.8815 0.8787 88.88
8 0.8787 0.8787 0.8787 88.88
90.88910.8943 0.8916 90.00
10 0.8882 0.8937 0.8908 90.00
11 0.8806 0.8888 0.8844 89.44
12 0.8787 0.8787 0.8787 88.88
13 0.8830 0.8858 0.8844 89.44
14 0.8752 0.8808 0.8778 88.88
15 0.8771 0.8822 0.8795 88.88
16 0.8847 0.8872 0.8859 89.44
17 0.8882 0.8937 0.8908 90.00
18 0.8898 0.8735 0.8807 89.44
19 0.8830 0.8858 0.8844 89.44
20 0.8806 0.8888 0.8844 89.44
Ave ra ge 0.8832 0.8835 0.8831 89.33
Bold values indicate the highest values of classification metrics
86.94
87.59
89.26
88.15
87.50
86.00
87.00
88.00
89.00
90.00
0; 4 1; 3 2; 2 3; 1 4; 0
eulaV ssentiF egare
vA
c1; c2
Test Results for c1and c2
Fig. 7 Test results for the values
of c
1
and c
2
J Med Syst (2019) 43:220 Page 9 of 23 220
0.81
0.82
0.83
0.84
0.85
0.86
0.87
0.88
0.89
0.9
PRE
K2 K3 K4 K5 K10
ecn
a
mrof
r
eP (%)
Partition
Protocol
a
0.81
0.82
0.83
0.84
0.85
0.86
0.87
0.88
0.89
0.9
REC
K2 K3 K4 K5 K10
ec
n
a
m
ro
fre
P(%)
Partition Protocol
b
0.81
0.82
0.83
0.84
0.85
0.86
0.87
0.88
0.89
0.9
FM
K2 K3 K4 K5 K10
ecnamr
ofr
eP (%)
Partition Protocol
c
83
84
85
86
87
88
89
90
91
ACC
K2 K3 K4 K5 K10
ecn
a
mrof
reP
(%)
Partition Protocol
d
Fig. 8 Results of the
classification obtained by
the IAPSO method using optimal
parameters and differentprotocols
(K=2,3,4,5,and10)appliedthe
CWD data set: a:PRE,bREC, c
FM, and d: ACC. Note: Trial No.
21 is the average of all of
the previous 20 trials
220 Page 10 of 23 J Med Syst (2019) 43:220
(success and failure). In the first step, AIRS was applied
on separate data sets to check the performance of the
algorithm (see Table 10 in Appendix 4). It can be noted
from Table 10 that AIRS has yielded 91.11% (K4) and
82.22% (K2) for the cryotherapy and immunotherapy data
sets, respectively. This means that the AIRS algorithm has
provided comparable results according to other published
works [7]. The results obtained using the classical AIRS
algorithm on the CWD data set are presented in Table 3.
It can be noted from Table 3that classical AIRS has
provided a better performance for the Bsuccess^class and
a weaker performance for the Bfailure^class. In other
words, AIRS showed good classification results for only
one class. Therefore, we tried to enhance the performance
of AIRS using the IAPSO technique. In the rest of the
study, we considered the weighted average of each proto-
col for the AIRS method.
Results of experimental protocol 2: IAPSO
In this experiment, the parameters of AIRS were optimized
using the proposed methodology (IAPSO). To get the best
values of our parameters, the proposed methodology was test-
ed for the number of particles, the number of iterations, the
value of inertia weight, and finally the values of c
1
and c
2
.The
test results for the number of particles are shown in Fig. 4.
It can be seen from Fig. 4that the average fitness value
increases with an increase in the number of particles.
However, upon reaching a certain point, the fitness value stops
augmenting. The fitness values tend to be stable when the
number of particles is above 25, which means that 25 is the
optimal number of particles. The second tests was for the
number of iterations that necessary to produce the optimal
solution. The number of iterations used in this test scenario
ranged from 5 to 50. The number of particles used was 25, the
Fig. 9 Comparison of performances of different ML algorithms in terms of ‘PRE’for the CWD data set
Fig. 10 Comparison of
performances of
different ML algorithms in terms
of ‘REC’for the CWD data set
J Med Syst (2019) 43:220 Page 11 of 23 220
weight of inertia was 0.5, and the values of c
1
and c
2
were 2.
Test results for the number of iterations are shown in Fig. 5.
In Fig. 5, it can be seen that the average fitness value
increases when the number of iterations increases. This
increase in fitness value occurs for iterations 5 to 20 but
tends to be stable when the number of iterations is above
20. This condition indicates that the number of optimal
iterations is 20. In the third step, the value of inertia
weight was tested. This valueisusedtodeterminethe
amount of inertia weights needed to produce the optimal
solution. The values of inertia weights used in this test
scenario were0.4, 0.5, 0.6, 0.7, 0.8, and 0.9. The number
of particles used was 25, the number of iterations was 20,
and the values of c
1
and c
2
were 2. Test results of this test
areshowninFig.6.
Figure 6shows that the best value of inertia is equal to
0.7 and the average fitness value decrease when larger
values of inertia weight were applied. Larger values of
inertia weight make the exploration process faster by
searching globally. Hence, it gets stuck into a local opti-
mum solution. Smaller values of inertia weights make the
particles more likely to perform local searching or exploi-
tation [72]. In this study, the best found value of inertia
weight was 0.7. The last test was for the values of c
1
and
c
2
. The values of c
1
used in this test scenario were 0, 1, 2,
3, and 4 with the condition c
1
+c
2
= 4. The number of
particles used was 25, the number of iterations was 20,
and the value of inertia weight was 0.7. The test results
for the values of c
1
and c
2
are shown in Fig. 7.
Figure 7shows that the best parameters used to gen-
erate the optimal solution are when the values of c
1
and
c
2
are 2. With a smaller value of c
1
, the convergence to a
certainsolution will occur quickly so that it has a tenden-
cy to produce a local optimal solution because the
Fig. 11 Comparison of performances of different ML algorithms in terms of ‘FM’for the CWD data set
Fig. 12 Comparison of
performances of different
ML algorithms in terms of ‘ACC’
for the CWD data set
220 Page 12 of 23 J Med Syst (2019) 43:220
process of updating leads to a social component.
Conversely, with a smaller value of c
2
, fewer particles
interact with each other because the process of updating
leads to a cognitive component of each particle. This
causes the particles to remain in the area around their
initial position [80]. Based on the test results shown in
Figs. 4,5,6and 7, it can be seen that the optimal pa-
rameters of IAPSO are as follows: the number of parti-
cles is 25, the number of iterations is 20, the value of the
inertia weight is 0.7, the value of c
1
is 2, and the value
of c
2
is also 2.
Tab le 4shows an example of the classification of WD
using the optimal parameters obtained from the test re-
sults.SinceAPSOisastochasticalgorithm,theexperi-
ment was performed 20 times to get the average values of
PRE, REC, FM, and ACC. It should be noted that K10
helped provide the best average accuracy.
It can be noted from Table 4that the best accuracy is 90%
obtained using the proposed methodology with K10.
Moreover, the average accuracy in Table 4is 89.33%.
Results of experimental protocol 3: IAPSO on different
partition protocols
The third experiment was conducted to evaluate the perfor-
mance of the proposed methodology with different partition
protocols (K2, K3, K4, K5, and K10). Therefore, AIPSO-
AISR was applied with these five protocols on the CWD data.
Test results for the inertia weights are shown in Fig. 8.
Fig. 8shows that K10 has indicated the best average
accuracy compared to other four protocols with an aver-
age accuracy of 89.33%. In addition, K10 has
provided the highest average value of FM as compared
to the other four protocols with the average of 88.3%.
Moreover, we applied 5 classical ML techniques (BN,
MLP, J48, RF, and H-LVQ) on the new data set to check
the performance of the proposed methodology. It may be
noted that for each of these classifiers, the highest value
of each metric was selected. The results are illustrated in
Figs. 9,10,11,and12.
It can be noted from Figs. 9-12 that the proposed
IAPSO for AIRS algorithm has yielded the highest PRE,
REC, FM, and ACC rates as compared to classical clas-
sifiers. Thus, we could argue that the IAPSO for AIRS
algorithm is very effective for an early diagnosis of WD.
Performance evaluation
Benchmarking improved APSO against AIRS
Since the improvement of AIRS was one of the main objec-
tives of this study, we compared the performance of classical
AIRS and optimized AIRS using the proposed
IAPSO evolutionary method. Thus, the obtained resultsin
first, second, and third experiments were compared. The im-
provement obtained using IAPSO with AIRS isimportant (see
Table 5 Performance analysis of AIRS versus IAPSO for AIRS
Protocol Improvement
PRE REC FM ACC (%)
K20.068↑
*
0.065 ↑0.065 ↑7.22 ↑
K30.11 ↑0.11 ↑0.11 ↑11.67 ↑
K40.059↑0.066 ↑0.062 ↑6.67 ↑
K50.085↑0.082 ↑0.084 ↑8.89 ↑
K10 0.10 ↑0.10 ↑0.10 ↑8.89 ↑
Average 0.0844 ↑0.0846 ↑0.0842 ↑8.668↑
*
Note: ↑means that the value was improved after using the proposed
model
Bold values indicate the highest values of classification metrics
Fig. 13 Bar chart demonstrating
the mean classifier accuracies for
different classifiers (for K = 5,
T = 20)
J Med Syst (2019) 43:220 Page 13 of 23 220
Table 5), which is a key point for medical and healthcare
subjects.
As Table 5shows, the accuracy of IAPSO with AIRS is
much higher than that of the traditional AIRS algorithm for all
protocols (K2,K3,K4,K5,andK10). Moreover, it can be seen
that not only the accuracy values, but those of the other met-
rics (PRE, REC, and FM) have also improved. These results
demonstrate that K3 achieved the maximum optimization in
terms of all metrics (PRE, REC, FM, and ACC), followed by
K10.
Moreover, the system classifier accuracy (η(c)) calculation
for all parameters was evaluated by using all five sets of pro-
tocols and all trial sets for each classifier. The formula for η(c)
is given by Eq. 4:
91
92
93
94
95
96
97
98
99
AIRS for
Immunotherapy
AIRS for
Cryotherapy
AIRS for
combined data
IAPSO-AIRS for
Immunotherapy
IAPSO-AIRS for
Cryotherapy
IAPSO-AIRS for
combined data
94.32
95.58
97.9
98.87 99 98.68
Reliability index (RI)%
dlohserhT
Fig. 14 Comparison of all
classifiers over different data sets
based on RI for K=2, 3 4, 5, and
10 and T=20
Table 6 Comparison of performances of the proposed method against previous studies using the same WD data
Author and Year Data type / Data Size # Selected
Features (input)
Partition type Best Classifier ACC (%)
Akben [1] (2018) cryotherapy / 90 6 10-fold The fuzzy informative
images + decision trees
94.40
Akben [1] (2018) immunotherapy / 90 7 10-fold The fuzzy informative
images + decision trees
90.00
Uzun et al. [6] (2018) cryotherapy / 90 6 NA KNN (K=7) 80.00
Uzun et al. [6] (2018) immunotherapy / 90 7 NA KNN (K=7) 80.00
Khozeimeh et al. [7] (2017) cryotherapy / 90 6 10-fold Fuzzy Logic 80.70
Khozeimeh et al. [7] (2017) immunotherapy / 90 7 10-fold Fuzzy Logic 83.33
Guimarães et al. [81] (2018) immunotherapy / 90 7 70% for training
and 30% for testing
Fuzzy neural networks 83.33
Alizadehsani et al. [82] (2018) cryotherapy / 90 4 NA LIBSVM 91.11 ± 6.67
Alizadehsani et al. [82] (2018) immunotherapy / 90 5 NA LIBSVM 88.89 ± 6.33
Nugroho et al. [83] (2018) Merged data / 180 6 90% for training
and 10% for testing
C4.5 + RFFW 87.22
Jia et al. [84] (2019) immunotherapy / 90 7 10-fold LDA+ AFS + C4.5Tree 80.73 ± 0.0262
Junio Guimarães et al. [85] (2019) cryotherapy / 90 6 10-fold Fuzzy neural network
(FNN)
88.64
Junio Guimarães et al. [85] (2019) immunotherapy / 90 7 10-fold Fuzzy neural network
(FNN)
84.32
Proposed Method cryotherapy / 90 6 5-fold (K5) IAPSO + AIRS 94.44
Proposed Method immunotherapy / 90 7 2-fold (K2) IAPSO + AIRS 84.44
Proposed Method CWD / 180 7 10-fold (K10) IAPSO + AIRS 90.00
average (89.33)
220 Page 14 of 23 J Med Syst (2019) 43:220
ηcðÞ¼∑k¼K
k¼1∑t¼T
t¼1ηc;k;t
ðÞ
KTð4Þ
The results are illustrated in Fig. 13.
Reliability analysis
The reliability index (RI) of the CAD/ML system helps to
evaluate the performance of a CAD/ML system. RI is the ratio
of the standard deviation of the classification accuracy and the
mean of the classification accuracies for all protocols. The
reliability index (named ξ
N
) of the system is computed using
Eq. 5:
ξN%ðÞ¼ 1−σN
μN
100 ð5Þ
where σ
N
is the population standard deviation and μ
N
is the
mean of all accuracies for K= 2, 3 4, 5, and 10 and T=20.
Note that the population standard deviation can be calculated
as follows:
σ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
N∑
N
i¼1
xi−μðÞ
2
sð6Þ
where x
i
represents an individual value, μrepresents the
mean/expected value, and Nrepresents the total number
of values. Here, we consider all accuracies obtained by
different protocols for each method. For example, we
carried out IAPSO with AIRS 20 times for each protocol
(K= 2, 3 4, 5, and 10) on the new WD data set. Then, all
accuracies were used to calculate RI (100 accuracies in
total). The results are presented in Fig. 14.
As Fig. 14 shows, the RI values for the proposed meth-
od applied on the new WD data set is 98.68%, while the
accuracy for the immunotherapy and cryotherapy data sets
are 98.87% and 99%, respectively.
Discussion
Claims
This study proposed a novel evolutionary-based system
to classify the WD treatment response. The proposed
methodology was applied with AIRS classifier. It can
be seen from Table 5that our IAPSO method significant-
ly improved the performance of traditional AIRS classi-
fier in the classification of the WD treatment response.
Khozeimeh et al. [7] used the same data sets; however,
they applied their methods on each data set separately. We
combined two original data sets and created a new data set
with more records, as suggested in [7], to apply various
machine learning algorithms. Uzun et al. [6] investigated
the same data set in their study and emphasized the need
to have a bigger data set. Our model showed valuable
performance results for early diagnosis of WD. The risk
stratification can be improved further by adding the path-
ophysiological factors.
Benchmarking
Several prediction studies were carried out on the clas-
sification and diagnosis of WD. Their results are shown
in Table 6.We compared our results with previous stud-
ies in the literature (see Table 6). In this regard, differ-
ent aspects of each paper are considered: (i) type of
data used in each paper, (ii) the size of data, (iii) the
number of various features used during design of the
ML system, (iv) partition type of each data, (v) type
of the algorithm applied during the training/testing
steps, and (vi) performance measures (metrics).
In Table 6, all reported studies have used original 90
records. Hence, we applied our new method on the same
original data to compare our performance with prior stud-
ies. We have obtained the highest performance using the
combination of the IAPSO and AISR methods for cryo-
therapy data with the accuracy of 94.44% (K5), while our
methods yielded the second highest accuracy of 84.44%
(K2) for immunotherapy data. In other words, our new
method had the best average accuracy. Also, we have
obtained the accuracy of 90% using our IAPSO-AISR
method with 180 data set.
A special note on classifiers
The proposed classifier is the main part of the proposed
CAD system. We have selected six different algorithms
includingAIRS,BN,MLP,J48,RF,andH-LVQthat
have been tested using five different partitioning proto-
cols (K2, K3, K4, K5, and K10). The main objective was
to investigate the performance of AIRS and then com-
pare it with other selected classifiers. To optimize the
performance of AIRS, we have used an improved adap-
tive PSO (IAPSO) technique, added random injection,
and applied the hybridization mechanism with simulated
annealing. The proposed method showed the best
J Med Syst (2019) 43:220 Page 15 of 23 220
performance compared to AIRS and other traditional
classifiers. We have observed that the protocol K10 ap-
plied on the new data has provided the best perfomances
in many cases. Our findings indicate that the IAPSO for
AIRS method was the best classifier followed by RF,
while the MLP and J48 classifiers have performed
averagely.
Strengths, weaknesses and extensions: perspectives
for future work
This research introduced the risk stratification system,
calledIAPSOforAIRS,whichwasshowntoclassify
accurately a group of 180 WD patients affected by two
types of warts: common and plantar warts. Current re-
search presents an evolutionary-based technique that
provides an average classification accuracy of 89.33%
when 10-fold cross validation is used (see Table 4).
Nevertheless, the proposed ML-based system still has
potential for optimization. Furthermore, many other
classification techniques can be used instead of AIRS.
In this study only PSO was used instead of other evo-
lutionary algorithms (EAs). Hence, in our future work,
we aim to use other EAs, such as GA [86], genetic
programming, covariance matrix adaptation evolution
strategy (CMA-ES), and cellular evolutionary algorithm.
Moreover, we aim to apply different ensemble tech-
niques (e. g., bootstrap aggregating (bagging), boosting,
stacking, voting, Bayesian model combination) either
with AIRS or even with the IAPSO for AIRS algorithm.
In the future, we also intend to apply different types of
fuzzy logic methods, such as type-2 fuzzy logic, learn-
ing vector quantization (LVQ) technique in our pro-
posed methodology [87–91], genetic ensembles of clas-
sifier [92] as well as different levels of genetic optimiz-
er [93].
Conclusion
An accurate evaluation of treatment response to wart
disease (WD) is a challenging task for physicians. In
the current study, we investigated the treatment response
in immunotherapy and cryotherapy: two popular WD
treatment methods. We combined the two original wart
data sets and created a new data set (each had 90 re-
cords), comprising 180 records. The proposed IAPSO
method returned the precision, recall, F-measure and
accuracy values of 0.8908, 0.8943, 0.8916 and 90%,
respectively, using the K10 protocol. The average im-
provement of precision, recall, F-measure and accuracy
over the artificial immune recognition system (AIRS)
were 8.44%,8.46%,8.42% and 8.668%,respectively.
In the future, the described IAPSO system can be fur-
ther improved using deep learning approach if more
WD-related data become available. We also intend to
improve the performance of the proposed method by
testing other optimization and evolutionary techniques,
such as genetic algorithms, memetic algorithms, ant col-
ony optimization, bees algorithm, artificial bee colony
algorithm, and Cuckoo search using huge data.
Compliance with ethical standards
Conflict of interest Author Moloud Abdar declares that he has no
conflict of interest. Author Vivi Nur Wijayaningrum declares that
she has no conflict of interest. Author Sadiq Hussain declares that
he has no conflict of interest. Author Roohallah Alizadehsani
declares that he has no conflict of interest. Author Pawel
Plawiak declares that he has no conflict of interest. Author U
Rajendra Acharya declares that he has no conflict of interest.
Author Vladimir Makarenkov declares that he has no conflict of
interest.
Ethical approval We used two secondary datasets taken from the
UCI public website (http://archive.ics.uci.edu/ml/datasets/
Immunotherapy+Dataset)and(http://archive.ics.uci.edu/ml/
datasets/Cryotherapy+Dataset+). No ethics approval is required for
these datasets.
Animal studies This article does not contain any studies with animals
performed by any of the authors.
Informed consent Informed consent was obtained from all individual
participants included in the study.
220 Page 16 of 23 J Med Syst (2019) 43:220
Appendix 1
List of Abbreviations/Symbols
Table 7 Shows the list of all
Abbreviations/Symbols used in
this study
SN Abbreviations/Symbols Description
1 ML Machine Learning
2 DM Data Mining
3 IDE Integrated Development Environment
4 C Totalnumberofclassifiers(7)
5KTotal number of partition protocols (5)
6TTotalnumberoftrials(20)
7K2 Protocol of partition (1/2 samples for training and 1/2 for testing)
8K3 Protocol of partition (2/3 samples for training and 1/3 for testing)
9K4 Protocol of partition (3/4 samples for training and 1/4 for testing)
10 K5 Protocol of partition (4/5 samples for training and 1/5 for testing)
11 K10 Protocol of partition (9/10 samples for training and 1/10 for testing)
12 AIRS Artificial Immune Recognition System
13 IAPSO Improved Adaptive Particle Swarm Optimization
14 ACC Accuracy
15 PRE Precision
16 REC Recall
17 FM F-Measure (F
1
Score)
18 N Total data size for the data set
19 k-NN k-Nearest Neighbors
20 BN Bayes Network
21 MLP Multilayer Perceptron
22 J48 J48 Classifier
23 RF Random Forest
24 H-LVQ Hierarchical LVQ
25 CAD Computer-Aided Diagnosis
26 CDSS Clinical Decision Support System
27 CDSA Clinical Decision Support Algorithm
28 UCI University of California, Irvine
29 AT Affinity Threshold
30 RI Reliability index
31 ξ
N
Reliability index for data set
32 σThe population standard deviation
33 μThe mean of all accuracies for all protocols
34 ag The antigen
35 MC The memory cell set
36 ClonalRate The number of mutated clones
37 HyperClonalRate The number of mutated clones a memory cell
38 maxstim The maximum stimulation value
39 minstim The minimum stimulation value
40 min(p) The minimum value of a feature
41 max(p) The maximum value of a feature
42 v max
i
The maximum velocity limit of a particle in the search space
43 v min
i
The minimum velocity limit of a particle in the search space
44 x max
i
The maximum value limit of a parameter
45 x min
i
The minimum value limit of a parameter
46 AFF Affinity
47 EUC-DIS Euclidean Distance
48 AT A delegate of the average affinity over whole training data set
49 STI Stimulation
50 NC Number of Clonal
51 CR Clonal Rate
52 HCR Hyper Clonal Rate
53 ab.r The constant clonal rate (ab.resource)
54 ARGM Argmax
54 d′The mapped value of d
J Med Syst (2019) 43:220 Page 17 of 23 220
Appendix 2
Flowchart of the proposed IAPSO for AIRS
Algorithm 1: IAPSO for AIRS
Inputs:
Wart disease data set
Outputs:
E –evaluation coefficients and confusion matrix
P –classification of wart disease (WD) treatment response
Procedure of IAPSO
For each particle
Initialize particle position randomly
Initialize particle velocity
End
Do
For each particle
Calculate fitness value using Procedure of AIRS
If the fitness value > the personal best fitness value (Pbest)
Set current value as the new Pbest
End if
End for
Choose a particle with the best fitness value of all the particles as Gbest
For each particle
Update particle velocity by calculating the minimum and maximum velocity limit
Update particle position
End for
If 20% Pbest in the current iteration >Pbest in the previous iteration
Increase the value of c1
Else
Decrease the value of c1
End if
If Gbest in the current iteration >Gbest in the previous iteration
Increase the value of c2
Else
Decrease the value of c2
End if
If iteration % 3 = 0
Generate a new particle with random position and velocity
End if
While maximum iteration is not attained
Procedure of AIRS
For each antigen
Select memory cell that has highest affinity to antigen from memory cell pool
Create a pool of B-cells, which consists of the offspring of the selected memory cell
Do
Clone and mutate most highly stimulated B-cell
Remove least stimulated B-cells
While stimulation level > threshold
If affinity of the best B-cell > affinity of best memory cell
Add the best B-cell to the memory pool
Remove the memory cell from the memory pool
End if
End for
220 Page 18 of 23 J Med Syst (2019) 43:220
Appendix 3
Initialization of particle position
In this study, a particle consists of 9 dimensions that
describetheparameterstobeusedintheAIRSalgo-
rithm. These values are randomly selected in different
value ranges for each parameter. The range of values for
each parameter used in this study are presented in
Tab le 8:
An example of particle initialization is shown in Table 9.
Table 9shows that a particle consists of 9 dimensions:
affinity threshold scalar, clonal rate, hypermutation rate, mu-
tation rate, total resources, stimulation threshold, ARB cell
pool size, memory cell pool size, and k-NN with different
values. These values are then used as parameters in AIRS.
Fitness value
The fitness function is used to find out how well the
position has been found by each particle. A particle that
has a high fitness value indicates that the position of the
particle is close to the optimal solution. The higher the
fitness value of a particle, the closer the particle posi-
tion to the optimal solution. The fitness value can be
obtained after applying the values of a particle to the
parameters used in AIRS. AIRS will conduct a training
process using these parameters. After the training pro-
cess is done, the test process will be performed by
calculating the suitability of the classification result with
the actual class data. Therefore, the fitness value will be
calculated using the accuracy, as shown in ACC.
Moreover, the performances of the methods are evaluat-
ed using the other metrics such as precision, recall, and
F-Measure, which are indicated as PRE, REC, and FM.
ACC ¼TP þTN
TP þFP þFN þTN &PRE ¼TP
TP þFP
REC ¼TP
TP þFN &FM ¼2PRE REC
PRE þREC
8
>
<
>
:
where, True Positive (TP) is the value obtained when
the actual class of the data was 1 (Success) and the
predicted was also 1 (Success), True Negative (TN) is
the value obtained when the actual class of the data was
0 (Failed) and the predicted was 0 (Failed), False
Positive (FP) is the value obtained when the actual class
of the data was 0 (Failed) and the predicted was 1
(Success), False Negative (FN) is the value obtained
when the actual class of the data was 1 (Success) and
the predicted was 0 (Failed).
Table 8 Mapping between the immune system and AIRS
Parameter Range
Affinity Threshold Scalar (ATS) [0, 1]
Clonal rate [1, 19]
Hypermutation rate [1, 19]
Mutation rate [0, 1]
Total resources [150, 300]
Stimulation threshold [0, 1]
ARB cell pool size [0, 10]
Memory cell pool size [0, 10]
k-NN [1, 7]
Table 9 Example of a particle
ATS Clonal
Rate
Hypermutation
Rate
Mutation
Rate
Tot al
Resources
Stimulation
Threshold
ARB Cell
Pool Size
Memory Cell
Pool Size
k-NN
0.1545 10.0 18.0 0.2398 112.0 0.0709 3 5 1
J Med Syst (2019) 43:220 Page 19 of 23 220
Appendix 4
References
1. Akben, S. B., Predicting the success of wart treatment methods
using decision tree based fuzzy informative images.
Biocybernetics and Biomedical Engineering., 2018.
2. Kyriakis, K. P., Pagana, G., Michailides, C., Emmanuelides, S.,
Palamaras, I., and Terzoudi, S., Lifetime prevalence fluctuations
of common and plane viral warts. J. Eur. Acad. Dermatol.
Venereol. 21(2):260–262, 2007.
3. Hogendoorn, G. K., Bruggink, S. C., Hermans, K. E.,
Kouwenhoven, S. T. P., Quint, K. D., Wolterbeek, R. et al.,
Developing and validating the Cutaneous WARTS (CWARTS) di-
agnostic tool: a novel clinical assessment and classification system
for cutaneous warts. Br. J. Dermatol. 178(2):527–534, 2018.
4. Karadaghy, B., Kamal, N. M., and Othman, N., Prevalence and
Knowledge about Genital and other Warts among Students in
Sulaimani Polytechnic University: A cross-sectional study.
Kurdistan Journal of Applied Research:96–101, 2018.
5. Loo, S. K. F., and Tang, W. Y. M., Warts (non-genital). BMJ
Clinical Evidence, 2009.
6. Uzun, R., İşler, Y., & Toksan,M., Choose of wart treatment method
using Naive Bayes and k-nearest neighbors classifiers. In: 2018
26th Signal Processing and Communications Applications
Conference (SIU). IEEE, 2018.
7. Khozeimeh, F., Alizadehsani, R., Roshanzamir, M., Khosravi, A.,
Layegh, P., and Nahavandi, S., An expert system for selecting wart
treatment method. Comput. Biol. Med. 81:167–175, 2017.
8. Al-antari, M. A., Al-masni, M. A., Choi, M. T., Han, S. M., and
Kim, T. S., A fully integrated computer-aided diagnosis system for
digital X-ray mammograms via deep learning detection, segmenta-
tion, and classification. Int. J. Med. Inform. 117:44–54, 2018.
9. Saba, L., Jain, P. K., Suri, H. S., Ikeda, N., Araki, T., Singh, B. K.
et al., Plaque tissue morphology-based stroke risk stratification
using carotid ultrasound: a polling-based PCA learning paradigm.
J. Med. Syst. 41(6):98, 2017.
10. Acharya, U. R., Faust, O., Sree, S. V., Molinari, F., Garberoglio, R.,
and Suri, J. S., Cost-effective and non-invasive automated benign &
malignant thyroid lesion classification in 3D contrast-enhanced ul-
trasound using combination of wavelets and textures: a class of
ThyroScan™algorithms. Technology in Cancer Research &
Treatment 10(4):371–380, 2011.
11. Boi, A., Jamthikar, A. D., Saba, L., Gupta, D., Sharma, A., Loi, B.
et al., A Survey on Coronary Atherosclerotic Plaque Tissue
Characterization in Intravascular Optical Coherence Tomography.
Curr. Atheroscler. Rep. 20(7):33, 2018.
12. Araki,T.,Ikeda,N.,Shukla,D.,Jain,P.K.,Londhe,N.D.,Shrivastava,
V. K. et al., PCA-based polling strategy in machine learning framework
for coronary artery disease risk assessment in intravascular ultrasound:
A link between carotid and coronary grayscale plaque morphology.
Comput. Methods Prog. Biomed. 128:137–158, 2016.
13. Maniruzzaman, M., Rahman, M. J., Al-MehediHasan, M., Suri, H.
S., Abedin, M. M., El-Baz, A., and Suri, J. S., Accurate Diabetes
Risk Stratification Using Machine Learning: Role of Missing Value
and Outliers. J. Med. Syst. 42(5):92, 2018.
14. Maniruzzaman, M., Kumar, N., Abedin, M. M., Islam, M. S., Suri,
H. S., El-Baz, A. S., and Suri, J. S., Comparative approaches for
classification of diabetes mellitus data: Machine learning paradigm.
Comput. Methods Prog. Biomed. 152:23–34, 2017.
15. Shrivastava, V. K., Londhe, N. D., Sonawane, R. S., and Suri, J. S.,
A novel and robust Bayesian approach for segmentation of psoria-
sis lesions and its risk stratification. Comput. Methods Prog.
Biomed. 150:9–22, 2017.
16. Gulshan, V., Peng, L., and Coram, M., (2016). Development and
Validation of a Deep Learning Algorithm for Detection of Diabetic
Table 10 The initial
results obtained by using
the AIRS algorithm applied on
the two original data sets
Class The Cryotherapy The Immunotherapy
PRE REC FM ACC (%) PRE REC FM ACC (%)
K2 Success 0.745 0.905 0.817 –0.867 0.915 0.890 –
Failure 0.897 0.729 0.805 –0.600 0.474 0.529 –
Weighted Average 0.826 0.811 0.810 81.11 0.810 0.822 0.814 82.22
K3 Success 0.860 0.881 0.871 –0.792 0.859 0.824 –
Failure 0.894 0.875 0.884 –0.231 0.158 0.187 –
Weighted Average 0.878 0.878 0.878 87.77 0.674 0.711 0.690 71.11
K4 Success 0.870 0.952 0.909 –0.827 0.944 0.882 –
Failure 0.955 0.875 0.913 –0.556 0.263 0.357 –
Weighted Average 0.915 0.911 0.911 91.11 0.770 0.800 0.771 80
K5 Success 0.765 0.929 0.839 –0.822 0.845 0.833 –
Failure 0.923 0.750 0.828 –0.353 0.316 0.333 –
Weighted Average 0.849 0.833 0.833 83.33 0.723 0.733 0.728 73.33
K10 Success 0.783 0.857 0.818 –0.790 0.901 0.842 –
Failure 0.864 0.792 0.826 –0.222 0.105 0.143 –
Weighted Average 0.826 0.822 0.822 82.22 0.670 0.733 0.694 73.33
Bold values indicate the highest values of classification metrics
220 Page 20 of 23 J Med Syst (2019) 43:220
Retinopathy in Retinal Fundus Photographs. JAMA. 316(22):
2402–2410, 2016.
17. Nayak, D. R., Dash, R., and Majhi, B., Brain MR image classifica-
tion using two-dimensional discrete wavelet transform and
AdaBoost with random forests. Neurocomputing 177:188–197,
2016.
18. Majid, A., Ali, S., Iqbal, M., and Kausar, N., Prediction of human
breast and colon cancers from imbalanced data using nearest neigh-
bor and support vector machines. Comput. Methods Prog. Biomed.
113(3):792–808, 2014. https://doi.org/10.1016/j.cmpb.2014.01.
001.
19. Shi, L., Xi, L., Ma, X. M., Weng, M., and Hu, X. H., A novel
ensemble algorithm for biomedical classification based on ant col-
ony optimization. Appl. Soft Comput. 11:5674–5683, 2011.
20. Arabasadi, Z., Alizadehsani, R., Roshanzamir, M., Moosaei, H.,
and Yarifard, A. A., Computer aided decision making for heart
disease detection using hybrid neural network-genetic algorithm.
Comput. Methods Prog. Biomed. 141:19–26, 2017.
21. Kim, J. K., and Kang, S., Neural network-based coronary heart
disease risk prediction using feature correlation analysis. Journal
of Healthcare Engineering 2017:1–13, 2017.
22. Danaee, P., Ghaeini, R., and Hendrix, D. A., A deep learning ap-
proach for cancer detection and relevant gene identification. Pac.
Symp. Biocomput. 22:219–229, 2016.
23. Kourou, K., Exarchos, T. P., and Exarchos, K. P., Machine learning
applications in cancer prognosis and prediction. Comput Struct
Biotechnol J 13:8–17, 2015.
24. Summers, M. J., Madl, T., Vercelli, A. E., Aumayr, G., Bleier, D.
M., and Ciferri, L., Deep Machine Learning Application to the
Detection of Preclinical Neurodegenerative Diseases of Aging.
DigitCult - Scientific Journal on Digital Cultures 2(2):9–24, 2017.
https://doi.org/10.4399/97888255088952.
25. Chikh, M. A., Saidi, M., and Settouti, N., J. Med. Syst. 36:2721,
2012. https://doi.org/10.1007/s10916-011-9748-4.
26. Abdar, M., and Yen, N. Y., Design of A Universal User Model for
Dynamic Crowd Preference Sensing and Decision-Making
Behavior Analysis. IEEE Access 5:24842–24852, 2017.
27. Xia, Y., Liu, C., Li, Y., and Liu, N., A boosted decision tree ap-
proach using Bayesian hyper-parameter optimization for credit
scoring. Expert Syst. Appl. 78:225–241, 2017.
28. Hassoon, M., Kouhi, M. S., Zomorodi-Moghadam, M., & Abdar,
M., Rule Optimization of Boosted C5. 0 Classification Using
Genetic Algorithm for Liver disease Prediction. In: Computer and
Applications (ICCA), 2017 International Conference on (pp. 299–
305). IEEE, 2017.
29. Abdar, M., Zomorodi-Moghadam, M., Das, R., and Ting, I. H.,
Performance analysis of classification algorithms on early detection
of liver disease. Expert Syst. Appl. 67:239–251, 2017.
30. Abdar, M., Yen, N. Y., and Hung, J. C. S., Improving the Diagnosis
of Liver Disease Using Multilayer Perceptron Neural Network and
Boosted Decision Trees. Journal of Medical and Biological
Engineering:1–13, 2017.
31. Zulvia, F. E., Kuo, R. J., & Roflin, E., An Initial Screening Method
for Tuberculosis Diseases Using a Multi-objective Gradient
Evolution-Based Support Vector Machine and C5. 0 Decision
Tree. In: Computer Software and Applications Conference
(COMPSAC), 2017 IEEE 41st Annual (Vol. 2, pp. 204–209).
IEEE, 2017.
32. Bai, Q., Analysis of particle swarm optimization algorithm.
Computer and information science 3(1):180, 2010.
33. Zhan, Z. H., Zhang, J., Li, Y., and Chung, H. S. H., Adaptive
particle swarm optimization. IEEE Transactions on Systems,
Man, and Cybernetics, Part B (Cybernetics) 39(6):1362–1381,
2009.
34. Toğan, V., and Daloğlu, A. T., Optimization of 3d trusses with
adaptive approach in genetic algorithms. Eng. Struct. 28(7):1019–
1027, 2006.
35. Alkeshuosh, A. H., Moghadam, M. Z., Mansoori, I. A., and Abdar,
M., Using pso algorithm for producing best rules in diagnosis of
heart disease. 2017 International Conference on Computer and
Applications (ICCA):306–311, 2017. https://doi.org/10.1109/
COMAPP.2017.8079784.
36. Paul, A., Shill, P., Rabin, M. et al., Adaptive weighted fuzzy rule-
based system for the risk level assessment of heart disease. Appl.
Intell. 48:1739, 2018. https://doi.org/10.1007/s10489-017-1037-6.
37. Manikandan, R., and Saravanan, V., A novel approach on Particle
Agent Swarm Optimization (PASO) in semantic mining for web
page recommender system of multimedia data: a health care per-
spective. Multimed. Tools Appl., 2019. https://doi.org/10.1007/
s11042-018-7141-8.
38. Chen, X., Yao, X., Tang, C., Sun, Y., Wang, X., and Wu, X.,
Detecting Parkinson’s Disease Using Gait Analysis with Particle
Swarm Optimization. In: Zhou, J., Salvendy, G. (Eds), Human
Aspects of IT for the Aged Population. Applications in Health,
Assistance, and Entertainment. ITAP 2018. Lecture Notes in
Computer Science. Vol. 10927. Cham: Springer, 2018.
39. Adem, K., Hekim, M., and Demir, S., Detection of hemorrhage in
retinal images using linear classifiers and iterative thresholding ap-
proaches based on firefly and particle swarm optimization algo-
rithms. Turk. J. Electr. Eng. Comput. Sci., 2019. https://doi.org/
10.3906/elk-1804-147.
40. Wang, H., and Lv, Y., Smart pathological brain detection system by
predator-prey particle swarm optimization and single-hidden layer
neural-network. Multimed. Tools Appl., 2016. https://doi.org/10.
1007/s11042-016-4242-0 (Online).
41. Yang, C.-H., Yang, H.-S., and Chuang, L.-Y., PBMDR: A particle
swarm optimization-based multifactor dimensionality reduction for
the detection of multilocus interactions. J. Theor. Biol. 461:68–75,
2019. https://doi.org/10.1016/j.jtbi.2018.10.012.
42. Sakri, S. B., Rashid, N. B. A., and Zain, Z. M., Particle swarm
optimization feature selection for breast cancer recurrence predic-
tion. IEEE Access 6:29637–29647, 2018.
43. Zhang, Y., Wang, S., Sui, Y., Yang, M., Liu, B., Cheng, H., Sun, J.,
Jia, W., Phillips, P., and Górriz, J. M., Multivariate Approach for
Alzheimer's Disease Detection Using Stationary Wavelet Entropy
and Predator-Prey Particle Swarm Optimization. Journal of
Alzheimer's Disease: JAD 65(3):855–869, 2018.
44. Zeng, N., Qiu, H., Wang, Z., Liu, W., Zhang, H., and Li, Y., A New
Switching-Delayed-PSO-Based Optimized SVM Algorithm for
Diagnosis of Alzheimer’s Disease. Neurocomputing, 2018.
https://doi.org/10.1016/j.neucom.2018.09.001.
45. Nekouie, A., and Moattar, M. H., Missing value imputation for
breast cancer diagnosis data using tensor factorization improved
by enhanced reduced adaptive particle swarm optimization.
Journal of King Saud University - Computer and Information
Sciences, 2018. https://doi.org/10.1016/j.jksuci.2018.01.006.
46. Wang, X., Wang, G., and Wu, Y., An Adaptive Particle Swarm
Optimization for Underwater Target Tracking in Forward Looking
Sonar Image Sequences. IEEE Access 6:46833–46843, 2018.
47. Sahu, C., Kumar, P. B., and Parhi, D. R., An Intelligent Path
Planning Approach for Humanoid Robots Using Adaptive
Particle Swarm Optimization. International Journal on Artificial
Intelligence Tools 27(05):1850015, 2018.
48. Song, Z., Liu, B., & Cheng, H., Adaptive particle swarm optimiza-
tion with population diversity control and its application in tandem
blade optimization. Proceedings of the Institution of Mechanical
Engineers, Part C: Journal of Mechanical Engineering Science,
0954406218776680, 2018.
49. Khozeimeh, F., Jabbari Azad, F., Mahboubi Oskouei, Y., Jafari, M.,
Tehranian, S., Alizadehsani, R., and Layegh, P., Intralesional
J Med Syst (2019) 43:220 Page 21 of 23 220
immunotherapy compared to cryotherapy in the treatment of warts.
Int. J. Dermatol. 56(4):474–478, 2017.
50. Cryotherapy Dataset Data Set, https://archive.ics.uci.edu/ml/
datasets/Cryotherapy+Dataset+. Accessed in April 2018.
51. Immunotherapy Dataset Data Set, https://archive.ics.uci.edu/ml/
datasets/Immunotherapy+Dataset, Accessed in April 2018.
52. Kennedy, J., Eberhart, R. (1995). Particle swarm optimization. In
Neural Networks, 1995. Proceedings, IEEE International
Conference on (pp. 1942–1948). IEEE, 2011.
53. Kennedy, J., Particle swarm optimization. In Encyclopedia of ma-
chine learning (pp. 760–766). Springer US.
54. Sengupta, S., Basak, S., & Peters, R. A., Data Clustering using a
Hybrid of Fuzzy C-Means and Quantum-behaved Particle Swarm
Optimization. In Computing and Communication Workshop and
Conference (CCWC), 2018 IEEE 8th Annual (pp. 137–142).
IEEE, 2018.
55. Engelbrecht, A. P., Computational Intelligence. England: John
Wiley & Sons. Inc., 2007.
56. Ahmed, H., and Glasgow, J., Swarm intelligence: concepts, models
and applications. Kingston: School Of Computing, Queens
University Technical Report, 2012.
57. Shi, Y., & Eberhart, R., A modified particle swarm optimizer. In
Evolutionary Computation Proceedings, 1998. IEEE World
Congress on Computational Intelligence., The 1998 IEEE
International Conference on (pp. 69–73). IEEE, 1998.
58. Hatanaka, T., Korenaga, T., Kondo, N., & Uosaki, K., Search per-
formance improvement for PSO in high dimensional space. In
Particle Swarm Optimization. IntechOpen, 2009.
59. Kawakami, K., and Meng, Z. Q., Improvement of particle swarm
optimization. PIERS Online 5(3):261–264, 2009.
60. Liu, H. R., Cui, J. C., Lu, Z. D., Liu, D. Y., and Deng, Y. J., A
hierarchical simple particle swarm optimization with mean dimen-
sional information. Appl. Soft Comput. 76:712–725, 2019.
61. Jain, Y. K., and Bhandare, S. K., Min max normalization based data
perturbation method for privacy protection. International Journal of
Computer & Communication Technology 2(8):45–50, 2011.
62. Meera, P. S., and Hemamalini, S., Optimal Siting of Distributed
Generators in a Distribution Network using Artificial Immune
System. International Journal of Electrical and Computer
Engineering (IJECE) 7(2):641–649, 2017.
63. de Castro, L. N., and Timmis, J., Artificial immune systems: A
novel paradigm to pattern recognition. Artificial Neural Networks
in Pattern Recognition 1:67–84, 2002.
64. Draman, N. A., Wilson, C., and Ling, S., Bio-inspired audio
content-based retrieval framework (B-ACRF). World Academy of
Science. Eng. Technol. 29(5):785–790, 2009.
65. Timmis, J., Neal, M., and Hunt, J., An artificial immune system for
data analysis. Biosystems 55(1–3):143–150, 2000.
66. Porselvi, S., Balaji, A. N., and Jawahar, N., Artificial immune sys-
tem and particle swarm optimisation algorithms for an integrated
production and distribution scheduling problem. International
Journal of Logistics Systems and Management 30(1):31–68, 2018.
67. Thumati, B., and Sarangapani, J., Authors’Reply to BComments on
BA Novel Fault Diagnosis and Prediction Scheme Using a
Nonlinear Observer With Artificial Immune System as an Online
Approximator^^. IEEE Trans. Control Syst. Technol. 26(1):377–
378, 2018.
68. Nikoo, M. R., and Kerachian, R., Wave Height Prediction Using
Artificial Immune Recognition Systems (AIRS) and Some Other
Data Mining Techniques. Iranian Journal of Science and
Technology, Transactions of Civil Engineering 41(3):329–344,
2017.
69. Brownlee, J., Artificial immune recognition system (airs)-a review
and analysis. Center for Intelligent Systems and Complex Processes
(CISCP), Faculty of Information and Communication Technologies
(ICT), Swinburne University of Technology, Victoria, 2005.
70. Petitjean, F., Buntine, W., Webb, G. I., and Zaidi, N., Accurate
parameter estimation for Bayesian network classifiers using hierar-
chical Dirichlet processes. Mach. Learn. 107(8–10):1303–1331,
2018.
71. Mahadevan, S., and Rebba, R., Validation of reliability computa-
tional models using Bayes networks. Reliab. Eng. Syst. Saf. 87(2):
223–232, 2005.
72. Gardner, M. W., and Dorling, S. R., Artificial neural networks (the
multilayer perceptron)—a review of applications in the atmospheric
sciences. Atmos. Environ. 32(14–15):2627–2636, 1998.
73. Bayat, F. M., Prezioso, M., Chakrabarti, B., Nili, H., Kataeva, I.,
and Strukov, D., Implementation of multilayer perceptron network
with highly uniform passive memristive crossbar circuits. Nat.
Commun. 9(1):2331, 2018.
74. Drazin, S., & Montag, M., Decision tree analysis using weka.
Machine Learning-Project II, University of Miami, 1–3, 2012.
75. Satyanarayana, N., Ramadevi, Y., & Chari, K. K., High blood pres-
sure prediction based on AAA using J48 classifier. In Signal
Processing and Communication Engineering Systems (SPACES),
2018 Conference on (pp. 121–126). IEEE, 2018.
76. Probst, P., and Boulesteix, A. L., To Tune or Not to Tune the
Number of Trees in Random Forest. J. Mach. Learn. Res.
18(181):1–18, 2018.
77. Pal, M., Random forest classifier for remote sensing classification.
Int. J. Remote Sens. 26(1):217–222, 2005.
78. Waizumi, Y., Kato, N., Saruta, K., and Nemoto, Y., High speed and
high accuracy rough classification for handwritten characters using
hierarchical learning vector quantization. IEICE Trans. Inf. Syst.
83(6):1282–1290, 2000.
79. Janghel, R. R., Tiwari, R., & Shukla, A., Breast Cancer Diagnostic
System using Hierarchical Learning Vector Quantization. In IJCA
Proceedings on National Seminar on Application of Artificial
Intelligence in Life Sciences, 2013.
80. Sauermann, R., Boja, D., Kirschbaum, F., and Nelles, O., Particle
Swarm Optimization for Automotive Model-Based Calibration.
IFAC Proceedings Volumes 43(7):733–738, 2010.
81. Guimarães, A. J., Araujo, V. J. S., de Campos Souza, P. V., Araujo,
V. S., & Rezende, T. S., Using fuzzy neural networks to the predic-
tion of improvement in expert systems for treatment of immuno-
therapy. In: Ibero-American Conference on Artificial Intelligence
(pp. 229–240). Springer, Cham, 2018.
82. Alizadehsani, R., Abdar, M., Jalali, S. M. J., Roshanzamir, M.,
Khosravi, A., & Nahavandi, S., Comparing the Performance of
Feature Selection Algorithms for Wart Treatment Selection. In
Proceedings of The International Workshop on Future
Technology (pp. 6–18), 2018.
83. Nugroho, H. W., Adji, T. B., and Setiawan, N. A., Random forest
weighting based feature selection for c4. 5 algorithm on wart treat-
ment selection method. International Journal on Advanced Science,
Engineering and Information Technology 8(5):1858–1863, 2018.
84. Jia, W., Deng, Y., Xin, C., Liu, X., and Pedrycz, W., A classification
algorithm with Linear Discriminant Analysis and Axiomatic Fuzzy
Sets. Mathematical Foundations of Computing 2(1):73–81, 2019.
85. Junio Guimarães, A., Vitor de Campos Souza, P., Jonathan Silva
Araújo, V., Silva Rezende, T., and Souza Araújo, V., Pruning Fuzzy
Neural Network Applied to the Construction of Expert Systems to
Aid in the Diagnosis of the Treatment of Cryotherapy and
Immunotherapy. Big Data and Cognitive Computing 3(2):22, 2019.
86. Pławiak, P., and Rzecki, K., Approximation of phenol concentration
using computational intelligence methods based on signals from the
metal-oxide sensor array. IEEE Sensors J. 15(3):1770–1783, 2015.
87. Olivas, F., Amador-Angulo, L., Perez, J., Caraveo, C., Valdez, F.,
and Castillo, O., Comparative study of type-2 fuzzy particle swarm,
bee colony and bat algorithms in optimization of fuzzy controllers.
Algorithms 10(3):101, 2017.
220 Page 22 of 23 J Med Syst (2019) 43:220
88. Valdez, F., Vazquez, J. C., Melin, P., and Castillo, O., Comparative
study of the use of fuzzy logic in improving particle swarm optimi-
zation variants for mathematical functions using co-evolution.
Appl. Soft Comput. 52:1070–1083, 2017.
89. Gaxiola, F., Melin, P., Valdez, F., Castro, J. R., and Castillo, O.,
Optimization of type-2 fuzzy weights in backpropagation learning
for neural networks using GAs and PSO. Appl. Soft Comput. 38:
860–871, 2016.
90. Olivas, F., Valdez, F., Castillo, O., and Melin, P., Dynamic param-
eter adaptation in particle swarm optimization using interval type-2
fuzzy logic. Soft. Comput. 20(3):1057–1070, 2016.
91. Melin, P., Amezcua, J., Valdez, F., and Castillo, O., A new neural
network model based on the LVQ algorithm for multi-class classi-
fication of arrhythmias. Inf. Sci. 279:483–497, 2014.
92. Pławiak, P., Novel genetic ensembles of classifiers applied to myo-
cardium dysfunction recognition based on ECG signals. Swarm and
Evolutionary Computation 39:192–208, 2018.
93. Książek, W., Abdar, M., Acharya, U. R., and Pławiak, P., A novel
machine learning approach for early detection of hepatocellular
carcinoma patients. Cogn. Syst. Res. 54:116–127, 2019.
Publisher’sNote Springer Nature remains neutral with regard to jurisdic-
tional claims in published maps and institutional affiliations.
J Med Syst (2019) 43:220 Page 23 of 23 220
A preview of this full-text is provided by Springer Nature.
Content available from Journal of Medical Systems
This content is subject to copyright. Terms and conditions apply.
