Int. J. Biomedical Engineering and Technology, Vol. 34, No.

1, 2020 57

Machine learning algorithms using binary

classification and multi model ensemble techniques
for skin diseases prediction

Vikas Chaurasia and Saurabh Pal*

Department of Computer Applications,
VBS Purvanchal University,
Jaunpur, UP, India
Email: chaurasia.vikas@gmail.com
Email: 2drsaurabhpal@yahoo.co.in
*Corresponding author

Abstract: Skin disease has more touchiness as compared to any other disease.
Regular skin issues are dermatitis. The main focus of this research paper will be
on dermatology database which contains different eryhemato-squamous
diseases class as psoriasis, seboreic dermatitis, lichen planus, pityriasisrosea,
cronic dermatitis and pityriasisrubrapilaris. Each record is a collection of 33
attributes which are linear values and one attribute of them is nominal. The
75% of the dataset utilise for demonstrating and keep down 25% for approval.
The purpose of this article is to achieve the best-performing classifier that will
communicate in the collection of dermatological information. Therefore, k-
nearest neighbours and support vector machines are used. By using ten-fold
cross validation and assess calculations utilising the accuracy metric. This is a
gross metric which will prove the developed model is best one.

Keywords: eryhemato-squamous; k-nearest neighbours; KNN; classification

and regression trees; CARTs; support vector machines; SVMs; ensemble
methods.

Reference to this paper should be made as follows: Chaurasia, V. and Pal, S.

(2020) ‘Machine learning algorithms using binary classification and multi
model ensemble techniques for skin diseases prediction’, Int. J. Biomedical
Engineering and Technology, Vol. 34, No. 1, pp.57–74.

Biographical notes: Vikas Chaurasia holds an MSc in Math and MCA from
UNSIET VBS Purvanchal University, U.P., India. He is currently a PhD
student and teaching assistant in the Computer Applications Department at the
V.B.S. Purvanchal University, where he teaches data mining, mathematics and
computer organisation. His research interests focus on the data mining
techniques to predict diseases. He has published more than 12 international
research papers related to disease perdition using data mining methods in
reputed journals. His area of research includes data mining, machine learning,
python programming, deep learning and artificial intelligence.

Saurabh Pal received his MSc in Computer Science in 1996 and obtained his
PhD in 2002. He then joined the Department of Computer Applications, VBS
Purvanchal University, Jaunpur as a Lecturer. Currently, he is working as Head
and Associate Professor. He has authored more than 53 research papers in
international/national conference/journals as well as four books and also guides

Copyright © 2020 Inderscience Enterprises Ltd.

58 V. Chaurasia and S. Pal

research scholars in computer science/applications. He is an active member of

CSI, Society of Statistics and Computer Applications and working as member
of editorial board for more than 15 international journals. His research interests
include bioinformatics, data mining, and artificial intelligence.

1 Introduction

“India suffers today in the estimation of the world, more through the world’s
ignorance of her achievements than in the absence or insignificance of these
achievements.” (Mukhopadhyay, 2016)
Nearly 1,500 BC a medical document on skin ailments Ebers Papyrus was found in
ancient Egypt. It portrays different skin maladies, including rashes, ulcers and tumours,
and recommends medical procedure for balms to care for the afflictions (Hartmann,
2016). From that point to now the skin sickness portion has indicated colossal
development. The predominance of skin malady in India is 10 to 12% of the all out
populace with eczema and psoriasis being the significant benefactors. Because of
contamination, bright light, and an unnatural weather change, photosensitive skin issue
like tanning, colour obscuring, sunburn, skin malignant growths, and irresistible
infections are expanding at a quicker pace. A 1% decrease in ozone prompts a 2 to 4%
expansion in the occurrence of tumours. The seriousness of developing skin illnesses in
India is additionally underlined by the way that the World Health Organization (WHO)
has included skin infection under the most widely recognised non-transferable maladies
in India. What’s more, there is an absence of offices that give thorough skin related
medicines under one rooftop. “The circumstance is additionally compounded by the low
accessibility of dermatologists in India. At present, there are around 6,000 dermatologists
obliging a populace of more than 135 carore. This implies for each 100,000 individuals,
just 0.49 dermatologists are accessible in India when contrasted with 3.2 in numerous
conditions of the US.” Different tertiary consideration private setups come up short on
the capacity to treat incessant, hereditary and paediatric skin diseases. Likewise, their
capacity to give complete derma pathology and immunopathology units is additionally
restricted. Non-careful restorative administrations given by independent healthy skin
focuses take into account just a little portion of skin treatment and a specific section of
the general public. Consequently, there is a critical need of exhaustive healthy skin setups
giving all healthy skin medications under one rooftop.
With the quick advancement of computer helped procedures in late years, use of
machine learning strategies is playing an undeniably essential job in the skin infection
analysis, and different forecast calculations are being investigated persistently by
analysts. When we take a shot at a machine learning venture, we regularly end up with
various great models to look over. Each model will have distinctive execution qualities.
Utilising re sampling techniques like cross validation, we can get a gauge for how precise
each model might be on inconspicuous information. We should almost certainly utilise
these appraisals to pick a couple of best models from the suite of models that we have
made. When we have another dataset, it is a smart thought to envision the information
utilising diverse methods so as to take a gander at the information from alternate points of
view. A similar thought applies to display determination. We should utilise various
distinctive methods for taking a gander at the assessed exactness of your machine
 Machine learning algorithms using binary classification 59

learning calculations so as to pick the couple of calculation to settle. An approach to do

this is to utilise representation strategies to demonstrate the normal precision, change and
different properties of the appropriation of model correctness.
A large number of researches have done work on skin disease. Skin disease can be
evaluated on the basis of different classifiers used on skin disease datasets and find good
accuracy. To improve the accuracy many researchers applied ensemble techniques. The
ensemble techniques are used to enhance the accuracy of predicted model by combing
different classifiers. To achieve higher accuracy several researches use the feature
selection techniques.

2 Literature survey

There are various studies has done to solve skin disease prediction and developed model
to achieve best accuracy.. They all offer the clinical highlights of erythema and scaling
with almost no distinctions. The illnesses in this gathering are psoriasis,
seboreicdermatitis, lichenplanus, pityriasisrosea, perpetual dermatitis and
pityriasisrubrapilaris (Güvenir et al., 1998). There have been a few examinations given an
account of erythemato-squamous diseases finding. These investigations have connected
diverse techniques to the given issue and accomplished distinctive characterisation
correctness’s. Among these investigations the main work on the differential analysis of
erythemato-squamous illnesses was shown Table 1.
Table 1 A few investigations which have dealt with skin disease mining

Author Year Method Classification accuracy

Bojarczuk et al. 2001 A constrained-syntax genetic 96.64%
programming C4.5 89.12%
Chang and Chen 2009 Decision tree 80.33%
Neural network 92.62%
Guvenir et al. 1998 VFI5 96.2%
Guvenir and Emeksiz 2000 Nearest neighbour classifier 99.2%
Naïve Bayesian classifier
VFI5
Übeylı and Güler 2005 ANFIS 95.5%
Nanni 2006 LSVM 97.22%
RS 97.22%
B1_5 97.5%
B1_10 98.1%
B1_15 97.22%
B2_5 97.5%
B2_10 97.8%
B2_15 98.3%
60 V. Chaurasia and S. Pal

Table 1 A few investigations which have dealt with skin disease mining (continued)

Author Year Method Classification accuracy

Polat and Güneş 2009 C4.5 and one-against-all 96.71%
Übeyli 2009 CNN 97.77%
Übeyli and Doğdu 2010 K-mean clustering 94.22%
Lekkas and Mikhailov 2010 Evolving fuzzy classification 97.55%
Xie and Wang 2011 IFSFS and SVM 98.61%
Amarathunga et al. 2015 AdaBoost 85% for eczema
BayesNet 95% for impetigo
J48 85% for melanoma
Badrinath et al. 2016 FELM 98.21%
ELM 95.87%
Hybrid AdaBoost 99.26%
Maghooli et al. 2016 CART 93.69%
Zhou et al. 2017 ANN 92.4%
Idoko et al. 2018 CART 94.84%
ARFCMC 75.96%
ANFIS 95.50%
AEC 97.32%
FNN 98.37%
Zhang et al. 2018 ANN 96.8%
Verma et al. 2019a Bagging 98.56%
AdaBoost 99.25%
Gradient boosting 99.86%
Verma et al. 2019b CART 94.17%
SVM 96.93%
DT 93.82%
RF 97.27%
GBDT 96.25%
Ensemble method 98.64%

In this paper, we have used two types of machine learning classification algorithms first
linear algorithms and second nonlinear algorithms. Linear type of algorithms includes
logistic regression (LR) and linear discriminant analysis (LDA) and nonlinear type of
algorithms includes k-nearest neighbours (KNN), support vector machines (SVMs),
classification and regression trees (CARTs) and Gaussian Naive Bayes (NB).
Four ensemble methods for improving the performance of algorithms are used in
which two are boosting algorithms gradient boosting (GBM) and AdaBoost (AB) and two
are bagging type extra trees (ETs) and random forests (RFs) to develop a multi model
ensemble model.
In last the proposed machine learning binary classification and multi model ensemble
methods are used to enhance the output of ensemble techniques.
 Machine learning algorithms using binary classification 61

3 Methods

The methodology of the proposed binary classification machine learning and multi model
ensemble is shown in Figure 1. In first stage, features are selected for the
erythemato-squamous dataset, in this process we have select the important features we
get best dataset. Then, the best dataset is divided into k groups of training and testing
datasets. This division of training and testing dataset is done using k-fold cross validation
process. Then, we will evaluate algorithms using the accuracy metric. This is a gross
metric that will give a quick idea of how correct a given model is. We apply the base
learners and then evaluated the performance of base classifiers. Now all the same base
learners are evaluated with standardised scaled copy of dataset. In next section, an
ensemble model is used to improve the accuracy by using four different machine learning
algorithms; two boosting and two bagging methods. Finally compare the predictions after
algorithm tuning with ensemble models to aim of reducing the generalisation error and
procuring a more accurate result.

Figure 1 Flowchart of the proposed methodology

3.1 Feature selection methods

Utilisation of dermatology information with an expanding number of highlights and data
makes it all the more difficult to create classification models. Feature determination is
where we consequently select that feature in our data that contribute most to the forecast
variable or yield in which we are intrigued. Having unessential highlights in our data can
diminish the exactness of numerous models. There are three advantages of performing
highlight determination before displaying our data improve accuracy, i.e., less deceptive
data implies demonstrating precision makes strides, reduce training time, i.e., less data
implies that algorithms train quicker and reduce over fitting, i.e., less repetitive data
implies less chance to settle on choices dependent on noise (https://scikit-learn.org/
stable/modules/feature_selection.html).
This database contains 34 properties, 33 of which are linear esteemed and one of
them is nominal. The differential analysis of erythemato-squamous ailments is a genuine
issue in dermatology. They all offer the clinical highlights of erythema and scaling, with
next to no distinctions. The infections in this gathering are psoriasis, seboreic dermatitis,
lichen planus, pityriasisrosea, cronic dermatitis, and pityriasisrubrapilaris. Typically a
biopsy is fundamental for the conclusion however lamentably these maladies share
numerous histopathological includes as well. Another trouble for the differential
conclusion is that a infection may demonstrate the highlights of another illness toward the
start arrange and may have the trademark highlights at the accompanying stages. Patients
were first assessed clinically with 12 highlights. A short time later, skin tests were taken
62 V. Chaurasia and S. Pal

for the assessment of 22 histopathological highlights. The estimations of the

histopathological highlights are controlled by an examination of the examples under a
magnifying instrument. In the dataset developed for this space, the family ancestry
include has the esteem 1 if any of these infections has been seen in the family, and 0
generally. The age highlight essentially speaks to the period of the patient. Each and
every other component (clinical and histopathological) was given a degree in the scope of
0 to 3. Here, 0 demonstrates that the highlight was absent, 3 demonstrate the biggest sum
conceivable, what’s more, 1, 2 demonstrate the relative middle of the road esteems. The
names and id quantities of the patients were as of late expelled from the database.

3.2 Cross validation

For some, characterisation models, the multifaceted nature might be administered by
numerous parameters. So as to accomplish the best forecast execution on new
information, we wish to find appropriate values of the multifaceted nature parameters that
lead to the ideal model for a specific application. On the off chance that information are
ample, at that point a basic path for model determination is to separate the whole
information into three subsets, the training set, the validation set and the test set. A scope
of models are prepared on the training set, thought about and chose on the validation set,
and finally evaluated on the test set. Among the diverse complex models that have been
prepared, the one having the best prescient execution is chosen, which is a compelling
model approved by the information in the validation set. In a pragmatic application, be
that as it may, the supply of information for preparing and testing is constrained,
prompting an increment of the speculation mistake. A way to deal with decreasing the
speculation mistake and averting over-fitting is to utilise cross approval.

Figure 2 Techniques to validate models/classifiers

 Machine learning algorithms using binary classification 63

The method of ten-fold cross validation utilised in this paper is outlined in Figure 2. We
will part the stacked informational index into two, 80% of which we will use to prepare
our models and 20% that we will keep down as an validate informational collection. We
will utilise ten-fold cross validation to appraise accuracy. This will part our dataset into
ten parts, train on 9 and test on 1 and rehash for all blends of train-test parts. We are
utilising the measurement of precision to assess models. This is a proportion of the
quantity of accurately anticipated occasions isolated by the absolute number of cases in
the dataset duplicated by 100 to give a rate.

3.3 Evaluation of algorithms

In the wake of pre-processing of the informational collections, we evaluate the
expectation execution of six famous order techniques towards finding of skin malady.
Specifically, we apply LR, LDA, KNN, CARTs, Gaussian NB, SVMs as first-arrange
order models. These six order strategies are of high exactness in practical applications
and are assessed as follows.
• LR studies about the relationship between a categorical dependent and a set of
independent (informative) factors. LR is utilised when the dependent variable has
just two qualities, for example, 0 and 1 or yes and no. LR contends with discriminant
investigation as a technique for examining categorical-response factors. LR is
increasingly adaptable and more qualified for displaying most circumstances than is
discriminant examination. This is on the grounds that LR does not accept that the
independent variables are normally distributed, as discriminant investigation does. It
plays out an extensive remaining examination including symptomatic lingering
reports and plots. It can play out an autonomous variable subset choice pursuit,
searching for the best relapse show with the least free factors. It gives certainty
interims on anticipated qualities, and gives ROC bends to help decide the best cut-off
point for arrangement. It permits to validate results via automatically classifying
rows that are not utilised amid the investigation.
• LDA is a strategy utilised in pattern recognition, machine learning and statistics how
to locate a combination of linear features that finds at least two or more groups of
occasions. LDA works on independent variables when the estimations made for
every perception are continuous. Categorical independent variables are managed as
the proportional system is discriminant correspondence analysis (Abdi, 2007;
Perriere and Thioulouse, 2003). When groups are known from the priori discriminant
examination is utilised. Each case must have a score on somewhere around one
quantitative indicator measures, and a score on a group measure. In straightforward
terms, discriminant work examination is order - the demonstration of circulating
things into gatherings, classes or classifications of a similar sort.
• KNN is a technique for classification and regression (Altman, 1992). k-NN is a kind
of instance-based learning or lazy learning, where the function is only approximated
locally, and all calculations are postponed to classification. In the case of K-NN, the
input contains the k closest training examples in the feature space. The output of
K-NN depends on whether it is used for classification or regression: The k-NN
64 V. Chaurasia and S. Pal

algorithm is one of the simplest algorithms of all machine learning algorithms.

Whether it is classification or regression, a useful technique can be used to assign
weights to the contributions of neighbours so that closer neighbours and farther
neighbours contribute more to the average. For example, a common weighting
scheme is to give each neighbour a weight of 1/d, where d is the distance to the
neighbour. Neighbours are obtained from a set of objects whose class (for k-NN
classification) or object attribute values (for k-NN regression) are known. Think of it
as a training set for the algorithm, although no explicit training steps are needed.
• CART is usually used for data mining, and its purpose is to build a model that
predicts the estimation of the target among some input estimates. It is one of the
predictive demonstration methods used in statistics, data mining and machine
learning. A tree model in which the target variables can be arranged in discrete
masses is called a classification tree. In these tree structures, the leaves represent the
class names, and the branches represent the highlighting of the class names. A
selection tree whose target variable can have constant quality is called a regression
tree. In decision surveys, decision trees can be used to present decisions and
decisions visually and explicitly.
• NB is an explicitly utilised when the highlights have continuous values. NB can be
stretched out to real-valued attributes, most usually by expecting a Gaussian
distribution. This expansion of NB is called Gaussian NB. Different capacities can
be utilised to gauge the distribution of the data, however the Gaussian is the most
effortless to work with in light of the fact that we just need to estimate the mean and
the standard deviation from training data.
• SVM is a supervised learning model with a related learning algorithm that analyses
data for classification and regression analysis. Supervised learning is not possible, if
the data is unlabeled, then unsupervised learning technique is needed. This method
tries to naturally cluster different data into groups, and maps the new data among
created groups. SVM construct a set of hyperplanes or a hyperplane. SVM can be
used for regression, classification and outlier detection. Intuitively, there is a
maximum distance (the so-called functional margin) to the nearest training data point
of any class, because generally the larger the margin, the lower the generalisation
error of the classifier.
This rundown is a decent blend of simple linear (LR and LDA), nonlinear (k-NN, CART,
NB, SVM) algorithms. We reset the arbitrary number seed before each run to guarantee
that the assessment of every algorithms is performed utilising the very same data parts. It
ensures the outcomes are straightforwardly comparable.

3.4 Ensemble methods

In this research paper ensemble method is used as a method to find the accuracy of the
skin disease dataset to improve the performance of algorithms. We will evaluate four
different ensemble machine learning algorithms, two boosting, AB and GBM and two
bagging methods RF and ET see Figure 3.
 Machine learning algorithms using binary classification 65

Figure 3 An illustration of the ensemble techniques structure (see online version for colours)

3.4.1 General procedure of boosting

The term boosting alludes to a group of algorithms that can change over weak learner to
strong learner. Naturally, a weak learner is simply marginally superior to irregular guess,
while a strong learner is extremely near perfect execution. If the appropriate response is
positive, any weak learner is conceivably able to be boosted to a strong learner, especially
it is very general to find weak learner however hard to get strong learner. Boosting works
via preparing a lot of learners consecutively and consolidating them for expectation,
where the later learners focus more around the errors of the prior learners. A general
boosting method is described below.
Input: Distribution of sample S;
Learning base algorithm A;
Total number of learning rounds N.
Process:
1 S1 = S. #Initialisation of distribution
2 for n = 1, …, N:
3 hn = A(Sn); #A weak learner trained from distribution Sn
4 µn = Px~Sn(hn(x) ≠ f(x)); #Evaluation of the error of hn
5 Sn+1 = adjustment of distribution (Sn, µn)
6 end
Output: H(x) = combine outputs ({h1(x), …, hn(x)})
66 V. Chaurasia and S. Pal

3.4.2 Bagging methods

The name Bagging originated from the contraction of bootstrap aggregating. As the name
suggests, the two key elements of bagging are bootstrap and aggregation. For a training
informational collection, one probability is by all accounts examining various non-
covered data subsets and after that preparation a base learner from every one of the
subsets. In any case, since we do not have unending training data information, such a
procedure will deliver exceptionally little and unrepresentative examples, prompting poor
execution of base learners. Bagging receives the bootstrap circulation for creating diverse
base learners. It applies bootstrap inspecting to get the data subsets for preparing the base
learners. A general bagging method is describe below.
Input: Distribution of sample S = {(x1, y1), (x2, y2), …, (xm, ym)};
Learning base algorithm A;
Total number of learning rounds N.
Process:
1 for n = 1, …, N:
2 hn = A(S, Sbr) #Sbr is the bootstrap distribution
3 end


N
Output: H(x) = arg max (h n (x) = y)
n =1

y∈Y

4 Results

Here we utilised distance-based algorithms like KNN and SVMs We have utilised
ten-fold cross validation. The dataset is not excessively little and this is a decent standard
test saddle setup. We will assess algorithms utilising the exactness metric. This is a gross
metric that will give a fast thought of how right a given model is. Making a standard of
act on this issue and spot-check various distinctive algorithms. We will choose a suite of
various algorithms fit for dealing with this classification issue. The algorithms all
utilisation defaults tuning parameters.
On comparing the algorithms mean accuracy values are given in Table 2.
Table 2 Output of evaluating algorithms

Algorithms Mean accuracy values

LR 0.979425 (0.022806)
LDA 0.962299 (0.024175)
KNN 0.855747 (0.051314)
CART 0.935057 (0.028180)
NB 0.890230 (0.072177)
SVM 0.921034 (0.027220)
 Machine learning algorithms using binary classification 67

It is always learned to look at the distribution of accuracy values calculated across cross
validation folds. The results demonstrate a tight distribution for LR which is empowering,
recommending low difference. The poor outcomes for KNN are amazing. See Figure 4.
It is conceivable that the changed appropriation of the attributes is affecting the
exactness of algorithms, for example, KNN. In the following segment we will rehash this
spot-check with a standardised copy of the training dataset.

Figure 4 Plots of algorithm comparison (see online version for colours)

4.1 Evaluation of algorithms with standardise data

We speculate that different distributions of the original data may adversely affect the
capabilities of some algorithms. How can we evaluate similar algorithms with a
standardised copy of the dataset. This is where the data changes and the ultimate goal is
that the average estimated value for each attribute is zero and the standard deviation is 1.
Similarly, we need to keep a strategic distance from data leaks when changing data. A
good way to avoid data leaks is to use a pipeline of standardised data and build a model
for each fold in the cross-validation test bundle.
Table 3 Output of evaluating algorithms on the scaled dataset

Algorithms Mean accuracy values

ScaledLR 0.972529 (0.025776)
ScaledLDA 0.962299 (0.024175)
ScaledKNN 0.969195 (0.018533)
ScaledCART 0.938391 (0.029970)
ScaledNB 0.869655 (0.087102)
ScaledSVM 0.969310 (0.023769)

We can see that LR is as yet progressing admirably. We can likewise observe that the
standardisation of the data has lifted the aptitude of SVM to be the most precise
algorithm tried up until this point. See Table 3.
Figure 5 plots the distribution of the accuracy scores using box and whisker plots.
68 V. Chaurasia and S. Pal

Figure 5 Scaled algorithm comparison (see online version for colours)

The outcomes propose delving further into the LR and SVM algorithms. Almost
certainly, setup past the default may yield significantly increasingly precise models.

4.2 Tuning of algorithms

In this area we research tuning the parameters for two algorithms that show high accuracy
from the spot-checking in the past segment: LR and SVM.

4.2.1 Tuning LR
We use parameter C as our regularisation parameter. Parameter C = 1/λ. Lambda (λ)
controls the exchange off between enabling the model to expand its unpredictability as
much as it needs with attempting to keep it straightforward. For instance, if λ is
exceptionally low or 0, the model will have enough capacity to build its complexity (over
fit) by relegating enormous qualities to the loads for every parameter. In the event that, in
the other hand, we increment the estimation of λ, the model will tend to under fit, as the
model will turn out to be excessively straightforward. Parameter C will work a different
way. For little estimations of C, we increment the regularisation quality which will make
basic models which under fit the information. For enormous estimations of C, we low the
intensity of regularisation which impels the model is permitted to expand its multifaceted
nature, and in this way, over fit the data. See table 4.
Table 4 Results of tuning LR on the scaled dataset

Best: 0.976027 using {‘C’: 10}

0.972603 (0.025510) with: {‘C’: 0.01}
0.969178 (0.032460) with: {‘C’: 0.1}
0.972603 (0.025740) with: {‘C’: 1}
0.976027 (0.022005) with: {‘C’: 10}
0.976027 (0.022005) with: {‘C’: 100}

We can see that the ideal setup is C = 10. This is interesting as the algorithm will make
forecasts utilising the most comparative example in the training dataset alone.
 Machine learning algorithms using binary classification 69

4.2.2 Tuning SVM

We can adjust two key parameters of the SVM algorithm, namely the estimated value of
C and the kernel type. The default setting for SVM is to use a radial basis function (RBF)
kernel with C esteem set to 1.0. We will perform a grid search using a normalised copy of
the ten-fold cross-validation and training dataset. We will try various less difficult kernel
types and C, because their bias is not very big, but biased. See Table 5.
Table 5 Results of tuning SVM on the scaled dataset

Best: 0.972603 using {‘C’: 0.3, ‘kernel’: ‘sigmoid’}

0.965753 (0.026474) with: {‘C’: 0.1, ‘kernel’: ‘linear’}
0.571918 (0.081588) with: {‘C’: 0.1, ‘kernel’: ‘poly’}
0.815068 (0.047240) with: {‘C’: 0.1, ‘kernel’: ‘rbf’}
0.880137 (0.054414) with: {‘C’: 0.1, ‘kernel’: ‘sigmoid’}
0.952055 (0.027061) with: {‘C’: 0.3, ‘kernel’: ‘linear’}
0.780822 (0.058694) with: {‘C’: 0.3, ‘kernel’: ‘poly’}
0.958904 (0.033551) with: {‘C’: 0.3, ‘kernel’: ‘rbf’}
0.972603 (0.025740) with: {‘C’: 0.3, ‘kernel’: ‘sigmoid’}
0.952055 (0.016721) with: {‘C’: 0.5, ‘kernel’: ‘linear’}
0.797945 (0.059439) with: {‘C’: 0.5, ‘kernel’: ‘poly’}
0.969178 (0.028384) with: {‘C’: 0.5, ‘kernel’: ‘rbf’}
0.972603 (0.025510) with: {‘C’: 0.5, ‘kernel’: ‘sigmoid’}
0.945205 (0.027150) with: {‘C’: 0.7, ‘kernel’: ‘linear’}
0.815068 (0.066129) with: {‘C’: 0.7, ‘kernel’: ‘poly’}
0.972603 (0.025510) with: {‘C’: 0.7, ‘kernel’: ‘rbf’}
0.965753 (0.030612) with: {‘C’: 0.7, ‘kernel’: ‘sigmoid’}
0.955479 (0.026727) with: {‘C’: 0.9, ‘kernel’: ‘linear’}
0.835616 (0.076927) with: {‘C’: 0.9, ‘kernel’: ‘poly’}
0.969178 (0.023781) with: {‘C’: 0.9, ‘kernel’: ‘rbf’}
0.969178 (0.028384) with: {‘C’: 0.9, ‘kernel’: ‘sigmoid’}
0.952055 (0.027422) with: {‘C’: 1.0, ‘kernel’: ‘linear’}
0.845890 (0.056523) with: {‘C’: 1.0, ‘kernel’: ‘poly’}
0.969178 (0.023781) with: {‘C’: 1.0, ‘kernel’: ‘rbf’}
0.969178 (0.028384) with: {‘C’: 1.0, ‘kernel’: ‘sigmoid’}
0.958904 (0.025789) with: {‘C’: 1.3, ‘kernel’: ‘linear’}
0.890411 (0.029697) with: {‘C’: 1.3, ‘kernel’: ‘poly’}
0.969178 (0.023781) with: {‘C’: 1.3, ‘kernel’: ‘rbf’}
0.972603 (0.025510) with: {‘C’: 1.3, ‘kernel’: ‘sigmoid’}
0.962329 (0.028579) with: {‘C’: 1.5, ‘kernel’: ‘linear’}
0.900685 (0.035452) with: {‘C’: 1.5, ‘kernel’: ‘poly’}
0.969178 (0.023781) with: {‘C’: 1.5, ‘kernel’: ‘rbf’}
0.972603 (0.025510) with: {‘C’: 1.5, ‘kernel’: ‘sigmoid’}
70 V. Chaurasia and S. Pal

Table 5 Results of tuning SVM on the scaled dataset (continued)

Best: 0.972603 using {‘C’: 0.3, ‘kernel’: ‘sigmoid’}

0.965753 (0.026623) with: {‘C’: 1.7, ‘kernel’: ‘linear’}
0.904110 (0.036536) with: {‘C’: 1.7, ‘kernel’: ‘poly’}
0.969178 (0.023781) with: {‘C’: 1.7, ‘kernel’: ‘rbf’}
0.969178 (0.023781) with: {‘C’: 1.7, ‘kernel’: ‘sigmoid’}
0.965753 (0.026623) with: {‘C’: 2.0, ‘kernel’: ‘linear’}
0.910959 (0.030793) with: {‘C’: 2.0, ‘kernel’: ‘poly’}
0.969178 (0.023781) with: {‘C’: 2.0, ‘kernel’: ‘rbf’}
0.972603 (0.025510) with: {‘C’: 2.0, ‘kernel’: ‘sigmoid’}

We can see the most precise design was SVM with a sigmoid kernel and a C estimation
of 0.3. The accuracy 97.26 % is apparently not as much as what LR could accomplish.

4.3 Ensemble methods

Another way we can improve the algorithm’s ability to execute on this issue is to use
integrated strategies. We will use a similar test solution as before for ten-fold
cross-validation. Because each integration algorithm relies on decision trees that are less
sensitive to information distribution, there is no data standardisation for this situation. See
Table 6.
Table 6 Output of evaluating algorithms

Algorithms Mean accuracy values Time (sec)

AB 0.5763 (0.0599) 0.04
GBM 0.9572 (0.0386) 0.04
RF 0.9575 (0.0344) 0.05
ET 0.9687 (0.0407) 0.06

Figure 6 Plots of ensemble algorithms comparison (see online version for colours)
 Machine learning algorithms using binary classification 71

We can see that both bagging procedures give solid precision scores in the low ‘90s (%)
with default designs. We can plot the accuracy scores over the cross validation folds. See
Figure 6.

4.4 Finalisation of model

From the results, LR is the strongest guarantee for the low complexity and stable model
of the dermatological dataset. In this section, we will end the model by preparing the
model on the entire training dataset, and make predictions for retaining the validation
dataset to confirm our findings. Part of the research result is that LR performance is better
when the dataset is normalised, so the mean of all features is estimated to be zero and the
standard deviation is 1. We can determine this from the entire training dataset and apply
equivalent changes to the input attributes in the validation dataset.
We can use more type of ensemble methods bucket of models and staking ensemble
techniques but since in this paper we used only homogeneous type of classifiers,
therefore we have used AB, GBM, RF and ET ensemble techniques.
We can see that we accomplish an exactness of about 99% on the held-out validation
dataset. A score that more and improved to our desires evaluated above amid the tuning
of LR. See Table 7.
Table 7 Output of evaluating SVM on the validation dataset

Accuracy score 0.9864

Confusion matrix  [11 0 0 0 0 0] 
 [0 11 0 0 0 0] 

 [0 0 13 0 0 1] 
 
 [0 0 0 4 0 0] 
[0 0 0 0 24 0]
 
[0 0 0 0 0 10] 

5 Discussion

In view of the results, we see that the proposed single classifier technology can produce
satisfactory results and outperform multi-model integration methods in disease prediction
(Chaurasia et al., 2018a, 2018b). Due to the complexity of the analysis and the need for
biopsy, auspicious and precise determinations are essential. Along these lines, improving
the accuracy of predictions through the application of computer-supported programs is
very helpful in treating diseases.
In the survey, we examined the six separately functioning classification models and
multi-model integration methods proposed in this paper. Six classifiers have been
established and promoted, which are commonly used for anticipated skin diseases. As
perceptually indicated, for SVM and LR, each classifier may rank behind the other on the
same dataset. A similar situation may occur on different classifiers, which indicates that
each technology has its own weaknesses compared to other technologies. It is this
perception that has inspired us to propose a procedure that combines various classifiers in
order to obtain proof of an increasingly accurate and fair arrangement. Our results on the
72 V. Chaurasia and S. Pal

dataset show that the six classifiers acting on the dataset alone have higher accuracy than
the multi-model integration method.
Moreover, a random check of the accuracy and standardisation accuracy of the
classifier shows that the prediction performance of the dataset displayed by a single
classifier is unstable. From the accuracy of the six classifiers, we selected SVM and LR
from the standardised datasets, which are very accurate. After adjusting the LR and SVM,
the accuracy will be improved. By going beyond to fit the output of the four classifiers,
our proposed technique will continue to prepare the weights for each classifier. In this
strategy, the implementation of the classifier with higher accuracy is more effective, and
the interference data of the classifier with lower accuracy is excluded. In this way, the
advantages of each classifier are fully considered and utilised, and better prediction
execution is obtained.
Additional tests were performed to confirm the expected predictions of better LR
performance findings after normalising the dataset. We can compute from the entire
training dataset and apply equivalent transformations to the input attributes in the
validation dataset. We need to come up with six unique classification models that require
higher computational costs. In order to overcome this limitation to a certain extent, we
applied a feature determination method in the data pre-processing stage, which can
significantly reduce the running time and improve the prediction accuracy in a similar
time. In general, the selection of features in the disclosure of imperative dermatomes and
pathological studies deserves more consideration.
Ensemble techniques are used to improve the results of the prediction of skin disease.
Four ensemble techniques are used AB, GBM, RF and ET. These techniques improve the
accuracy score up to a maximum of 99% in the case of ET ensemble method. Instead of
using these three techniques we can also used Bucket of models or stacking ensemble
techniques but it is future plan to test the predictions.

6 Conclusions

Dermatology is a disturbing and real medical problem worldwide. Although machine

learning strategies have become increasingly common for disease prediction, no one
technology can outperform all others. In this article, we show six different order models
and multi-model integration methods to handle the prediction of skin diseases. The
results show that differential expression analysis is very important to reduce the
dimensionality of the data and select effective data. Along these lines of thinking, the
accuracy of prediction can be enlarged and the calculation time can be greatly reduced.
The multi-model integration approach at that time used the expectations of many
different classification models as input. The classification technique can reduce the
generation errors and obtain more data by using the main organisation prediction as the
focus. Moreover, by using classification technology, it turned out to find an incredible
connection between the classifiers, in this way enabling ordering strategies to complete
better predictions.
 Machine learning algorithms using binary classification 73

References
Abdi, H. (2007) ‘Discriminant correspondence analysis’, Encyclopedia of Measurement and
Statistics, pp.270–275, Sage, Thousand Oaks (CA).
Altman, N.S. (1992) ‘The American statistician. An introduction to kernel and nearest-neighbor
nonparametric regression’, Journal of the American Statistical Association, Vol. 46, No. 3,
pp.175–185.
Amarathunga, A.A.L.C., Ellawala, E.P.W.C., Abeysekara, G.N. and Amalraj, C.R.J. (2015) ‘Expert
system for diagnosis of skin diseases’, International Journal of Scientific & Technology
Research, Vol. 4, No. 1, pp.174–178.
Badrinath, N., Gopinath, G., Ravichandran, K.S. and Soundhar, R.G. (2016) ‘Estimation of
automatic detection of erythemato-squamous diseases through AdaBoost and its hybrid
classifiers’, Artificial Intelligence Review, Vol. 45, No. 4, pp.471–488.
Bojarczuka, C.C., Lopesb, H.S. and Freitasc, A.A. (2001) ‘Data mining with constrained-syntax
genetic programming: applications in medical data set’, in Data Analysis in Medicine and
Pharmacology (IDAMAP-2001), a Workshop at Medinfo-2001, London, UK.
Chang, C.L. and Chen, C.H. (2009) ‘Applying decision tree and neural network to increase quality
of dermatologic diagnosis’, Expert Systems with Applications, Vol. 36, No. 2, pp.4035–4041.
Chaurasia, V., Pal, S. and Tiwari, B.B. (2018a) ‘Prediction of benign and malignant breast cancer
using data mining techniques’, Journal of Algorithms & Computational Technology, Vol. 12,
No. 2, pp.119–126.
Chaurasia, V., Pal, S. and Tiwari, B.B. (2018b) ‘Chronic kidney disease: a predictive model using
decision tree’, Int. J. Eng. Res. Technol., Vol. 11, No. 11, pp.1781–1794.
Güvenir, H.A. and Emeksiz, N. (2000) ‘An expert system for the differential diagnosis of
erythemato-squamous diseases’, Expert Systems with Applications, Vol. 18, No. 1, pp.43–49.
Güvenir, H.A., Demiröz, G. and Ilter, N. (1998) ‘Learning differential diagnosis of
erythemato-squamous diseases using voting feature intervals’, Artificial Intelligence in
Medicine, Vol. 13, No. 3, pp.147–165.
Hartmann, A. (2016) ‘Back to the roots-dermatology in ancient Egyptian medicine’, JDDG:
Journal der Deutschen Dermatologischen Gesellschaft, Vol. 14, No. 4, pp.389–396.
https://scikit-learn.org/stable/modules/feature_selection.html.
Idoko, J.B., Arslan, M. and Abiyev, R. (2018) ‘Fuzzy neural system application to differential
diagnosis of erythemato-squamous diseases’, Cyprus Journal of Medical Sciences, Vol. 3,
No. 2, pp.90–97.
Lekkas, S. and Mikhailov, L. (2010) ‘Evolving fuzzy medical diagnosis of Pima Indians diabetes
and of dermatological diseases’, Artificial Intelligence in Medicine, Vol. 50, No. 2,
pp.117–126.
Maghooli, K., Langarizadeh, M., Shahmoradi, L., Habibi-koolaee, M., Jebraeily, M. and
Bouraghi, H. (2016) ‘Differential diagnosis of erythmato-squamous diseases using
classification and regression tree’, Acta Informatica Medica, Vol. 24, No. 5, p.338.
Mukhopadhyay, A.K. (2016) ‘Dermatology in India and Indian dermatology: a medico-historical
perspective’, Indian Dermatology Online Journal, Vol. 7, No. 4, p.235.
Nanni, L. (2006) ‘An ensemble of classifiers for the diagnosis of erythemato-squamous diseases’,
Neurocomputing, Vol. 69, Nos. 7–9, pp.842–845.
Perriere, G. and Thioulouse, J. (2003) ‘Use of correspondence discriminant analysis to predict the
subcellular location of bacterial proteins’, Computer Methods and Programs in Biomedicine,
Vol. 70, No. 2, pp.99–105.
Polat, K. and Güneş, S. (2009) ‘A novel hybrid intelligent method based on C4.5 decision tree
classifier and one-against-all approach for multi-class classification problems’, Expert Systems
with Applications, Vol. 36, No. 2, pp.1587–1592.
74 V. Chaurasia and S. Pal

Übeyli, E.D. (2009) ‘Combined neural networks for diagnosis of erythemato-squamous diseases’,.
Expert Systems with Applications, Vol. 36, No. 3, pp.5107–5112.
Übeyli, E.D. and Doğdu, E. (2010) ‘Automatic detection of erythemato-squamous diseases using
k-means clustering’, Journal of Medical Systems, Vol. 34, No. 2, pp.179–184.
Übeylı, E.D. and Güler, I. (2005) ‘Automatic detection of erythemato-squamous diseases using
adaptive neuro-fuzzy inference systems’, Computers in Biology and Medicine, Vol. 35, No. 5,
pp.421–433.
Verma, A.K., Pal, S. and Kumar, S. (2019a) ‘Prediction of skin disease using ensemble data mining
techniques and feature selection method – a comparative study’, Applied Biochemistry and
Biotechnology, Vol. 190, No. 2, pp.1–19.
Verma, A.K., Pal, S. and Kumar, S. (2019b) ‘Classification of skin disease using ensemble data
mining techniques’, Asian Pacific Journal of Cancer Prevention, Vol. 20, No. 6,
pp.1887–1894.
Xie, J. and Wang, C. (2011) ‘Using support vector machines with a novel hybrid feature selection
method for diagnosis of erythemato-squamous diseases’, Expert Systems with Applications,
Vol. 38, No. 5, pp.5809–5815.
Zhang, X., Wang, S., Liu, J. and Tao, C. (2018) ‘Towards improving diagnosis of skin diseases by
combining deep neural network and human knowledge’, BMC Medical Informatics and
Decision Making, Vol. 18, No. 2, p.59.
Zhou, H., Xie, F., Jiang, Z., Liu, J., Wang, S. and Zhu, C. (2017) ‘Multi-classification of skin
diseases for dermoscopy images using deep learning’, in 2017 IEEE International Conference
on Imaging Systems and Techniques (IST), IEEE, October, pp.1–5.