CHAPTER 2

LITERATURE REVIEW
In this digital era, due to the rapid advancement in science and technology, enormous
amounts of healthcare data are generated from healthcare applications through variable
technologies like embedded systems, smart health devices, and computers. Machine learning
algorithms are emerging as efficient technologies in the health care arena, which can be used
efficiently in the early diagnosis of disease by identifying meaningful patterns from the data
[1]. In recent years, addressing the issue of chronic disease has become a global challenge. It
takes hours for a physician to identify a chronic disease efficiently. The traditional method of
manually diagnosing diseases has a number of limitations such as it is prone to human errors,
expensive and time-consuming.

Machine learning algorithms can be combined with feature selection algorithms to

circumvent these restrictions. A successful integration of machine learning and feature
selection based frameworks can aid physicians in forecasting the risk of chronic diseases at an
early stage. The chapter discusses the numerous issues that persist in traditional approaches
and the motivational objectives for resolving these issues through the proposed research work.

The chapter provides a comprehensive overview of the research work that has already
been carried out in the areas of early detection and prediction of chronic diseases such as CKD,
CVD, and CDM. As discussed in the introduction section, inter-connected chronic disease has
become a serious threat to human life, deteriorating the quality of human life. These diseases
are considered as a global burden of disease as almost $5 trillion of amount will be spent yearly
on treating these diseases across the globe[35]. Heart disease has become the number one
deadliest chronic disease as 71% of death in a developing country is due to one or the other
heart diseases. India is considered as the world's capital of diabetes, as over 50 million people
are suffering from these metabolic disorders. At least 10% of the global adult population is
affected due to CKD[36]. According to WHO, globally, 60% of the chronic disease burden will
be in developing countries. One of the remedies would be predicting and preventing the disease
in the initial stage itself. Due to digitalization and a new revolution in the diagnostic field, a
massive volume of clinical data is now available in electronic form. The mathematical model
can be integrated with healthcare informatics and diverse algorithms to increase the quality of
healthcare services. Machine learning algorithms are emerging as a useful solution in the
healthcare domain. The efficiency of the ML algorithm depends on how effectively the feature
27
selection process is carried out[37]. Hence feature selection is considered as part and parcel of
machine learning techniques. The proposed research work seeks to apply efficient feature
selection algorithms in combination with machine learning algorithms to early predict the
chronic diseases CKD, CDM, and CVD.

This chapter aims to compile a list of the various state-of-the-art works proposed in the
early detection of chronic diseases using various machine learning algorithms. The chapter is
subdivided into two subsections. The first section discusses several feature selection methods
for selecting the best features from the chronic disease dataset. The second section provides an
overview of numerous research works that use machine learning techniques to predict chronic
diseases at their premature stage.

2.1 Traditional Machine Learning Algorithms

Machine learning is a term that refers to an algorithm's ability to learn and improve
based on its experience without requiring explicit programming. Predictive intelligence tasks
can be achieved using several machine learning techniques, which are classified essentially as
supervised or unsupervised algorithms. Figure 2.1 below shows the various types of machine
learning algorithms. As represented in the figure, the supervised machine learning techniques
are utilized for problems related to classification and regression. In supervised learning, the
labeled input F is mapped to output using equation (2.1) below.
𝐷 = {(𝐹 , 𝑂 )} 𝑁 [38] (2.1)
𝑖 𝑖 𝑖=1

where D represents the training set, and N represents the cardinality associated with training
instances. In each example, 𝐹𝑖 is the dimensional vector representing the feature instance, and
𝑂𝑖 represents the class label. The models are trained with labeled examples in this case. The
data is fed into the system as input as a test set. One of the most important tasks is to compare
the actual output and what is desired. The type of job being performed can further separate
supervised learning into two categories: classification and regression, which are subsets of
classification[39].
In the classification task, the output variable Oi is nominal or categorical, whereas, in
the regression task, the output variable Oi is real-valued or scalar in nature. The Support vector
machine(SVM), Naive Bayes, K nearest neighbor algorithms(KNN), and Decision Tree are

28
 most often used machine learning algorithms for classification tasks. The algorithms logistic
regression and linear regression are frequently used for regressive problems. The goal of the
unsupervised learning method is to uncover interesting patterns in a large amount of data. The
unsupervised algorithms can be represented without the class label using the equation (2.2).
𝐷 = {(𝐹𝑖)}𝑖=1 [40] (2.2)

Association analysis is the unsupervised algorithm that can be used to discover interesting
patterns from huge databases. Clustering is the process of identifying and grouping similar data,
such as putting clients with similar purchasing habits together. The K-means algorithm,
Hierarchical algorithm, and Expectation maximum algorithm are three of the most often used
unsupervised learning algorithms. Market basket analysis is performed in combination with the
development of analysis rules that define the data in which consumers who purchase product x
also prefer to purchase product Y. In machine learning, there may be a situation in which just a
little quantity of data is labeled out of a substantial quantity of data that has been collected (x),
which is known as a semi-supervised state (y). The section below explains the properties of a
few frequently used machine learning algorithms.

Figure 2.1: Types of Machine Learning Algorithm

29
Decision Tree Classifier

In machine learning, a decision tree classifier is a tree-based classifier that can predict
continuous or discrete values depending on the data types in the training set. The decision tree
used to predict continuous values is a regression tree, whereas the decision tree used to predict
discrete values is a classification tree[41]. Test conditions are applied to each node, starting with
the root node and progressing to the relevant branches based on the conclusion. The leaf node
is related to the class labels that have been assigned to it. This algorithm has an issue with the
construction of an optimal decision tree. The tree must be formed by selecting the most
appropriate attribute for each step of the algorithm's execution. There are numerous ways to
select the most advantageous features from a given dataset. Information gain is considered as
one such parameter to construct a decision tree[42]. The information gain uses entropy as a
parameter to obtain the information about the class and tuples of the dataset. The entropy of
given dataset D with n distinct value of given class can be calculated using equation (2.3).
𝐸𝑛𝑡𝑟𝑜𝑝𝑦 = 𝐼𝑛𝑓𝑜(𝐷) = − ∑𝑛𝑖=1 𝑞𝑖log2(𝑞𝑖) [43] (2.3)

Where 𝑞𝑖 represents the probability of a given instance having class value as i. When
constructing the decision tree, the attribute that provides the greatest amount of information
gain is used as a split attribute, and this process continues.

Support Vector Machine

It is a supervised learning technique that uses hyperplanes to distinguish between

distinct classes of data. To differentiate between the classes in n-dimensional space, the goal is
to construct a hyperplane with the greatest possible margin of error. The greater the margin of
error, the lower the error in the classification outcome. SVM uses the support vector and the
margin to determine the best hyperplane. SVMs can be used for both regression and
classification tasks, depending on the implementation task. SVM algorithm is implemented
using a technique called as kernel trick, which will transform the data in low dimensional space
to the higher dimension. The separating hyperplane can be represented using equation (2.4)
below.

𝑊 · 𝑓 + 𝑏𝑖 = 0 [44] (2.4)

30
where W indicates the weighted feature vector and bi represents the bias. The several kernel
types of SVM are the radial function (RBF), linear kernel, and polynomial kernel. The linear
kernel is the most common type of kernel used in SVM, where data can be separated using
unique lines. The generalized form of the linear kernel is a polynomial kernel where the data
are separated in nonlinear space. The RBF kernel is used if the data needs to be separated into
infinite-dimensional space. The advantage of using SVM is that it is highly accurate, which
works well with high-dimensional data. SVM classifier uses a limited amount of memory as it
uses only the subset of the training instances. But the drawback with SVM is, it will consume
more time during the training phase with the large dataset. The SVM classifier will not work
well with the overlapped classes.

Naive Bayes Classifier

In probability theory, the Naive Bayes classifier is a probabilistic classifier which is based on
the Bayes theorem. The algorithm assumes that features in the given class are independent
of each other. As per the Bayes theorem, an event concerning posterior probability can be
represented using the equation (2.5).

𝑃(𝑃𝑟 𝐸) = 𝑃(𝐸 𝑃𝑟)·𝑃(𝑃𝑟) [45] (2.5)

𝑃(𝐸)

31
(𝑃𝑟 𝐸) is the hypothesis of posterior probability, P(Pr) indicates the prior probability, and
𝑃(𝐸 𝑃𝑟)is likelihood, and P(E) is evidence. Based on the data distribution, there are different
types of Naive Bayes classifiers to choose from, including multinomial Naive Bayes, Gaussian
Naive Bayes, and Bernoulli Naive Bayes. Whenever the values associated with a class are not
discrete, the classification is determined using the Gaussian Naive Bayes algorithm.
Multinomial Naive Bayes is the preferable method if the events are created concerning a certain
frequency. If the features are independent of each other where Boolean variables describe the
event, the Bernoulli Naïve Bayes models are applied. The Naïve Bayes classifier performs very
well under the circumstances of unrealistic assumptions about the data. In addition to being
simple to develop, robust, and fast, the method may be applied to a wide range of applications,
including disease prediction and making treatment options for diseases. The Naive Bayes
algorithms can work with lesser amounts of data and deal with both discrete and continuous
data types. Because the algorithm is founded on the concept of probability theory, it is both
faster and more accurate at making real-time forecasts.

K-Nearest Neighbours Classifier

K-Nearest Neighbours is a classifier based on a lazy learner that does not generate the
model until the test date arrives. The KNN calculates the proximity between the train and test
instances to discover its K nearest neighbors. After gathering the nearest neighbors, the forecast
will be judged by a simple majority of these. If the value of k is 10, then the prediction will be
based on a total of ten nearest neighbors. The concept of Euclidean distance is used to calculate
the distance between the closest neighbors, as indicated in equation (2.6).
𝑑(𝑠, 𝑟) = √(𝑠1 − 𝑟1)2 + (𝑠2 − 𝑟2)2+. . . +(𝑠𝑛 − 𝑟𝑛)2 [46] (2.6)
In equation (2.6) above, s and r represent the two different feature instances. In the KNN, the
value of k must be priorly known. Once the value of k is known, the decision value with the
unknown data is found using averaging or polling method. The polling method considers the
class value of k nearest neighbors using the majority method. The averaging method considers
the average decision value with k nearest neighbors. The algorithm traverses the entire dataset

32
to determine the distance between each training observation. The KNN algorithm is
nonparametric as it makes no assumptions about the data. The KNN training phase consumes
minimal time because there is no exclusive training phase as it uses all the training data in the
testing phase. The algorithm works well with the nonlinear data because it does not make any
assumptions about the data. The algorithm is versatile and simple. The algorithm is very
expensive because it consumes more memory space since it uses all the training during the
testing phase. It is sensitive to outliers and slower with the computational complexity of big N.

The primary difficulty with the KNN method is determining the value of k. The greater
the value of k, the less noise will be generated during the prediction or classification process.
The best value of k can be chosen by applying the heuristic method. If the model must cope with
a huge amount of data, the KNN technique is more error-prone. The algorithm will be effective
if the dimensional reduction process is applied before KNN is applied. The dimensional
reduction process will overcome the drawback of the curse of the dimensionality problem. The
KNN algorithm can also be used for the regressive task to estimate the data related to continuous
variables. To perform regressive tasks, KNN uses the weighted-average method. The KNN
algorithm can also be used for binary classification tasks such as disease prediction. For the
binary classification task, the KNN algorithm uses a bootstrap method to find the optimal value
of k.

Logistic and Linear Regression

The logistic regression model is based on statistical theory and is used to determine the
relationship between independent and dependent variables. Although the model is mostly
utilized for binary classification problems, it can also perform multiclass classification.
Logistic regression is one of the simplest algorithms available. It can be used for various
classification tasks such as disease prediction and spam identification[47]. Logistic regression
techniques include binary logistic regression, ordinal logistic regression, and multinomial
logistic regression. Binary logistic regression requires binary dependent variables, i.e. either '0'
or '1'. Multinomial regression is mostly used for multiclass classification. The dependent
variable is unordered and can be classified as type A, B, or C. In the ordinal regression, the
dependent variables can be categorized based on the quantitative significance, such as "poor",
"good", and "bad". The binary logistic regression algorithm uses sigmoid as an activation

33
 function to map the prediction in terms of probabilities using the equation (2.7) below.

Sig(δ) [48] (2.7)

where 𝛿 represent the input variable, e is a natural log, and the value of 𝑆𝑖𝑔(𝛿) must be binary
variable '0' and '1'.For the discrete class, the decision boundary can be set as 𝑝𝑟𝑜𝑏 ⩾ 0.5, 𝑐 = 1
𝑎𝑛𝑑 𝑝 < 0.5, 𝑐 = 0 where c indicates the class label. In case the threshold ranges between
0.5 and 0.7, the observation is classified as positive, or else it will be classified as negative. The
algorithm uses cross-entropy or log loss as a cost function. A variety of optimization strategies
can be used to reduce the cost. Gradient optimization is an example of this type of optimizer.
In multinomial regression, softmax is utilized as the activation function, with the vectors
serving as the input function. In Linear regression, the linear approach is used in modeling,
which forms the relationship between independent and dependent variables. The algorithm
looks at the statistical relationship between these two variables. Like Logistic regression, linear
regression comes in different forms like simple regression and multivariable regression. The
simple regression can be denoted using the slope-intercept form as Y=bx+c where b and c are
variable, x is input to the algorithm, and Y indicates the prediction outcome. The multivariable
regression model establishes the association between multiple explanatory variables. The
approach uses the mean square error (MSE) as a cost function to calculate the square of the
disparity between actual and predicted values. Gradient descent can be used to minimize the
cost of the MSE.

Ensemble Learning Algorithms

Ensemble learning is a machine learning technique in which a prediction job is carried out using
numerous learning algorithms. The learners which are used to perform ensemble are called as
base learners[49]. There are three popular ensembling methods. Bagging, Boosting, and
Ensemble. The Bagging approach uses an average of the prediction made by the different
classifiers to generate the final prediction result. In Boosting method, the weighted voting- based
approach is applied to obtain the prediction result. The ensembling technique can boost the
capacity of weak learners. The Bagging approach is analogous to the medical diagnosis made
by multiple doctors where the majority voting concept is used. The Bagging approach is also
known as bootstrap aggregation. The Boosting is analogous to weighted diagnoses by consulting

34
several doctors, which uses a series of classifiers to perform iterative learning. After each
classifier Ci learns the weights, the subsequent classifiers Ci+1 will pay more attention to miss
classified instances by Ci, and the final classifier C* combines the vote of the individual
classifier. The boosting approach is more accurate compared to the bagging approach. The two
important algorithms listed under Boosting approach are Gradient Boosting and Extreme
gradient Boost algorithm. The optimized version of the gradient boosting algorithm is XG Boost
which is faster and more accurate than gradient boost. The reason is, XG Boost applies the
regularization concept to the complex model through L1 and L2. The algorithm can able to deal
with sparse data. For faster processing, XG Boost can use multiple cores on GPU. XG Boost is
a cache-aware algorithm where it optimally uses the hardware. It uses a noncontinuous memory
block to get the gradient statistics using the row index.

Neural Network

An algorithm inspired by nature and meant to mimic the operation of the human brain is referred
to as a neural network in computer science. The three layers that constitute a neural network are
the input, output, and hidden layers. It has numerous hidden layers and a nonlinear activation
function, making it particularly effective. The method takes the input through the input layer,
passes it to the numerous hidden layers for processing, and then shows the result through the
output layer, which is a layer that is concealed from view. The activation function utilized in the
neural network will aid in the model's ability to forecast nonlinearly in the future. The sigmoid,
tanh, softmax, and ReLu activation functions are some of the most commonly applied in neural
networks. The main reason behind the usage of the sigmoid function is to transform the network
input between the range 0 and 1. The tanh activation function will map the network input
between the range (-1,1) and Rectifier linear unit(ReLu) map the input in the given interval [0,∞]
where all negatives are set to zero. The loss or cost function is used to determine the goodness
of the model with possible parameters. The MSE and Cross entropy loss functions are two of
the most commonly utilized loss functions in neural network models.

35
 The perceptron is the fundamental unit of a neural network's operation. In a neural
network, the parameters weight and bias are regarded to be learnable parameters. Each of the
neural nodes is associated with a particular input, as well as with the relevant weight and bias
value for that input. Weight and bias are changed in stages as training progresses to ensure that
the intended outcome is achieved. Neural networks are available in several configurations,
including the feed-forward neural network, the recurrent neural network, multilayer perceptron
learning, and convolution neural networks[50].

It is believed that a low bias value indicates that the implemented neural network makes
fewer assumptions about the output, whereas a high bias value suggests that the implemented
neural network makes a greater number of assumptions about the output. It is predicted that the
input with a lower weight will have no effect on the output but that the input with a higher
weight will have more effects on the prediction. The most basic sort of neural network is a
feed-forward neural network in which information only flows in one direction, from the input
layer to the hidden layer and then to the output layer. The network is devoid of any loops or
cycles that could be identified. The perceptron is the most fundamental type of neural network,
in which inputs are supplied directly to the output node. In a recurrent neural network
(RNN)[51], the input is transferred from the last unit, allowing activation to flow around the
loop and into the next unit. The internal memory of RNN keeps track of the data from
previously encountered patterns.

When it comes to document or sentence classification challenges, the RNN is most

commonly utilized. A convolutional neural network (CNN) is a deep learning model used in
artificial intelligence with several hidden layers. The CNN is mostly used for image and video
analysis, as well as for recommendation systems. The CNN has numerous layers in it, such as
convolutional, pooling, fully connected, and output layers. The convolution layer is the
foundation of the CNN algorithm. The convolution operation is applied to the input layer, which
can learn the feature and classify the data. To limit the number of resources required to process
the data, the pooling layer selects only the most significant characteristics from the input.

36
 It is possible to configure the pooling layer in two different ways: as the maximum
pooling layer or as the average pooling layer. When using the max-pooling layer, the maximum
amount of information can be extracted from the area of the picture that has been covered by
the kernel function in the preceding layer. This layer is in charge of extracting an average value
from the data set covered by the kernel function for each fraction of the data set covered by the
kernel function.
A "completely linked layer" is one in which each neuron from one layer is linked to every
neuron in the next layer. It is now widely acknowledged that deep learning may be a viable
method for predictive analytics[52]. It can be applied for classification, regression, and
clustering tasks. The process of learning on the considered data is performed by many neurons,
where each starts with a different weight than the others. It is possible to utilize deep learning
for various applications, including analyzing multimedia data (such as images, audio, and
video), natural language processing, and many more. Nature is the inspiration behind the
creation of the set of algorithms which is known as a nature-inspired algorithm. The
characteristics of these algorithms are used for the optimization problem of predictive
analytics[53]. The advantage of a neural network-based model is, it is robust and provides good
accuracy. But the problem with neural networks is the model needs to be trained with extensive
training examples to obtain good classification accuracy.
Clustering Algorithm

Clusters are objects that are grouped based on their resemblance. Since clustering
requires no prior knowledge, it is characterized as unsupervised learning. For the clustering
method to work, unlabeled data or decisional characteristics are required. A clustering method
may be classified into three types: hierarchical clustering, partition-based clustering, and
density-based clustering[54]. One of the advantages of the clustering algorithm is that outliers
can be easily tracked. The partition-based clustering algorithm forms the clusters by generating
partitions, and then clusters are evaluated by using some criteria. The algorithms like k-means,
k medoids, and CLARANS belongs to the category of partition-based clustering. The
hierarchical-based clustering algorithm generates the clusters by applying the hierarchical
decomposition on the set of objects. The set of criteria is used while decomposing the object
based on the hierarchy. The algorithms like Diana , BIRCH , and Agnes belong to the category
of hierarchical clustering. If the clusters are formed based on the connectivity and density of
37
the objects, it is known as density-based clustering. DBSCAN and optics are two techniques
that fall under the category of density-based clustering. The objects' granularity structure can
also be used to construct clusters. As a result, it falls within the umbrella of a Grid-based
strategy. STING, CLIQUE, and wave clusters are examples of typical Grid-based approach
methods. The K means clustering algorithm is one of the most often used clustering algorithms.
The algorithm works as follows. Initially, the items are subdivided into k subsets. The seed
point for each item is calculated using the cluster's centroid. The items are allocated to the
cluster depending on their proximity to the closest seed point. The process is continued until
no change occurs in assigning all objects to the cluster. Other subcategories of clustering
approaches include model-based clustering, frequent pattern-based clustering, constraint-based
clustering, and link-based clustering.

The clustering algorithm must satisfy certain requirements while forming the
clusters. The clustering algorithm must be scalable where it should be able to deal with a large
set of objects. The clustering algorithm must be able to deal with a variety of attributes like
numeric attributes, interval scaled attributes, etc. One of the important characteristics of the
clustering algorithm is that it is insensitive to the order of the input while forming the cluster.
The clustering algorithm must be capable of working with high-dimensional data like clinical
data. The clustering algorithm must be able to form the cluster under various conditions and
constraints.

2.2 Feature Reduction with High Dimensional Data

The data which emerges from the clinical test contains numerous features which
illustrate the properties and behaviors of the particular disease, known as multidimensional
data. As the size of the dimension increases, it adds noise and redundancy to the dataset, which
further reduces the accuracy of the prediction[55]. As part of the preparatory step, data is pre-
processed to remove noise, outliers, missing data, and duplicates from the data set, among other
things. When it comes to machine learning, one of the most critical strategies is the feature
selection or reduction process. The efficiency of the machine learning model will be diminished
when an excessive number of irrelevant features are included in the dataset. By retaining only
the relevant features, the performance of the machine learning classifier can be improved. The

38
data is subjected to undergo the feature selection process immediately following the pre-
processing stage and prior to the building of the machine learning model. Feature reduction
techniques are frequently classified into two categories: feature transformation and feature
selection. The feature transformation procedure reduces the dimension of a feature by
converting it from a different form to another while retaining its originality. The principal
component analysis (PCA) algorithm is a good example of feature transformation[56]. Rather
than retaining all the features, the selection process selects only those features that positively
impact the prediction result.

The search strategy and classifier environment are used further to classify the
feature selection process. The feature selection process belongs to the category of search
procedures are Exhaustive feature selection, Heuristic-based feature selection, and Meta
heuristic-based feature selection. When using exhaustive feature selection, all potential subsets
of the features are generated first, and then each subset is examined until the best subset of the
features is found. However, the disadvantage of this strategy is that it takes a long time to
generate the subsets and can only be used for datasets with minimal features. To obtain the
bestfeatures, the heuristic-based feature selection technique uses the concept of
randomness[57]. The advantage of the heuristic method is that it takes less time to choose the
features because the algorithm's complexity is polynomial. Heuristic feature selection
techniques such as forward selection and backward elimination are good examples of this. To
derive the optimal features, the metaheuristic algorithm uses the concept of total randomness,
where the answer is independent of the problem. The meta-heuristic approach is used to find a
solution to 'n' issues, resulting in a near-optimal solution for any problem. The best examples
of feature selection utilizing metaheuristics are evolutionary algorithms like genetic algorithms
and particle swarm algorithms. To simplify the classification process, various feature selection
algorithms are applied, including the filter, wrapper, and hybrid methods.

The filter-based strategy does not employ a classifier to choose features; instead, it uses
the statistical idea to generate suboptimal features. Although the filtered method is faster, it is
inferior to the wrapper method in terms of quality. The chi-square test method is a great
example of filter-based feature selection. When picking features, the wrapper technique
employs a classifier. The classifier will be given with the specified features as input from the

39
dataset. The wrapper technique selects the fitness of the features by determining the accuracy
attained with the classifier. The wrapper method consistently outperforms the filter method in
terms of accuracy, although it is slower than the filter-based feature selection process. The
wrapper method can also be combined with a heuristic or metaheuristic approach. The filter
and wrapper selection methods can be combined as the hybrid feature selection algorithm. If
the dataset is multidimensional in form, applying a hybrid feature selection technique is
recommended to obtain the optimal result with the machine learning algorithm as it is imposed
on the minimized dataset. The data size is reduced in the hybrid approach by applying the filter
and wrapper method. The outcome of the filter method is given as input to the wrapper method,
where it will assess the credibility of the features by applying the possible hypothesis using a
machine learning algorithm.

2.3 Feature Selection Algorithm and Application in Chronic

Disease Prediction

It is common to consider feature selection as a pre-processing step in machine learning

to reduce the size of the input dataset. Identifying the most relevant risk factors of a specific
disease is critical for the early prediction of chronic disease[58]. The disease datasets are the
result of a various diagnostic test, which is usually multidimensional in form. All the features
in the data set may not be essential from the perspective of the prediction of the disease. By
identifying relevant features, redundant and unnecessary noisy attributes can be removed from
the clinical dataset, which further produces quick and quality results. By applying the feature
selection, the comprehensibility of data can be enhanced further. This facilitates better data
visualization, reducing the time required to train the machine learning algorithm [59]. In recent
days, many FS algorithms have been put forward in healthcare to get the optimal features from
the diagnostic dataset for predicting diseases like CVD, CDM, and CKD.

2.3.1 Feature Selection Algorithm for the CVD Data

The following section provides an overview of the various work accomplished in the
early prediction of CVD using different feature selection algorithms.

Ottom et al. [60] have proposed an Accumulated feature rank-based approach for CVD

40
prediction. The proposed algorithm selected the features from the dataset based on their rank.
The rank was assigned to the different features of the dataset using the concept of Gain Ratio.
All the features with a rank<1 were considered as non-optimal. Hence they have been removed
from the dataset. The suggested work successfully removed four features from the considered
dataset, which were found irrelevant and redundant. The disease was predicted using the
proposed technique with 83% accuracy. Wosiak et al.[61] selected the most prominent features
from the CVD dataset using Reversed Correlation technique. The algorithm selected the best
features by calculating the inverse correlation coefficient between the number of instances and
the features. The experiment was conducted on three real data sets of CVD cases. The proposed
method selected a total of 14 features out of 30 and achieved an accuracy of 79%. Perez-Riverol
et al.[62] applied a Quasi constant-based feature elimination technique for the prediction of
CVD. The Quasi constant eliminated 99.9% of the traits with low variance that were otherwise
comparable. The experiment was conducted using the Arrhythmia dataset of the UCI repository
and achieved an accuracy rate of 75.6%.

Amin et al.[63] have proposed Combinational feature selection for the prediction of
CVD. The CVD dataset was collected from the UCI repository. The data set had 303 instances
with 76 attributes in it. The Rapid miner studio was used to conduct the experiment. The
algorithm formed a Subset of combinational features where each subset had three attributes.
Thus a total of 8100 feature combinations were formed, which are tested for accuracy. The
performance of each of the feature combinations was tested. The algorithm achieved an
accuracy of 78.20% using the proposed approach. Awan et al.[64] used a two-sided chi-square
test to choose the features from the CVD dataset, which they found effective. One of the
primary goals of the study was to find clinically important variables from the CVD dataset that
were connected with the death of heart failure patients. The features whose p<0.05 were
retained. The algorithm achieved an AUC rate of 0.52 and a specificity of 62.5%. Joshi et
al.[65] have selected features from heart disease using entropy partition-based feature selection
method combined with clustered decision tree. The features were partitioned based on the
entropy value. The classifier performance for each entropy partition was calculated. Hence
feature dimensions were reduced using the entropy-based partition method without
compromising the prediction accuracy. Takci et al.[66] have applied a fisher filter-based
algorithm on the CVD data. The fisher score was calculated by finding the difference between
the expected and observed values. The main objective was to select the features with

41
maximized information with minimized variance. The experiment was conducted by using a
stat log dataset. The dataset had 270 records with 13 features in it. The algorithm selected eight
critical features, and the algorithm achieved an accuracy of 83.3%. Haq et al. [67] used the
mrMR algorithm to extract optimal features from CVD data. The proposed algorithm selected
features which will influence the target label. The features obtained with the proposed scheme
were minimal redundant and were relevant to the class label. The pairwise redundancy was
computed by checking the features one at a time in each cycle. The algorithm selected a total
of 6 features out of [Link] Minh et al.[68]developed a novel technique for selecting latent
attributes from CVD data. The best features were determined based on ranks and weights using
graph weighting in conjunction with probabilistic latent semantic analysis (PLSA). The SVM
classifier was used in conjunction with the approach for selecting infinite latent features. The
algorithm predicted the disease with an 89% accuracy. Xu et al.[69] introduced a CFS Subset
Evaluation approach in conjunction with on Best-First-Search to minimize the dimensionality
of a CVD dataset. The algorithm had four modules. The first phase was subset generation which
was responsible for the generation of the subsets. The second module was subset evaluation to
evaluate the generated subset by considering the merit of the features. The third phase was
search termination to terminate the process as stopping criteria, and the last phase was result
verification to validate the result. The CFS-oriented feature extraction technique was used in
association with the random forest algorithm to make early disease predictions utilizing CVD
data. The algorithm attained an accuracy of 91%.

Liu et al.[70] have applied Selection via Unique Relevant Information (SURI)
algorithm for the CVD prediction. The SURI assigned weightage to each feature according to
the scoring function. The relevant information was further divided as unique, relevant
information (URI) and overlapped relevant information. Thus, features with a mutual
information value greater than zero were considered as unique features. The algorithm selected
a total of 7 unique features out of 13 and achieved an accuracy of 83%. Wang et al.[71] have
implemented a Kernel Canonical Correlation Analysis (KCCA) based algorithm in
combination with mutual information (MI) algorithm for the CVD prediction. The strategy
mentioned above was designed to eliminate joint redundancy between candidate features and
to boost the relevance of target class labels and feature candidates. The Cholesky kernel matrix
approximation was also used to optimize the constructed algorithm's efficiency. The KCCA
was compared to typical feature selection algorithms in terms of productivity. The suggested

42
KCCA algorithm performed better than the previous techniques as per the experimental results.

Dhananjay et al.[72] implemented a novel CatBoost(CB) feature ranking model to

extract pertinent features from a CVD disease dataset. The CB algorithm was further
synthesized through the Prediction value change strategy. The results of the experiments
showed that the proposed technique was more accurate in its predictions than existing
approaches. Jayaraman et al.[73] proposed a bio-inspired hybrid technique that combined
cuckoo search with particle bees and neural networks to choose the features from a CVD
dataset. The optimization task was performed using a particle bee optimizer. The selected
features were classified using a neural network algorithm. Kamel et al.[74] have developed a
grasshopper optimization algorithm to retrieve the important features from the CDM
[Link] experiment was conducted in MATLAB software. The prediction ability of the
filtered features was examined using the SVM algorithm. The proposed approach obtained
promising results while predicting the disease.

Usman et al.[75] have created two methods for feature selection from the CVD data set
that were related but distinct from one another: the cuckoo search algorithm and the cuckoo
optimizer algorithm. The algorithms were designed following the concept of the filter
approach. When compared to the cuckoo optimizer algorithm, the experimental findings
showed that the Cuckoo search method performed significantly better. Jothi Prakash et al.[76]
have proposed an enhanced evolutionary feature selection technique to identify significant
traits from the CVD dataset. The primary objective of the work was to increase the prediction
accuracy. A genetic algorithm was used in conjunction with LDA to optimize the process.
Validation of the results was performed using a variety of metrics, including precision,
accuracy, f1 score, and recall. The experimental results indicated that the algorithm performed
better than the existing techniques. Saqlain et al.[77] have applied a statistical approach based
on Fisher scores to extract the optimal features from a CVD data set. A fisher score and a
correlation coefficient score were assigned to each feature. The approach was evaluated using
a variety of metrics such as sensitivity, specificity, and F measure. The proposed approach was
found 82.17% accurate. Yekkala et al.[78] have proposed a hybrid approach combining rough
set theory and random forest algorithms to select features from CVD data. For feature selection,
a rough set theory-based algorithm was applied. The implemented algorithm helped in
distinguishing between features that were redundant and irrelevant. The random forest
algorithm is a decision tree-based method that generates many decision trees, and then the
43
voting strategy will be used to select the best one. The proposed hybrid technique achieved
better accuracy than the existing algorithms. Sharma et al.[79] have applied an updated
frequency-based bat algorithm to select the informative features from CVD data. The voting
based classifier was used to make predictions on filtered data. The experiment was conducted
using the dataset extracted from the UCI repository[80]. The believability of the acquired result
was assessed utilizing the parameters such as accuracy, precision, recall, and [Link]
Christo et al.[81] proposed Correlation Feature Selection(CFS) algorithm in blend with
bioinspired algorithms to accurately classify the CVD. Initially, Bioinspired Algorithms such
as Lion Optimization, Differential Evolution, and Glow-worm Swarm Optimization algorithms
were used to assess the fitness of the features. Each bio-inspired algorithm will generate the
subset of the features by considering the fitness factor. Finally, the concept of CFS was applied
to extract the informative features from these feature subsets. The advantage of the proposed
method is that it can properly deal with both redundant and relevant features. On the other hand,
the technique suffers from a lack of interaction between the search space and the hypothesis
[Link]-Zoubi et al.[82] applied Salp swarm-based algorithm to predict cardiac disease. The
advantage of this method was that the algorithm was heuristic in nature, making it adaptable to
almost any problem. The technique, however, has the drawback of being sluggish to
convergence. Khan et al.[83] have implemented a swarm-based Firefly algorithm to perform
the optimization task on the CVD dataset. However, the firefly algorithm had the issue of being
stuck in local optima.

Table 2.1 : Literature Review of Feature Selection Algorithm on CVD Data

Name of Feature Selection Characteristics Drawback

the Author Approach

Ottom et al. Accumulated Features were ranked It was inefficient

[60] Feature Selection based on the Gain ratio when dealing
with attributes
that have large
distinct values.

44
Wosiak et Reversed The approach determines Ignores the
al.[61] Correlation the mutual relation interaction
between the attributes and among classifiers
the number of instances
by applying the
correlation coefficient in
reverse order.

Perez- Quasi constant based The algorithm filtered the This resulted in
Riverol et feature selection similar features by more of type II
al. [62] algorithm checking the Quasi errors
constant.

Amin et Combinational The subset of the Computationally

al.[63] feature selection combinational features expensive
was formed.

Awan et al. Chi square-based A two-sided chi-square very sensitive to

[64] feature selection was applied where the p- the size of the
value determines feature dataset
significance

Joshi et Entropy-based Critical features were Lacks in

al.[65] feature selection extracted using the Generality
entropy-based partition
method.

Takci et Fisher filtering The features were Does not take into
al.[66] based feature selected using the fisher account the
selection score by finding the dependence of
distance between the one feature on
mean and sample mean of another.
each class per feature
divided by the variance

45
Haq et mRMR based The mrMR method It was a Bivariate
al.[67] feature selection minimizes duplication method that
framework and relevance by considered only
selecting the most pairwise
relevant features to the redundancy or
class label. relevance.

Hung Minh Infinite Latent Most significant features The proposed

et al.[68] feature selection were selected based on method became
weights and ranks using minimum variant
probabilistic latent. and unbiased
Towards
Semantic analysis(PLSA) estimators with
an increase in the
size of the
dataset.

Xu et CFS with Best first The subsets were Does not consider
al.[69] search generated using the CFS the dependency
subset evaluation of one feature on
algorithm. The merit of another
the features was
identified using the
relevance of the feature,
prediction ability, and
classification goal.

Liu et The relevant based A feature was considered Omitting the

al.[70] Information distinct if the mutual features that do
selection algorithm information between the not have mutual
feature and the label was information
not zero. might throw the
needed features.

46
 Wang et Kernel canonical Kernel canonical based Computationally
al.[71] correlational feature mutual feature selection expensive
selection approach uses the concept
of mRMR to reduce the
joint redundancy among
the already selected
features and candidate
features.

Elgin Correlation-oriented Effectively deal with Lack of

Christo et feature selection redundant and relevant interaction
al.[81] feature between search
space and
hypothesis space.

Al-Zoubi et Salp swarm Applies the strategy of the The proposed

al.[82] algorithm heuristic method model was slow
in convergence

Khan et Firefly algorithm Randomly generates the Getting stuck

al.[83] feasible solution by with local optima
considering the light
intensity of the firefly.

The table 2.1 above summarizes the numerous review of the literature work that has
been conducted using various feature selection algorithms on the CVD dataset. The feature
selection algorithms applied to the CVD dataset, such as accumulated feature selection
algorithms and reversed correlation based feature selection algorithms, have filtered the
important attributes from the disease dataset. Still, these approaches are found inefficient when
attributes with many distinct values are imposed on the algorithm. The conventional algorithms
such as chi square-based feature selection approach and entropy based feature selection applied
to these diseases were lacking in generality as they are very sensitive to the size of the dataset.
47
The traditional techniques such as Fisher filtering and Infinite latent feature selection
algorithms did not consider one feature's dependence on another. Numerous nature inspired
algorithms are applied to the CVD dataset, such as the Salp swarm algorithm and Firefly
algorithms, to reduce the dimensions of the dataset effectively. These algorithms used the
strategy of the heuristic approach to extract the features. But these approaches were found slow
in convergence which stuck with local optima. Hence, a sophisticated feature selection
approach is needed to overcome these drawbacks.

2.3.2 Feature Selection Algorithm for the CDM Data

This section provides an overview of some research work that is already carried out in
the early prediction of CDM using various feature selection algorithms.

Mahboob Alam et al.[84] obtained valuable feature insights from the CDM dataset by
applying the principal component analysis (PCA) algorithm. The experiment was conducted
using the National Institute of Diabetes dataset, which is publicly accessible via the University
of California, Irvine repository. The dataset was first pre-processed to eliminate missing values
and inconsistent data. The original dataset was then reduced to get data with a lower volume
that produced the same result. The PCA method was used to extract the important features from
the dataset. The dataset included features such as diastolic blood pressure, BMI, age, and
glucose. PCA was used in conjunction with the ANN classifier to obtain a 75% accuracy.
Monirul Kabir et al.[85] identified the best features from a CDM dataset using a constructive
method for feature selection (CAFS). The algorithm partitioned the entire dataset into
correlated and dissimilar feature groups. In combination with CAFS, a feed-forward Neural
Network (NN) with a three-layer architecture was chosen to perform classification tasks. The
proposed CAFS evaluated features at each epoch of the neural network. The addition and
deletion of features were determined by the NN validation result. The usefulness of the features
was determined using the validation set's classification accuracy. The approach achieved an
accuracy of 76%.

Abitha et al.[86] used a hybridized feature selection method to extract the best features
from the CDM Dataset. The hybrid features were chosen using PSO and Rough Set Theory.
The PSO algorithm was used to evaluate the fitness of the features, and then Rough Set Theory
was applied to choose the features. The method successfully predicted the illness with 91%
[Link] et al. [87]implemented a novel Sigmis attribute selection algorithm to predict
48
CDM disease. The method was initiated by computing the Pearson correlation coefficient
among the continuous features and the target label. Then, using the t-test, the statistical
correlation between the features was determined. If the value of 't' exceeds the level of
significance α = 0.05, the feature was considered as important, and it was selected. The
experimental results indicated that the suggested technique forecasted the illness with a
precision of 86% when six features from the dataset were used. Li et al.[88] developed the Fast
Feature Selection Method based on the Coefficient of Variation for CDM prediction. It is a
feature selection method that prioritizes the attributes based on their dispersion of the data.
Attributes with a low data dispersion were rejected and considered as less important. The
information about the data variation was obtained by analyzing the coefficient of variation in
terms of standard deviations to the mean of each feature. The algorithm was combined with a
multi-layer perceptron and attained an accuracy of 77%. Maniruzzaman et al.[89] developed a
method for selecting features from CDM data based on the concept of mutual information. The
method detected the mutual information between the two features using the joint probability
distribution and marginal probability distribution techniques and therefore extracted the most
informative features from the dataset. The performance of the algorithm was evaluated using a
variety of machine learning methods, including the Artificial Neural Network and Naive Bayes.
The technique obtained an accuracy of 86%.Ye et al. [90] implemented a randomly restarted
incremental feature selection algorithm to determine which variables in the CDM dataset
influence the prediction accuracy. The statistical association evaluation algorithm was used to
identify the low-ranked features among the obtained features. The features were ordered
according to the feature that corresponded to a class label. The feature element was introduced
to the feature set until the classifier's accuracy decreases consistently. The suggested feature
selection technique was combined with an SVM and Decision Tree classifier. The proposed
technique predicted the disease with 89% accuracy. Rodríguez-Rodríguez et al.[91] developed
a sequential input selection method for the prediction of type 1 CDM. The concept of regression
coefficient was used to remove the features from the original data set using the concept of
backward elimination. The feature which minimizes the validation above the threshold error
was considered a significant feature. The data set obtained from the 25 diabetic patients (14
men and 11 women)was considered as input to the algorithm. Huang et al.[92] have
implemented the FSSMC based feature selection algorithm for the CDM data.

The usefulness of the features was evaluated by finding the relationship between

49
feature value and the disease outcome. Sneha et al.[93] have used irrelevant Correlation-based
feature selection to extract the best traits from the CDM dataset. Each attribute value was
compared to the value of an additional attribute. If there was a greater difference between the
two attributes, the latter was considered to be less significant. As a result, the best attributes
were chosen by randomly ordering the features. Nagpal et al.[94] implemented a modified fast
feature selection algorithm to extract the optimal features from the CDM dataset. A minimum
spanning tree was constructed to determine the relationships between the features. The results
of the experiment were compared using a variety of feature selection algorithms. In terms of
accuracy, Modifiedfast outperformed all other traditional techniques. Kavakiotis et al.[95]
presented a feature selection method based on Linear Discriminant Analysis (LDA) for the
prediction of CDM. The LDA is a method for reducing the dimension of data by projecting it
into a new feature space that maximizes class separability. The feature reduction was performed
by calculating the between-class variance and within-class variance. Finally, lower-
dimensional space will be obtained, which will maximize the class variance. The proposed
algorithm experimented with an SVM classifier. The experimental result indicated that the
approach obtained an accuracy of 86%.

Hegazy et al.[96]developed a bio-inspired chaotic salp swarm algorithm for extracting

key features from CDM data. In the proposed approach, chaotic maps were utilized to find the
location of the salp, which in turn was considered as an altered version of the basic salp swarm
algorithm. The use of a chaotic map increased the convergence speed. The leader salp led the
front of the chain, and the rest of the salp was treated as followers. The feature subsets were
assigned to each of the salps, and each salp was assessed for its fitness. The proposed method
was combined with the KNN classifier, which predicted the disease with 73% accuracy. The
presented approach, however, had the drawback of being slower in terms of convergence.
Mafarja et al.[97] developed a binary dragonfly-based method for the prediction of CDM. It is
an evolutionary algorithm that imitates the behavior of dragonflies. Step vectors were used to
represent the dragonfly's position. The fitness of the population was determined on every
iteration using the step vectors as X={x1,x2,..xn}. The Step vectors 0 and 1 were used to track
the dragonfly's position. The algorithm predicted the disease with 74.5% accuracy. Kemal
Akyol et al.[98] have introduced cascading feature extraction for the forecast of CDM. The
whole data set was first split into three equal subsets. The training and testing set was formed
with each of the created parts. Different feature selection algorithms were passed to these
50
subsets, such as iterative relief, stability, and recursive feature elimination. The feature
evaluation algorithms have experimented on ensemble classifiers such as gradient boosted
trees, AdaBoost, and random forest. The proposed approach achieved a 73% accuracy rate.
Anaraki et al.[99] have proposed a fuzzy rough set based frog algorithm for selecting the
optimal traits from a CDM dataset. Mashhour et al.[100] used a hybrid algorithm that combined

the firefly algorithm with chi-square. The suggested method aimed to decrease the
dimensionality of the given data while simultaneously extracting significant characteristics.
Firefly locations were calculated using the chi-square value. The categorization task was
carried out using the KNN algorithm. The firefly's intensity was calculated using the fitness
function as a parameter for extracting relevant features from the dataset. The suggested method
outperformed current approaches in terms of accuracy. Muhammad [Link].[101] have proposed
a PCA-based feature selection algorithm for extracting the optimal features from the CDM
Dataset and for predicting the disease's probability. Suresh et al.[102] have applied recursive
feature elimination technique to extract significant feature traits from CDM Data. The
drawback of this approach is that it is computationally expensive and does not consider any
correlation between the features. Saw et al.[103] have implemented PSO based evolutionary
algorithm as an optimizer for CDM prediction. The advantage of the approach was that the
algorithm optimized the problem by improving the candidate solutions. But the drawback with
this approach was that it had missed some informative features due to pbest and gbest updation
mechanisms. Alharbi et al.[104] have applied a Genetic Algorithm to get the important traits
from the CDM dataset. The advantage of the algorithm was that in each generation, algorithms
could get the optimal fit feature traits. However, the method has the disadvantage of being
computationally costly. Putri et al.[105]selected the best features from the CDM Data using a
Chi square-based statistical method. The drawback of the approach is that it is sensitive to the
size of the dataset.

51
 Table 2.2:Literature Review of Feature Selection Algorithm on CDM Data

Name of Feature Characteristics Drawback

the Author Selection
Approach

Mahboob Principle It is the orthogonal Higher computational

Alam et component transformation of data to cost as the size of the
al.[84] analysis reduce the dimension of data dataset increases
which retains only variation
present in the data.

Monirul Constructive The proposed technique used Considers only the

Kabir et based Feature a constructive approach in correlation among the
al.[85] selection combination with the features and ignores the
correlation information to relationship with the
select the relevant feature classifier.
traits by dividing the entire
dataset into correlated
feature groups and dissimilar
groups.

Abitha et The hybrid The concept of the Rough Computationally

al.[86] algorithm Set theory was enforced to expensive with a high
with a get the best traits for dimensional dataset as
combination classification. The fitness of in each generation of
of Rough Set the feature was evaluated the PSO, the fitness of
Theory and using the PSO algorithm. the features needs
PSO to be
evaluated
Blessie Sigmis The Sigmis will find the Performs better in
feature correlation between the accuracy only when
et al.[87]
selection continuous features and the data is normally
class label using a t-test. distributed

52
Li et al.[88] Fast Feature The proposed method The coefficient of
Selection applied the concept of variation is sensitive to
coefficient of variation small changes in mean
where attributes with low values thus when the
data dispersion were average value is zero,
disqualified. coefficient variation
approaches to infinity.

Maniruzza Mutual The mutual information Did not consider the

man et information- between the two features relationship between the
al.[89] based feature was identified by finding feature and class label
selection joint probability distribution into account
and marginal probability
distribution. Thus it detects
the most informative
features.

Ye et al.[90] Randomness The approach applied the Adding and removing

based statistical association of p the feature on every
incremental value to obtain the low iteration resulted in
feature ranked features. nesting problems.
selection
approach

Rodríguez- The The algorithm works based The approach is limited

Rodríguez sequential on the regression coefficient, to linear relationships
et al.[91] Input and features were removed if and easily affected by
selection the validation error was outliers.
algorithm smaller than a threshold.

53
ang et FSSMC The usefulness of the The approach did not
al.[92] Algorithm features was evaluated by guarantee the accuracy
finding the relation between
the feature value and the
outcome.

Sneha Correlation- The value of each of the The approach selected

based feature attributes was compared redundant features as
et al.[93]
selection with another attribute. If they did not consider the
there is more difference relationship between
between the attribute, the variables.
particular attribute was
considered as less
significant.

Nagpal et Modifiedfast The redundant features were Computational

al.[94] removed by constructing a complexity is very high
minimum spanning tree as a with a larger number of
weighted complete graph. features.

Kavakiotis Linear Data will be projected into a LDA algorithm will not
et al.[95] Discriminant new feature space to work properly if the
analysis maximize class separability. number of objects in
The feature reduction was various classes differs.
performed by calculating the
difference between class
variance.

Muhammad PCA Reduced the dimension of If the variances of

et al.[101] the data by computing the irrelevant features are
eigenvalue pairs and large, then extracted
covariance values of a given features are not suitable
feature for classification.

54
 Suresh et RFE Eliminated the features Found computationally
al.[102] recursively using the expensive with high
heuristic approach dimensional data as it
did not take into account
any correlation among
the features

Saw et PSO Evolutionary algorithm The approach Missed

al.[103] selects the features by some informative
candidate set generation features due to pbest and

gbest updation
mechanism

Alharbi et Genetic Features were selected Found Computationally

al.[104] Algorithm according to the level of expensive with the high
fitness in each generation dimensional dataset

Putri et Chi-square Features were filtered by Approach was found to

al.[105] applying the statistical chi- be sensitive to the size
square test of the dataset

The table 2.2 above summarises the various reviews of the literature on the application
of feature selection algorithms on the CDM dataset. Numerous feature selection algorithms are
proposed to identify the informative features from the CDM disease dataset. These algorithms
are found computationally expensive as the size of the dataset increases. An algorithm such as
constructive based feature selection considers only the correlation among the features and
ignores the relationship with the classifier. A few approaches are applied in hybrid
combination, such as rough set theory in combination with PSO to determine the relevant
features from the disease data. But the critical part of this approach is determining the fitness
of the features in each generation. The algorithms, including sigmis feature selection and the
fast feature selection, applied the concept of coefficient variation, where attributes with low
data dispersions are disqualified. But the coefficient of variation is sensitive to small changes

55
in mean values. Thus when the average value is zero, coefficient variation approaches to
infinity. Thus novel feature selection approach has to be implemented to overcome with these
drawbacks.

2.3.3 Feature Selection Algorithm for the CKD Data

This section summarises some of the research work that has already been conducted on
early CKD prediction using various feature selection algorithms.

Avci et al.[106] have introduced information acquisition-based element selection to

obtain the key features from a renal disease dataset. The acquired data was used to select
themost pertinent feature from the dataset. The effectiveness of the features processed utilizing
𝑛 the equation 𝐼𝑛𝑓𝑜𝐺𝑎𝑖𝑛(𝐹) = −∑𝑝𝑟𝑜𝑏(𝐶𝑖)log(𝑝𝑟𝑜𝑏(𝐶𝑖)) where prob(Ci) is the likelihood 𝑖=1
appropriation of the highlight of the ith feature. The algorithm chose 20 feature traits from the
dataset. The J48 algorithm was utilized as a classifier on a diminished dataset. The proposed
strategy improved the predictive accuracy while classifying the disease. Pitchumani
Angayarkanni et al.[107] implemented a technique known as Sequential Forward Selection
(SFS) to choose the best feature subset from the kidney disease dataset. The method chose
features repeatedly in response to the classifier's output. The suggested study utilized classifiers
such as Naive Bayes, K nearest neighbor, and support vector machines as basic classifiers. The
suggested method was accepted based on various estimations, including Accuracy, Precision,
and Recall. The disease was predicted with 85% accuracy using the proposed algorithm.
Aprilianto et al.[108] have applied the CFS algorithm to filter the important feature traits from
the CKD dataset. The CFS algorithm was further optimized using the Binary particle swarm
optimization algorithm. The dataset had 24 features, including symptoms, manifestations, and
risk factors associated with CKD. The SVM algorithm was used to build the predictive model.
The approach obtained an accuracy rate of 88.7%. Dankolo et al.[109 ] pioneered the use of
entropy-based feature selection to classify constant renal diseases. The suggested algorithm
selected ten features from a total of twenty-four features. The experiment was carried out
through Rapid excavator Studio.

Sara et al.[110] have applied Consistency based feature Evaluation techniques to predict
the kidney disease. The Class consistency-based logic was used as an assessment criterion by
the CS to filter the features from the dataset. The proposed approach aimed to condense the
dataset features into a subset that was dependent on the class larger part measure. An ANN was
56
used to handle the predictive task in conjunction with the described strategy. The presented
work produced a list of nine informative features from CKD datasets. The methodology was
found to be 80% accurate. Almasoud et al.[111] have combined ANOVA with Pearson
correlation to extract informative traits from the CKD informational dataset. To begin, the data
collection was cleaned and examined to eliminate any missing values. The ANOVA and
Pearson relationship measures were used to determine the relationship between the attributes.
The suggested method prioritized six informative features that would affect the prediction
result. A logistic regression-based algorithm was applied to predict the CKD [Link]
suggested strategy predicted the illness with a precision of 78% Shrivas et al.[112] have
implemented a Union-based feature evaluation algorithm for the prediction of CKD. The
fundamental point of the exploration was to improve the predictive analysis of CKD. Chi
Squared Attribute Evaluation, Genetic Search-J48, and One-R Attribute Evaluation were
utilized to rank the features of the dataset, and it was given as input to the UFS algorithm. The
prediction was performed using Naïve Bayes and the Multilayer Perceptron algorithm.

Liu et al.[113] developed a technique for predicting kidney disease using a chi-square-
based attribute selection approach. Mohammed et al.[114] implemented a fusion-based
attribute selection method for predicting CKD. The most important features were selected
based on the output of different feature selection algorithms. The primary features were chosen
using a voting-based approach, with the significance of the qualities evaluated using different
feature selection techniques. Polat et al.[115] used correlation-based Greedy stepwise attribute
selection to forecast kidney disease. The algorithm started with a null set. A forward selection
method was used to select the features, and then a backward elimination technique was used to
filter them. The prediction task was carried out using the SVM. The approach was found more
accurate than previous techniques. Dilli Arasu et al.[116] have proposed TOPSIS based feature
selector for the CKD prediction. It depends on the idea that a given feature should be limited
to mathematical separation from the objective of the variable. At first, the hypothetically best
solution was built. At that point, the distance between the features and their hypothetical best
and worst solutions was resolved for each segment in the dataset. The best characteristics were
chosen based on their priority. Elhoseny et al.[117] have developed a Density-based attribute
selection method to select the most appropriate features from a kidney disease dataset. The
DFS considered separating the features by considering correlation as a measure. At first, the
likelihood of each of the feature components was determined. The attributes were positioned

57
dependent on the overlap between the features. The proposed strategy accomplished a precision
of 87.3%. Cai et al.[118] have proposed a Sequential Forward Selection algorithm to separate
the best attributes from the CKD dataset. Initially, an empty feature set was created with the
given k features 𝐸𝑘 = {∅}. The optimal traits from the dataset were extracted using the greedy
search algorithm till it gave maximum accuracy. The attributes were chosen iteratively
dependent on the aftereffect of the classifier. The KNN was utilized as a classifier. The
proposed approach was assessed utilizing parameters such as Precision, Accuracy, and Recall.
The methodology achieved a precision of 85%. Senan et al.[119] have proposed RFE to filter
the feature traits from the CKD dataset. The dataset was compiled by utilizing 24 attributes
related to kidney disease. The prediction task was performed using various machine learning
algorithms such as SVM, KNN, and Decision Tree. The methodology obtained an accuracy
rate of 95%. Kadhum et al.[120] have proposed a combination of evolutionary wrapper
approaches to predict early CKD. The feature from the dataset was selected using a genetic
algorithm, and the prediction task was handled using SVM and an extreme learning machine
model. The approach achieved better accuracy than the conventional technique.

Wibawa et. al.[121] developed a CFS algorithm to select the features from the CKD
dataset. The predictive ability of the filtered features was examined using the Adaboost
algorithm. The implemented technique achieved the best precision rate than the traditional
techniques. Parthiban et al.[122] proposed a hybrid of filter, hybrid, and wrapper techniques to
extract the ideal features from the renal disease dataset. The algorithm was named as Hybrid
Filter Wrapper Embedded based feature selection algorithm. The filter algorithm was
implemented using Gain Ratio, Relief, Gini Index, and One-R algorithms. The Wrapper
approach was applied using the Improved Bat Algorithm. The proposed strategy outperformed
the conventional approach in terms of accuracy. Abdulla et al.[123] have proposed a mixed
integer program(MIP) based cost-effective model to select the informative features from the
CKD dataset. The proposed model improved the accuracy rate up to 10% and was also found
robust against the feature groups. Lambert et al.[124] proposed Oppositional-based FireFly
Optimization method to extract the features from the CKD dataset. The use of oppositional-
based learning improved the convergence rate of the FireFly algorithm. The disease prediction
task was performed using a Deep Neural Network. The experimental results indicated that the
suggested algorithm outperformed the existing algorithms in terms of accuracy.

Jain et al.[125] Implemented a two-phase hybrid feature extraction algorithm to filter

58
the attributes from the CKD dataset. The hybrid algorithm was formed using the combination
of PCA and the Relief algorithm. The predictive ability of the algorithm was examined using
the SVM classifier. The proposed approach obtained better accuracy than the conventional
approaches. Saha et al.[126] have implemented the Correlation-based feature evaluation
technique to get the ideal feature traits from the CKD dataset. The forward selection was
applied to choose the features, and Backward elimination was made used to filter the features.
The SVM algorithm was applied as a machine learning classifier to make predictions from
thedataset. The methodology correctly predicted the disease 91.65 % of the time. Jerlin Rubini
et al.[127] have applied the fruit fly optimization(FFOA) algorithm to determine the fitness of
the features from the kidney disease dataset. The FFOA chose the key features from the illness
dataset. The classification task was performed using multi kernel support vector machine. The
experiment was conducted using MATLAB software. The algorithm achieved an accuracy of
89.11%.Manonmani et al.[128] have presented a technique called density-based feature
selection (DFS) to filter the optimal attributes from the CKD disease dataset. The heuristic
approach rated the feature positions according to their density, with attributes regarded optimal
only if there is minimal overlap between the various class identifiers. The heuristic approach is
considered as one of the universal approaches which can be applied to provide the solution to
any problem irrespective of the domain of the application. However, to select the features from
the disease dataset specific approach is recommended. In the proposed study heuristic method
predicted kidney illness with a precision of 87.3%.

Table 2.3 : Literature Review of Feature Selection Algorithm on CKD Data

Name of the Feature Selection Characteristics Drawback

Author Approach

Avci Information gain-based Features were ranked The problem occurs

ranker search engine based on information with information
et al.[106]
gain gain when attributes
take a large number
of distinct values

59
 Sequential Forward The Greedy search Computationally
Selection algorithm selected the expensive with high
Pitchumani
important subset of the dimensional data
Angayarkan
features by iteratively
ni et al.[107]
applying the machine
learning classifier.

Aprilianto et Correlation-based The algorithm Ignores the

al.[108] considered the interactions among
Feature Selection
predictive capabilities of the classifier
each feature,
correlation, and
redundancy level among
features while selecting
the features.

Entropy-based feature Attributes were selected Possibility of

Dankolo
selection by measuring the overfitting
et al.[109]
randomness of the
features in the dataset.

Consistency oriented Obtain the subset set of Time complexity is

Subset Evaluation features through class higher while
Sara
majority measure selecting the subset
et al.[110]
of features

Almasoud Pearson correlation and Distance between the The test will be
ANOVA test feature variable was inaccurate if there is
et al.[111]
determined using the a greater difference
correlation measure. in standard
deviations between
groups.

60
Shrivas et Union-based feature The higher ranked Time-consuming
al.[112] selection features were
and complex
determined based on the
approach
factors Union of Info
Gain, Gain Ratio, Chi-
Squared

Liu et Chi square- The statistical test was very sensitive to the
al.[113] based feature selection applied to determine the size of the dataset
occurrence of two
events. Chi-Square
score measured the
deviation between the
expected counts and
observed Count.

Mohammed Fusion-based feature A voting based Did not guaranteed

selection technique was applied to the accuracy
et al.[114]
obtain the most
important attributes

Polat Correlation-based Features were obtained Less scalable with

Greedy stepwise search by forwarding selection large datasets
et al.[115]
engine and filtered useless
attributes by backward
elimination to find the
best feature subset.

61
 Dilli Arasu TOPSIS-based feature The best theoretical and Time-consuming
et al.[116] selection worst solutions were and greater time
computed by applying complexity
features, and the optimal
features were filtered
based on their priority.

Elhoseny et Feature selection based Applies the heuristic test Unable to deliver the
al.[117] on the DFS algorithm to evaluate the merit of optimal solution in
the feature most of the cases

As summarized in Table 2.3 above comprehensive review of literature has been

conducted related to the application of feature selection algorithms for the prediction of CKD.
The algorithms such as Information gain based feature selection, Sequential forward selection,
Pearson correlation based feature selection, and ANNOVA test are frequently applied to CKD
data. Even though these algorithms could retrieve the relevant attributes from the dataset, these
approaches are less scalable with a large dataset. The algorithm, such as fusion based feature
selection, determined the high-ranked features by applying the voting-based technique. Still,
they did not deliver the optimal solution in most of the cases. The algorithms, including
TOPSIS based feature selection and DFS feature selection, faced the time complexity issue and
did not deliver the optimal solution in most of the cases. A few of the statistical-based feature
selections applied to the CKD disease dataset, such as chi square test, entropy based feature
selection, and correlation based feature selection, are sensitive to the size and nature of the
dataset. These algorithms extracted ineffectual features from the data whenever there was a
small variation in the structure of the data. Hence, robust feature selection algorithms must be
implemented to overcome these drawbacks.

2.4 Chronic Disease Prediction using Machine Learning

Algorithms
Predicting the disorder in its initial phase was time-consuming for the physician, which
62
requires hours of effort. Physical examinations of the disease may be prone to errors[129].
There would be a need for an automated disease prediction system in the future that can
anticipate disease at an early stage. The disease can be early predicted utilizing machine
learning algorithms based on the symptoms of the disease. As of now, significant progress has
been made toward these objectives. The section below enumerates the various research work
proposed in the early prediction of disease using machine learning based approaches.

Figure 2.2: Generic Framework of Machine Learning for Chronic Disease Prediction

Figure 2.2 above depicts the generic framework for the early prediction of chronic
disease using machine learning algorithms. Healthcare data in its native format may contain
anomalies, missing values, and noise. If the same datasets are utilized for predictive analytics
the efficiency of the machine learning model may degrade. These data will be thoroughly
preprocessed to remove any irrelevant information. There are four steps involved in the data
pre-processing process. They are data cleaning, normalization of data, eliminating the missing
63
value, and transforming data. The data is first cleansed to reduce the noise, missing information,
and outliers. Data normalization approaches such as decimal scaling, z score normalization,
and min-max normalization are used to maintain the data within the right range. The data
transformation approach changes the format of diagnostic data to fit the needs of the ML
algorithm. An exploratory data analysis (EDA) method is used to conduct statistical analysis
on the preprocessed data. The dataset is filtered as a result of the preprocessing to extract the
most useful features, thus lowering the computing cost of further [Link], it
eliminates the underfitting, overfitting, and data imbalance problem with considered data which
may affect the outcome of the prediction. The preprocessed data is thoroughly trained using
the various machine learning algorithms. Once the ML model has been developed using clinical
data, it is exposed to an external and internal validation process to ensure that it is accurate.
This is done before passing the disease data as the test set for the model. Potential predictive
outcomes of the model include predicting the absence or presence of disease, assessing the risk
levels for a certain disease, and predicting the likelihood of a specific disease.

2.4.1 Machine Learning Algorithms for the Prediction of CVD

The following section summarises some of the research work that has already been
done on the early prediction of CVD using various ML-based algorithms, as well as some of
the findings from that study.

Maji et al.[130] have implemented a decision tree-based algorithm for heart disease
prediction. One of the principal reasons behind utilizing the decision tree was that it is
extremely simple to implement. The investigation was carried out in the WEKA environment.
A good outcome was obtained with this model when it was presented with limited data. When
the model was exposed to a large volume of data, it degraded because of the overfitting issue
with the decision tree. Gokulnath et al.[131] have proposed a Genetic algorithm-based SVM
Model to forecast coronary disease. The fit features were chosen by utilizing the Genetic
algorithm, and the prediction was performed utilizing SVM based algorithm. Kodati et al.[132]
developed the K means algorithm for cardiac illness prediction. The dataset was assembled
from the UCI archive. The various bunches of clusters were framed depending on the
estimation of k (2 to 8), and the ideal outcome was acquired with k=4. Dulhare et al.[133] have
applied the Naive Bayes algorithm to predict coronary failure utilizing a stat log informational
repository. The experimental results revealed that the approach was less accurate due to Type

64
II errors. Aldhyani et al.[134] have introduced a Soft clustering-based approach to the early
predict coronary diseases. The Rough K Means(RKM) algorithm was proposed to form the
clustered groups. The RKM improved the prediction performance by excluding the ineffectual
object. The data were grouped into upper and lower bound. The lower estimated objects were
considered as appropriate objects, and information placed with higher bound were redundant
objects. The aftereffect of RKM was compared with the classifier having different
parameterssuch as recall, precision, and sensitivity. Magesh et al.[135] have implemented a
Clustered decision learning (CDTL) approach for the early detection of coronary disease. The
suggested approach diminished the dimension of the features by 62%. The Random forest
algorithm was used as a classifier.
Ali et al.[136] proposed a chi-square-based deep neural network for cardiac disease
prediction. The relationship between the attribute variable and the class was determined using
the chi-square test. The DNN model was built in two layers, each of which had four neurons.
The proposed technique achieved a 91% accuracy rate. Akgül et al.[137] have set forward an
ANN model for early detection of coronary illness. A hybrid approach was developed to
improve prediction results by combining ANN and a Genetic algorithm. The experiment was
conducted utilizing the 'Cleveland' dataset. The validation results explored that the proposed
hybrid model outperformed C4.5, K-Nearest Neighbour, and Naive Bayes model in terms of
performance parameters such as precision, F-measure, and accuracy. Khourdifi et al.[138] have
proposed an updated evolutionary model for CVD prediction. The Fast Correlation-Based
Feature Selection was initially used to eliminate less significant features. The prediction task
was accomplished through the use of an ANN-based algorithm. The PSO was used in
conjunction with Ant colony optimizers to improve prediction accuracy. Eskandari et al.[139]
have implemented a combination of the Whale Optimization and Dragonfly algorithms to
predict coronary disease. The experimental findings showed that the suggested approach
obtained an accuracy of 88.89%. Jalali et al.[140] introduced evolutionary intelligence to
anticipate the early stages of heart disease. Using a correlation-based technique, the suggested
approach identified useful features from the dataset. A multilayer perceptron (MLP) was
deployed to accomplish the prediction goal. Through the usage of a Multi-Verse Optimization
algorithm(MVO), the predictive ability of the multilayer perceptron algorithm was boosted.
The hybrid MLP-MVO model improved the proposed model performance to forecast the illness
accurately. Castellanos-Garzón et al.[141] have made a study on applying a hybrid framework
for predicting chronic diseases, which combined a genetic algorithm with rule- based induction
65
(IF/THEN). The outcome of the approach was evaluated through well-known parameters such
as recall, accuracy, roc, and precision. Siddique Ibrahim et al.[142] have proposed an
evolutionary memetic algorithm in combination with an associative rule classifier for the
forecast of coronary illness. The proposed investigation was assessed by utilizing various
validation parameters. P et al.[143] have presented an integrated approach for the early
prediction of coronary disease. The proposed framework combined cuckoo search with the
rough set algorithm. Cuckoo search methods extracted the main features from the coronary
disease dataset. The rough set approach was applied to extract key feature traits from the
dataset. Further comparative analysis was carried out to find out the cons and pros of the
proposed approach.

Table 2.4: Literature Review of Machine Learning Algorithm on CVD Data

Name of the Machine Learning Characteristics Drawback

Author Approach

Maji et Decision Tree One of the basic and Performance

al.[130] simplest algorithms degrades due to
which are robust against overfitting of
noisy data trees

Gokulnath et Genetic algorithm- SVM algorithm Computationally

al.[131] based SVM Model experimented in expensive and
combination with time-consuming
Genetic algorithm

Kodati et K means clustering A lazy learner-based Determining the

al.[132] clustering approach was optimal value of
found computationally the K cluster is
faster by producing tight computationally
clusters. expensive.

66
Dulhare Naïve Bayes classifier The algorithm used the More of a Type II
Bayes theorem, which error due to the
et al.[133]
converged quickly with assumption of
minimal training data class conditional

independence

Magesh et al. Decision tree learning The proposed approach The feature
[135] based on the cluster to applied the data selection method
early predicts coronary distribution strategy to did not remove
artery disease. split the data based on the the ineffectual
target label through which features. So the
different method was
combinations on the class found to be
set have been made. The computationally
feature identification was intensive.
made using the concept of
entropy. The
classification was
performed considering
the value of entropy.

Ali et al . chi-square based deep The DNN with two The proposed
neural network model hidden layers containing approach
[136]
was implemented to four neurons achieved an necessitates a
forecast the CVD accuracy of 90%. massive number
of samples to
consider the chi-
square
approximation
test and to
converge the
model.

67
Khourdifi A fast correlation- The proposed approach Took a long time
based attribute drastically reduced the for the
et al.[138]
Selection scheme was feature's dimensions, and computation. The
applied in combination the accuracy rate was approach was
with the Neural
Network algorithm.
PSO and Ant colony- increased to 86% through found poor with
based methods were the PSO optimizer. generality.
utilized to optimize the
classifier.

Eskandari Whale Optimization The hybrid evolutionary The evolutionary-

and Dragonfly approach combined with based approach
et al.[139]
algorithm were the SVM classifier was expensive in
implemented in obtained an accuracy of terms of
combination with 88.89%. computational
Support Vector resources and
Machine was found less
scalable with the
large data set.

Jalali Heart disease MLP was used to perform Random iteration

prediction using MLP the classification task. A performed during
et al.[140]
and Multi-Verse multi-Verse Optimizer the training phase
Optimizer was used to improve the increased the
performance of MLP. The training time of
integration of MLP and the model.
MVO improved
prediction capabilities.

68
 Castellanos- Rule-based Classifier The feature fitness was Developing if-
Garzón with Genetic assessed through a then rules using
Algorithm Genetic algorithm, and the evolutionary
et al.[141]
the rule induction genetic approach
algorithm was used to was found to be
perform the prediction. slower and prone
The approach was found to less scalable
promising for the disease when exposed to
prediction task.

a large number of
samples in it.

Table 2.4 above summarizes the findings of numerous studies conducted in the area of
applying ML algorithms in the early prediction of CVD. The traditional algorithms such as
Decision Tree, SVM algorithms, K means clustering, and Naïve Bayes algorithms are
frequently applied for predicting CVD disease. However, the performance of the decision tree
algorithms was degraded due to the data overfitting issue. The Naïve Bayes classifier did not
obtain good accuracy with CVD data due to the assumption of class conditional independence,
which was made while training the model. The machine learning algorithms are also proposed
in combination with feature selection algorithms to attain good accuracy. But the feature
selection algorithms were not data specific, which did not produce fruitful results. Many of
researchers have also proposed nature-inspired algorithms to optimize the performance of the
machine learning algorithms. But in most cases, these algorithms were found to be slower and
prone to less scalable when exposed to a large number of samples in it. As a result, an efficient
machine learning framework capable of accurately predicting chronic diseases is required.

2.4.2 Machine Learning Algorithms for the Prediction of CDM

This section reviews the literature on the early prediction of CDM using various
machine learning algorithms, as well as findings from those studies.

Aiello et al.[144] have predicted the Type 1 CDM using the KNN algorithm. The

69
glucose profile was determined by using multiple KNN classifiers. The experiment was
conducted UVA/PADOVA simulator. Li et al.[145] conducted a thorough investigation on
deep learning-based models for forecasting CDM in its early stages. To evaluate the robustness
of the model, a ten-fold cross-validation was performed. The suggested approach yielded
satisfactory outcomes. Alehegn et al.[146] have applied an ensemble model in combination
with Decision Tree, Naive Bayes, and SVM algorithmsx for the early CDM prediction. The
data set containing an aggregate of 768 records belonging to diabetes patients was used to carry
out the experiment. The model was found to be 84% accurate. Islam et al., et al.[147] have put
forward the Bagged CART model for the prediction of CDM. The data generated through
Bangladesh healthcare demographic survey was used to conduct the experiment. The risk factor
of the disease was determined through the statistical t and chi-square test. The proposed
technique predicted the illness with an optical accuracy. Singh et al.[148] have implemented a
Stacked ensemble model by combining various meta learners such as Naïve Bayes, KNN, C4.5,
and logistic regression. The experiment was executed utilizing the Pima diabetes as the dataset.
The proposed meta-learning model accomplished an accuracy of 79% while predicting the
disease. Wang et al.[149] proposed rank-based weighted SVM for diabetic illness
identification. The dataset collected from a Chinese general hospital was utilized to perform
the investigation. The features were assigned with weight based on the significance of the
attributes. The correctness of the implemented approach was validated by utilizing the different
measures such as one error, hamming loss, recall, and precision. The proposed technique
identified the relationship between the features using the lasso technique. The approach
recorded good accuracy than the traditional method. Howsalya Devi et al.[150] have proposed
a hybrid model for the prediction of CDM. The model was implemented by combining Farthest
First-based clustering algorithm with Sequential Minimal Optimization. The clusters were
formed using the disease dataset by identifying the similarity between the features using the
Farthest First algorithm. The outliers in the dataset were omitted by applying the concept of
IQR. The SMO is an advanced version of the SVM algorithm, which will overcome the
drawback of the SVM by optimizing it. The experimental results stated that the hybrid model
recorded better precision and accuracy than the based model.

Jayashree et al.[151] have implemented a hybrid model of Hopfield Neural Network in

combination with a gravitational search algorithm to forecast CDM. Initially, a gravitational
search algorithm was built to select the informative features from the disease dataset, and then

70
the Hopfield Neural Network algorithm was applied to make predictions out of the data. Das
et al.[152] have proposed a ridge regression-based feature selection for the early prediction of
CDM. The prediction process was carried out using the extreme learning machine model. The
experiment results showed that the proposed approach achieved high precision while predicting
the disease. Gucen et al.[153]carried out a detailed study on a hybrid model of deep neural
network and Gaussian membership function for the early prediction of CDM. The fuzzification
of data was carried out using the Gaussian membership function as fuzzified data are always
superior to the standardized data to forecast the disease. The deep neural network model was
applied on top of fuzzified data. Thus, the new proposed methodology delivered a better result
in accuracy and precision than the traditional approaches.

Table 2.5 : Literature Review of Machine Learning Algorithm on CDM Data

Name of Machine Learning Characteristics Drawback

the Author Approach

Aiello et KNN algorithm Model Achieved No proper strategy

al.[144] Satisfactory result was used to
to Classify Type 1
determine the
CDM
optimal value of K.

Li et Deep learning Obtained satisfactory The approach faced

al.[145] result the issue of
overfitting and
underfitting while
setting the number
of neurons in the
hidden layers.

71
Alehegn et Gradient Boost Achieved higher Found
al.[146] Model accuracy Computationally
expensive when
applied to the high
dimensional dataset

Islam et Bagged CART Reduced the variance The proposed

al.[147] model with and the overfitting of the model experienced
statistical test t and constructed model a bias in the
chi-square for the prediction result
forecast of CDM due to the loss of
interpretability

Singh et A stacked ensemble Instead of using a single The approach was

al.[148] model using meta- classifier, many of the slower while
learners such as models were combined, generating the
KNN, Naïve Bayes, which was robust prediction result
C4.5, and logit- against the individual and more
boost for the model. computationally
prediction of CDM intensive than
predictions
generated through
the single model.

Wang et Weighted SVM Reduced the dimensions Did not consider the
al.[149] was implemented of the features relationship
for chronic diabetes extensively. The between the feature
disease prediction. approach was robust while assigning the
against the noise. weights.

72
 Hybrid The issue of The approach was
combination of oversampling was found
Howsalya
Farthest First (FF) handled using the computationally
Devi et
clustering Synthetic minority- expensive while
al.[150]
algorithm with based over-sampling forming the cluster.
Sequential Minimal technique. The outliers Obtained an
Optimization were removed by accuracy of 79%
(SMO) for the applying the strategy of with CDM
prediction of CDM. interquartile range. prediction.

Jayashree A hybrid algorithm The GSA reduced the The proposed

et al.[151] consisting of an number of attributes in model converged
evolutionary the diabetes dataset by slowly and also
gravitational search 60%, and the hybrid converged into a
algorithm (GSA) false pattern. It also
and a Hopfield faced the issue of
Neural Network combination obtained local minima
was proposed for an accuracy of 82%. during the feature
CDM prediction. selection phase.

Gucen et A hybrid approach The Gaussian The proposed

al.[153] of deep neural membership function neuro-fuzzy
networks and was used as a data approach showed a
Gaussian fuzzifier. The Deep slow convergence
membership Neural Network was rate due to using the
functions were used to perform gradient descent
applied for the early predictions out of mechanism while
detection of CDM. fuzzified data. The learning its internal
neuro-fuzzy approach parameters.
was found effective.

Table 2.5 summarises the different research that has been conducted utilizing machine
learning algorithms to predict CDM. The KNN algorithm was proposed to predict Type 1

73
diabetes, where the model achieved a satisfactory result with the prediction however there was
no proper strategy was used to determine the optimal value of K. The neural network-based
models used to forecast CDM diseases, such as deep learning and MLP, encountered issues
with data overfitting and underfitting when setting the hidden layer neurons. The ensemble
model used to predict the CDM disease, such as Gradient Boost Model, Bagged CART, and
stacked ensemble model was found to be slower while generating the prediction result and
more computationally intensive than predictions generated through the single model. The
machine learning algorithms were also applied in hybrid combinations such as Farthest First
clustering algorithm with Sequential Minimal Optimization and evolutionary gravitational
search algorithm with a Hopfield Neural Network to predict the CDM disease accurately. But
these approaches converged slowly and also converged into a false pattern. It was also faced
the issue of local minima during the model validation stage. As a result, a machine learning
framework capable of effectively forecasting chronic diseases is essential.

2.4.3 Machine Learning Algorithms for the Prediction of CKD

This section reviews the research on the early prediction of CKD using various machine
learning algorithms.

Jena et al.[154] have proposed Naïve Bayes and Decision tree-based models in
conjunction with a genetic algorithm to forecast CKD. A genetic algorithm was used to select
the attributes, while naive Bayes and decision tree methods performed the classification. The
dataset had a total of 24 features in it, containing 400 instances. The outcome of the model
before attribute selection and after attribute selection were compared. The genetic algorithm
selected 16 features from the dataset after conducting 100 iterations. The comparative analysis
indicated that the naïve Bayes algorithm acquired higher accuracy than the Decision tree.
Individuals who are suffering from heart disease are more probable to develop CKD. Alloghani
et al.[155] have proposed 12 different machine learning classifiers to forecast CKD. The dataset
containing details of 550 CKD patients were utilized to experiment. Out of the assessed
algorithm, the polynomial SVM achieved higher accuracy of 91%. Aly et al.[156] have
presented three - Stage based prescient models to predict CKD infection. The model was
developed using classifiers based on Bagging, Boosting, and Medium Trees. The proposed
model attained a higher accuracy than the stage-1 model. Harimoorthy et al.[157] developed
radial bias kernels using SVM for chronic disease prediction. The result obtained with the
suggested approach was analyzed to that of several other machine learning approaches. R
74
Studio was used to conduct the experiment. The result of the deployed system was measured
using several parameters. The results of the experiments depicted that the SVM radial bias
kernel performed better than all other methods in terms of accuracy and precision.

Sharaf-El-Deen D et al.[158] developed a twin model for chronic renal disease

prediction. A hybrid of Neural Networks and Case-Based Reasoning was used to construct the
twin model. The investigation was carried out utilizing a dataset collected from a Colombian
hospital. Around 34 lakh individuals were identified as CKD-contaminated patients using the
implemented approach. A comparison was made between the outcome of the proposed method
and those from existing machine learning algorithms like SVM and Random forest. The
experimental results demonstrated that the suggested method performed better than the other
two algorithms. Ghosh et al.[159] have utilized a variety of algorithms to predict CKD,
including SVM, gradient boosting classifier, and random forest. The experimental findings
indicate that the Gradient boosting classifier achieved a high level of accuracy compared to
other algorithms. Singh et al.[160] have implemented a rule-based classifier to forecast the
CKD disease. The predictive ability of the suggested approach was validated using a variety of
estimates, including precision, recall, and accuracy. According to the experimental results, the
proposed approach was superior to other approaches. Khamparia et al.[161] have applied a
Deep neural network-based technique to predict CKD. The model was implemented utilizing
a stacked autoencoder mechanism. The investigation was carried out in a MATLAB simulator.
The suggested technique achieved greater accuracy than the conventional methodologies.

Table 2.6: Literature Review of Machine Learning Algorithm on CKD Data

Name of the Machine Learning Characteristics Drawback

Author Approach

Jena et al. Genetic algorithm Bayes algorithm prediction result

[154] in combination with obtained nominal generated more of
Naïve Bayes and accuracy type II errors.
Decision tree-based
machine learning
algorithm for CKD
prediction

75
 Recursive feature The approach was The proposed model
elimination in performed well as was found harder
Alloghani et
combination with overfitting was avoided during the training
al.[155]
radial bias kernel- with the usage of the phase as determining
based SVM to gamma parameter of the a good set of
predict chronic RBF kernel. parameters requires
CKD, CDM, and an expensive Grid
CVD search.

Sharaf-El- A hybrid method Provided higher Computationally

Deen D et using Neural accuracy with expensive as it takes
prediction upon
al.[158] Networks and a long time during
extensive training with
Case-Based the training phase as
Reasoning (CBR) the feature instances. every feature in a
for the prediction of The complex CBR needs to be
CKD relationship between compared with
input and output reasoning
instances was
understood well while
generating the
prediction result.

Ghosh et The Gradient The proposed approach Took more time in

al.[159] boosting classifier was robust against the training phase
in combination with outliers and can due to the sequential
ANOVA and understand the building of the tree,
Pearson correlation nonlinear relationship and the result was
test for the between the feature unconstrained and
prediction of CKD instances very well. overfitted with high
dimensional chronic
data.

76
 Singh et Rule-based Performed well even Overfits with high
al.[160] classifier for the with less training data dimensional data.
prediction of CKD

Khamparia et Deep neural Obtained decent A large dataset was

al.[161] network using accuracy of 83% in considered necessary
stacked early prediction of for model training.
autoencoder CKD.
mechanism with
softmax as
activation function

Table 2.6 above summarises the outcomes of numerous explored studies utilizing
machine learning algorithms to predict CKD. The machine learning approaches such as Naïve
Bayes algorithm, Decision Tree, SVM algorithms, and Neural Network were applied in hybrid
combination with feature selection algorithms such as recursive feature elimination, ANNOVA
technique, and Pearson correlation to early predict the CKD disease. The Naïve Bayes
algorithm obtained nominal accuracy with prediction but generated more type II errors. The
hybrid method of Case-Based Reasoning and Neural Networks was proposed to forecast CKD
and provided higher accuracy with prediction upon extensive training with the feature
instances. The complex relationship between input and output instances was understood well
while generating the prediction result. But the approach was found computationally expensive
as it takes a long time during the training phase as every feature in a CBR needs to be compared
with reasoning. The Recursive feature elimination approach proposed in combination with
radial bias kernel-based SVM was performed well as overfitting was avoided with the usage of
the gamma parameter of the RBF kernel. But the model was harder during the training phase
as determining a good set of parameters requires an expensive Grid search. The Gradient
boosting classifier, proposed in combination with ANOVA and Pearson correlation, was robust
against outliers and can understand the nonlinear relationship between the feature instances
very well. But it took more time in the training phase due to the sequential building of the tree,
and the result was unconstrained and overfitted with high dimensional chronic disease data.
Hence, it is essential to develop an effective machine learning framework that is capable of
77
forecasting CKD diseases.

2.5 Research methods

The section focuses on the research gap that exists with conventional methodologies,
which are used to forecast chronic diseases in their early stages. Silva et al.[162] presented
study on conventional feature selection algorithm. The proposed study concluded that
traditional feature selection algorithms applied to the chronic disease datasets had retrieved the
important features and traits from the data. Still, the approaches are inefficient when attributes
with many distinct values are imposed on the algorithm. The conventional techniques were
lacking in generality as they are very sensitive to the size of the dataset. Nayar et al.[163] have
made a detailed study on the application of nature-inspired algorithms to optimize the
performance of machine learning. The study analyzed numerous nature-inspired algorithms
proposed for the optimization problem in the domain of early disease prediction. But the critical
part of this approach is determining the fitness of the features in each generation. Odili et
al.[164] found that most of the nature-inspired algorithms are computationally expensive,
which will take more time to converge. But in the area of chronic disease prediction, there is a
need for a feature selection approach that can produce the result quickly so that disease can be
predicted in real-time. Zhang et al.[165] have analyzed the application of hybrid feature
selection for the disease prediction problem. The study concluded that the hybrid feature
selection techniques had not delivered good results in most cases. Nasarian et al.[166]
presented the study on applying statistical feature selection techniques for coronary artery
disease. Zebari et al.[167] analyzed that, though the statistical feature selection techniques
provide efficient results in minimal time, it is sensitive to the small changes in mean values.
The study also found that the existing feature selection does not consider the interaction
between features and class labels, which leads to the extraction of useless features, further
reducing the accuracy of the implemented machine learning algorithms.

Exploratory data analysis is considered as one of the important stages of data

preprocessing. A deeper statistical analysis is performed to understand the nature and
underlying hidden assumptions from the data. Indrakumari et al.[168] The conventional

78
approaches often ignored the statistical analysis process, resulting in data anomalies during the
training phase of machine learning. Fatima et al.[169] analyzed that most of the approaches
resulted in falsified prediction as they retrieved the inconsistent patterns from the data due to
the lack of data exploration. Harshvardhan et al.[170] have stated that, to generate a better
prediction result, it is essential to analyze the hidden statistics and distribution of the disease
data, which is not precisely performed in the existing approaches and needs attention. Yashfi
et al.[171] suggested that it is always recommended to apply appropriate feature selection
strategies before the prediction task is performed using the machine learning algorithms.
Benhar et al.[172] have found that the performance of the machine learning algorithms always
depends on how effectively the preprocessing task is performed. The chronic disease dataset in
its original form will contain noise, outliers, and missing values. The presence of outliers in the
data may impact the prediction result. Jamgade et al.[173] have analyzed that many of the
machine learning approaches which are applied for the disease prediction problem have not
used effective strategies to handle the issue of data preprocessing. Meyer et al.[174] noticed
that few of the machine learning algorithms are compatible with categorical data. So the data
has to be converted to the appropriate format before it is imposed on the machine learning
algorithms using the data transformation techniques, which in turn depends on the structure
and nature of the data. Hence it is essential to perform a detailed statistical analysis of the data
for effective disease prediction.

Handelman et al.[175] have analyzed that, even though several machine learning
algorithms proposed to predict chronic diseases, most of the approaches have underperformed
when imposed on the high dimensional chronic disease dataset. Stiglic et al.[176] found that
many of the machine learning algorithms have shown a slow convergence rate and also
converged to the false pattern as these techniques have not analyzed the relationship between
the features of the dataset, which resulted in type II error while predicting the disease. The
study made by Tanha et al.[177] It is essential to build the generalized model as the nature of
the disease dataset changes based on the root cause. The study made by Kaur et al.[178]
analyzed that most of the machine learning algorithms applied in the disease prediction problem
have faced the issue of imbalanced data classification. When the same dataset is used to
perform the disease prediction, the machine learning algorithms favor the classes with major
instances belonging to a particular class over classes with fewer examples. This results in biased
prediction due to the loss of interpretability. The majority of the work which is proposed in the

79
area of chronic disease prediction has ignored this particular problem which has to be addressed
effectively. Table 2.7 below depicts the summary of the research gap.

Table 2.7:Summary of Research Gap

Gap References

Statistical analysis of ● Conventional approaches often ignored Indrakumari et al.[168]

the chronic disease data the statistical analysis process, which
Harshvardhan et
resulted in data anomalies issues during al.[170]
the training phase of machine learning
Benhar et al.[172]
in disease prediction problems.
● Inconsistent patterns extracted from
the disease data due to the lack of data

exploration resulted in falsified

predictions.
● Machine learning algorithms are
substantially dependent on the
distribution and statistics of the disease
data to generate better prediction
results.

80
Feature selection ● Most of the existing feature selection Silva et al.[162]
algorithms for the algorithms did not consider the
Zhang et al.[165]
chronic disease data interaction between chronical feature
instances and classes, which resulted Zebari et al.[167]
in the extraction of ineffectual
Odili et al.[164]
features
● Traditional feature selection
algorithms were found inefficient
when dealing with attributes of
disease datasets having large distinct
values.
● The applied feature selection
algorithms are limited to linear
relationships and are easily affected
by the outlier.
● The feature selection approach is
sensitive to the size and nature of the
disease dataset.

A machine learning ● The traditional machine learning Handelman et al.[175]

algorithm for the early algorithms are found to be
Jayashree et al.[151]
prediction of the computationally complex with the
chronic disease high dimensional chronic disease Khourdifi et al.[138]
dataset.

81
 ● The machine learning algorithms
converged slowly and also converged
to a false pattern, resulting in type II
error while predicting the disease.
● Most of the algorithms proposed are
poor with generality due to the
heterogeneous nature of the disease
dataset.
Algorithm for ● Machine learning algorithms Islam et al.[147]
imbalanced data proposed for chronic disease
Kaur et al.[178]
classification problem prediction are experience bias in the
of machine learning prediction result due to the loss of
algorithm interpretability.
● Most of the machine learning
algorithms proposed for chronic
disease prediction are not addressed
the issue of the data imbalance
problem.

2.6 Research Gap with Machine Learning Algorithms in Prediction of

Chronic Diseases

Table 2.8 below depicts the specific research gap identified with traditional machine
learning algorithms when applied to chronic disease datasets. The SVM algorithms are
extensively applied in the early prediction of chronic diseases. In [131] SVM algorithm was
applied in the early prediction of CVD disease. However, the algorithm was found to be
computationally expensive and consumes more time when applied to high dimensional disease
dataset. The KNN is considered as supervised algorithm which classifies the data based on
similarity. In [144], KNN was used to predict the CDM disease in its early stage. The main
problem with this algorithm is that there is no proper strategy to determine the optimal value
of K. The algorithm can easily be affected by an outlier as it cannot handle the missing values.
In[111] Logistic regression algorithm was applied to early predict the CKD disease. The
82
approach did not produce an accurate result as it failed to consider the linear relationship
between the variables. The decision tree is one of the simplest machine learning algorithms. In
[130], the decision tree algorithm was applied to predict CVD disease. The algorithm obtained
only moderate accuracy while predicting the disease because the algorithm does not produce
accurate results if the dataset has many un-correlated variables.

Table 2.8: Research Gap with Machine Learning in Prediction of Diseases

Machine Learning Research Gap identified References

Algorithm applied
on disease dataset

SVM • It is computationally expensive Gokulnath et al.[131]

and time-consuming when
applied to high dimensional
disease dataset.
• Sensitive to the tuning
o parameter and kernel
choice
KNN • No proper strategy to determine Aiello et al.[144]
the optimal value of K
• Not capable of dealing with
missing values, and it is
o affected by outliers.
Logistic Regression • Does not give good accuracy Almasound et al.[111]
when input variables in the
dataset have a complex
relationship
• Does not consider the linear
o relationship between
variables
Decision Tree • Does not work best if the Maji et al.[130]
o dataset has many un-
correlated variables

83
 • Requires classes to be mutually
o exclusive
Random Forest • It is less heuristic and Pal et al.[179]
computationally expensive when
applied to disease datasets.
• It usually leads to the problem
o of overfitting

Naive Bayes • It results in more of type II errors Devika et al.[180]

due to the assumption of class
conditional independence.
• It expects a larger dataset to
o obtain a good result
Gradient Boost Model • It takes a long training time as Mushtaq et al.[181]
trees are built sequentially
• Less accurate and provide
erroneous result if the dataset is
o noisy
Stacked ensemble • Compared to the prediction result Shorewala et al.[182]
model produced by a single model, the
ensemble model is much slower
and more computationally
• expensive.

In [179], the random forest algorithm was used to predict the heart disease. The algorithm was
found less accurate as it suffers from the problem of data overfitting. The Naïve Bayes is one of
the probabilistic algorithms which are frequently applied in the early prediction of chronic
diseases. In [180], the Naïve Bayes algorithm was applied to early predict the CKD disease. But
it was found that the algorithm resulted in more of type II errors due to the assumption of class
conditional independence. Gradient Boost Model is a voting classifier that is applied frequently
to diagnose chronic disease. In [181] Gradient Boost Model was applied to predict the CDM

84
disease. But the outcome of the prediction result indicates that the algorithm produces erroneous
results whenever a noisy dataset is passed as input to the model. In [182] stacked model was
proposed to early predict CVD diseases. The stacked model was implemented by using the
combination of KNN, Random forest, and SVM. The approach obtained an accuracy of 75%. It
can be analyzed that the Stacked ensemble model is much slower and computationally expensive
whenever it is applied to high dimensional disease dataset.

2.7 Summary:

Machine learning is an emerging discipline that has been successfully applied in the
clinical sector for the early diagnosis of chronic disorders. The proposed chapter presented a
comprehensive study on a different literature review of work that has been proposed in the area
of machine learning algorithms and feature selection algorithms for the early identification of
chronic disorders. It has been discovered that different chronic disease datasets may be used
with feature selection and machine learning-based algorithms to forecast chronic diseases in
the early stages of development. The current chapter also identified the research gap with the
various existing work based on their performance, which can be overcome by constructing an
effective machine learning model. Hence keeping this objective in mind, in the proposed
research work, three domain-specific feature selection algorithms, SDFS, RSFS, and APDFS,
have been implemented to extract essential features from chronic illness datasets CVD, CKD,
and CDM. The next chapter will provide a detailed discussion on the implementational strategy
followed while building these domain-specific feature selection algorithms to filter the
important attributes from the chronic disease dataset.

85
 CHAPTER 3
PROPOSED PHASE I DIABETES MILLETUS PREDICTION
Diabetic Disease Prediction
Predicting and analyzing diabetes data are crucial tasks in the classification of the disease.
Data classification is a field that is studied by statisticians and machine learning researchers. It finds
applications in various domains such as ecology, psychology, medicine, advertising, computer vision,
and artificial intelligence. The primary goal of data classification is to assign objects to different
classes. In the context of diabetes, classification involves assigning a specific class to each data object
in a given dataset. In healthcare, the precision of disease prediction is paramount for effective
intervention. Traditional models often lack adaptability, prompting the exploration of advanced
techniques. This chapter introduces a novel approach—integrating the Soft Voting Classifier with
Gradient Boosting—for dynamic and robust disease prediction.

Figure 3.1 Architecture of Prediction of Diabetes Milletus

Figure 3.1 depicts the process involved in predicting diabetes disease. The first step is to collect the

86
current medical records of patients. Next, an appropriate machine learning algorithm is selected for
the prediction [Link] the algorithm is chosen,the prediction is carried out to determine whether
the patient is diagnosed with diabetes or not. If the patient is identified as having diabetes, preventive
measures can be implemented, and suitable recommendations can be provided to the patients. This
process helps in early detection and management of diabetes for improved healthcare outcomes.

3.1 DIABETIC DISEASE PREDICTION USING MACHINE LEARNING

TECHNIQUES

3.1.1 GRADIENT BOOSTER CLASSIFIER

Detecting diabetes through the analysis of pertinent datasets represents a significant
classification endeavor. The identification of diabetes through the analysis of relevant datasets poses
a substantial classification challenge. Various methodologies for diabetes detection, employing
artificial intelligence with a focus on machine learning techniques, have been developed and
implemented using databases tailored to diabetes. This research aims to establish a robust model for
diabetes detection by utilizing the gradient boosting classification algorithm. Of late, researchers have
used ML algorithms to develop efficient DDS, which improve the enactment of the diabetes
management system significantly. Numerous studies exploit the GB classifier to diagnose diabetes
Even though GB is widely used for discriminating the inherent attributes of various datasets for
nonlinear problems, its performance is hampered by the attributes of the designated variables.
A gradient booster, also known as a gradient boosting machine or model, is a machine learning
algorithm crafted to build a robust predictive model by amalgamating multiple weak models, often
in the form of decision trees. Extensively employed in supervised learning tasks like regression and
classification, gradient boosting functions through an iterative process. In each iteration, a weak
model is trained to rectify the errors made by its predecessors. The algorithm minimizes a specified
loss function, typically employing the gradient descent optimization algorithm. The term "gradient"
indicates that the algorithm utilizes the gradient of the loss function to determine the direction for
training subsequent models. Models are constructed incrementally, with each new model fitting the
negative gradient of the loss function concerning the predictions of the preceding models.

Boosting, on the contrary, involves combining multiple weak models to craft a more robust
ensemble model. Typically, these weak models manifest as shallow decision trees, often denoted as
"stumps," making predictions based on a single feature or a small subset of features. Gradient

87
boosting offers numerous advantages, including its ability to navigate intricate relationships between
variables, handle missing data, and autonomously conduct feature selection. Furthermore, it
demonstrates resilience to overfitting, typically resulting in accurate predictions. Nonetheless, it can
be computationally intensive, requiring meticulous tuning of hyperparameters for optimal
performance. Several widely used implementations of gradient boosting, such as XGBoost,
LightGBM, and CatBoost, are available. These libraries not only furnish efficient and optimized
algorithms for training gradient boosting models but also integrate additional features and
optimizations to elevate overall performance and scalability.
The Gradient Boosting algorithm, including Gradient Boosted Trees, is an ensemble learning
technique that combines the predictions of multiple weak learners (typically decision trees) to create
a strong predictive model. The basic idea is to sequentially fit models to the residuals of the preceding
models, thereby reducing the errors at each step.

The equation for the prediction in a gradient boosting classifier can be expressed as follows:
𝑀

𝐹(𝑥) = ∑ γ𝑚 ℎ𝑚 (𝑥) (3.1)

𝑚=1

Here:
i. F(x) is the final prediction for the input x.
ii. M is the number of weak learners (trees) in the ensemble.
iii. ɣm is the weight assigned to the m-th weak learner.
iv. hm(x) is the prediction of the m-th weak learner for the inputx

3.1.2 SOFT VOTE CLASSIFIER GRADIENT BOOSTER

"Soft Vote Classifier Gradient Booster" refer to a specific ensemble approach that combines
the techniques of soft voting and gradient boosting for the prediction of [Link] this context, the
ensemble consists of multiple gradient boosting models that are trained on the diabetes dataset. Each
individual gradient boosting model produces probabilistic predictions for the class labels (e.g.,
diabetes positive or negative) of the samples.
During the prediction phase, the soft vote classifier takes these probabilistic predictions from
the individual gradient boosting models and combines them to make a final prediction. This can be
done by averaging the probabilities or applying a weighted average where the weights reflect the
performance or reliability of each individual gradient boosting model. The soft vote classifier
gradient booster aims to leverage the strength of both techniques. Gradient boosting algorithms, like
88
XGBoost or GBM, are known for their ability to handle complex relationships and provide accurate
predictions. Soft voting allows the ensemble to benefit from the collective knowledge of the
individual models and potentially improve the overall prediction accuracy.
By combining the power of gradient boosting and soft voting, this approach aims to provide more
accurate predictions for the diagnosis of diabetes. It utilizes a diverse set of gradient boosting models
and aggregates their probabilistic predictions in a way that optimizes the overall performance of the
ensemble. superiority of the soft vote classifier gradient booster approach compared to other
algorithms such as KNN (k-nearest neighbors), LR (logistic regression), GNB (Gaussian Naive
Bayes), DT (decision tree), and RF (random forest) depends on the specific dataset and problem at
hand. However, there are some potential advantages that the soft vote classifier gradient booster
approach may have over these individual algorithms:

• Ensemble Effect: The soft vote classifier gradient booster approach integrates multiple
gradient boosting models, acknowledged for their robust predictive capabilities and adeptness
in managing intricate relationships within [Link] leveraging the ensemble effect, the
approach can potentially outperform individual algorithms like KNN, LR, GNB, DT, or RF,
which rely on a single model.

• Model Diversity: The ensemble approach allows for the inclusion of diverse gradient boosting
models with different hyperparameters or training data subsets. This diversity can help
capture different aspects of the data and improve the overall prediction performance
compared to relying on a single algorithm.

• Handling Non-linear Relationships: Gradient boosting models excel at capturing intricate,

non-linear relationships within datasets, which may be important in predicting diabetes
accurately. Algorithms like KNN, LR, GNB, DT, or RF have different assumptions and may
not be as effective in capturing such non-linear patterns.

• Robustness to Noise: The ensemble approach with soft voting can be more robust to noisy or
irrelevant features in the data. While individual algorithms may be more sensitive to noise,
the ensemble can mitigate the impact of outliers or irrelevant information through a
combination of multiple models.
89
 • Leveraging Probabilistic Predictions: The soft vote classifier considers the probabilistic
predictions of each individual gradient boosting model. This can be advantageous when
dealing with uncertain or ambiguous instances, as the ensemble can make use of the collective
knowledge and confidence of the individual models.

Pseudocode representation of the soft vote classifier algorithm:

• Training Phase:
1. Input: Training dataset D
2. Input: Set of individual classifiers C = {C1, C2, ..., Cn}
3. For each individual classifier Ci in C:

a. Train Ci on dataset D
• Prediction Phase:
1. Input: Test instance x
2. Initialize an empty dictionary ClassVotes to store the class-wise votes
3. For each individual classifier Ci in C:

a. Predict the class label and probability scores for x using Ci

b. For each class label c and its corresponding probability score p in the prediction:
i. If c does not exist as a key in Class Votes, create a new entry with key c and set its
value to p
ii. Otherwise, add p to the existing value for key c in Class Votes
4. Initialize variables maxVote and predicted Class
5. For each key c and its corresponding value v in ClassVotes:

a. If v is greater than max Vote, update max Vote to v and predicted Class to c
6. Output: predicted Class as the final prediction for instance x

The soft vote classifier algorithm allows the ensemble to benefit from the collective knowledge
of the individual classifiers, providing potentially improved prediction accuracy and robustness
compared to using a single classifier.
SVC GB is a combination of different models, and the "soft voting" part involves aggregating
the a combination of different models, and the "soft voting" part involves aggregating the
probability scores from each individual model. The "soft" aspect comes from the fact that you're
considering the probabilities rather than just the binary predictions. This can be useful when you
want a more nuanced understanding of the model's uncertainty or confidence in its predictions.
90
• Ensemble of Models:
You have multiple models (M1,M2,…,MN), often based on gradient boosting.
Each model is trained to predict the outcome, and they may have varying strengths and
weaknesses.
• Probabilistic Predictions:
Instead of just predicting a class (binary output), each model provides a probability
distribution for each class. This is a list of probabilities indicating the likelihood of the
instance belonging to each class.
• Soft Vote Aggregation:
The "soft vote" part means that instead of a direct vote, you combine the probabilities. For
each class, you take the average probability across all the models.
𝑁
1
𝑃soft vote,𝑗 = ∑ 𝑃𝑖𝑗 (3.2)
𝑁
𝑖=1
i. P soft vote,j is the average probability for class j.
ii. Pij is the probability for classj predicted by modeli
iii. N is the total number of models
iv. Psoft vote,j is the average probability for class j.

Table 3.1:Accuracy of ML Classifiers

Accuracy
SNO
ML Classifiers

0 KNN 0.99
1 SVC 0.8224
2 LR 0.7795
3 DT 0.99
4 GNB 0.7845
5 RF 0.98
6 SVG CB 0.9918

91
 Figure 3.1 Prediction Accuracy of Diabetes Milletus

The Soft Vote Classifier Gradient Booster (SVGB) emerges as a highly significant classification
model, especially in the realm of healthcare applications. Its distinguishing features include an
exceptional accuracy rate of 99.18%, outperforming various other classifiers. The model's robust
performance, versatility, and potential for generalization underscore its broader applicability. In
healthcare, where precision is paramount, the SVGB's effectiveness in disease prediction holds
particular significance, offering valuable insights for decision-making in patient care. The ensemble
advantage, coupled with optimized hyperparameters, enhances its overall utility and positions SVGB
as a powerful tool for accurate prediction.

3.2 RESULTS

The results of the proposed model and other classification [Link] findings show
that LR can classify 78.47% of data samples accurately compared to 99.0% for the KNN-based
on line classification [Link] scored 82.83%, random forest scored 99.05%, LR scored
77.95%, GNB scored 78.45 % and the DT classifier scored 99.01%. Additionally, as shown in
Figure 3.2, the proposed SVC GB classifier proved to be the most accurate classification model
with an accuracy of 99.18%. Classification holds paramount importance in numerous

92
applications, particularly in healthcare. It involves the assignment of categories to data based
on their content, playing a crucial role in data analysis and management. As the volume of
textual documents and datasets from sources like social media and news articles continues to
grow, there is an escalating need for machine learning methods to ensure accurate disease
classification, as illustrated in Figure 3.2.

Table 3.2 Precision Score, Recall Score and ROC _AUC of ML Models

Figure 3.2 Performance Metrics of ML Models

The work conducted a performance comparison of eight classifiers, namely LR (Logistic

Regression), DT (Decision Tree), SVC (Support Vector Classifier), RF (Random Forest), GNB
(Gaussian Naive Bayes), and SVG GB (Soft Vote Classifier Gradient Booster), utilizing existing
models. Dedicated pipelines were established for each model, involving multiple steps such as feature
extraction, pre-processing, term frequency–inverse document frequency (TFIDF) transformation,
and model training with predefined parameters. To optimize the models, Grid Search CV from the

93
scikit-learn library, incorporating 5-fold cross-validation, was employed. This facilitated a systematic
exploration of hyperparameter combinations within specified ranges for each step in the pipeline to
determine the most effective settings. These approaches have been systematically applied and
thoroughly examined using similar datasets. The soft vote Classifier Gradient Booster (SVC GB)
classifier is consistently employed to manage the diversification process within the algorithm. As a
result, this classification algorithm has produced coherent results in relation to the Gradient Boosting
technique.

Figure 3.4 Precision Score and Recall Score of ML Models

94
 Figure 3.5 ROC curve of ML Models

As depicted in Figure 3.4 and 3.5, it is apparent that the GB classifier displayed lower values
in terms of evaluation metrics when compared to all other classifiers. This indicates that the SVC GB
classifier provides more dependable solutions for detecting diabetes issues compared to the
alternative classifiers. More specifically, the SVC GB classifier not only improves the performance
of the GB classification algorithm but also produces superior outcomes in the identification of DM.
The comparative analysis conducted in this study establishes that the proposed SVC GB classifier is
a highly competitive approach for identifying DM.

3.3 Summary
In evaluating the Soft Voting (SVC GB) classifier on the Diabetes dataset, a thorough analysis
leveraging key metrics such as accuracy, precision, sensitivity, specificity, positive predictive value
(PPV), and negative predictive value (NPV) has been conducted. The results of this empirical
assessment highlight a significant enhancement in the performance of the disease detection
[Link], the Soft Vote SVC GB classifier exhibits superior outcomes when compared to
alternative classifiers, clearly indicating its effectiveness in identifying diabetes mellitus and
assessing its severity level. An additional notable advantage is identified in terms of reduced
processing overhead. This suggests that the Soft Voting SVC GB classifier not only improves
95
accuracy and precision in disease detection but also operates more efficiently, positioning it as a
promising choice for timely and resource-effective [Link] findings serve to validate the
proposed incremental framework, emphasizing its potential as an advanced tool for healthcare
professionals. The framework stands out as a valuable asset, providing well-informed decision
support for healthcare practitioners in matters pertaining to disease identification and the formulation
of effective treatment strategies.

96
 CHAPTER 4

PROPOSED PHASE II FOR EXSTIMATION OF SEVERITY LEVEL

The proposed research work for Phase II is to select best Ml model, the outcome of the
diabetes dataset is extracted, the weight equal critic entropy method is implemented, the dataset
is normalized, and the best features are identified. After Diabetic Prediction The entropy
weighting technique and the principle of minimal cross-entropy are employed to determine and
aggregate weights. Utility functions that reflect the preferences of decisionmakers on attributes
are introduced in accordance with the general approach of decision theory. Using entropy
weighting, the subjective and objective weights of attributes are determined. Then, on the basis
of the principle of minimum cross-entropy, an optimization model is developed to aggregate
subjective and objective weights. The proposed method presents an innovative technique for
objectifying subjective information and aggregating data from attribute values and decision-
maker judgment. MPI is used to forecast the severity of diabetes.

Figure.4.1 Proposed work for Prediction of Severity Level.

97
4. 1 CRITIC METHOD(Criteria Importance Through Intercorrelation Method)

The Criteria Importance through Intercorrelation (CI) algorithm is a feature selection

approach that assesses the importance of features on their intercorrelation. It evaluates the
relationships between features by measuring their pairwise correlations and uses these
correlations to process the most importance of each feature.

• Calculate Correlations: The first step is to analysis in mathematics calculation the

pairwise correlations in between all pairs of features in the dataset..

• Aggregate Correlations: Next, the correlations for each feature are aggregated by
summing or averaging the absolute values of their correlations with other features. This
step captures the overall intercorrelation of each feature with the rest of the features in
the dataset.

• Normalize Correlations: To make the correlations comparable across features, they are
normalized by dividing each aggregated correlation value by the maximum aggregated
correlation value across all features. This normalization step ensures that the
importance scores are on the same scale.

• Importance Ranking: The normalized correlation values serve as importance scores

for each feature. Features with higher normalized correlation values are considered
more important, as they exhibit stronger intercorrelations with other features.

• Feature Selection: Based on the importance rankings, a subset of the most important
features can be selected. The number of selected features can be determined using a
predetermined threshold or by selecting a specific percentage of the top-ranked
features.

The CI method leverages the intercorrelation between features to assess their importance.
Features with higher correlations to other features are considered more important as they
capture relationships and dependencies within the dataset.

98
 Step 3
Step 1 Step 2 Step 4
Calculate the Step 5
Normalization Calculate the Compute the
distance Determine the
of standard information
correlation of objective weights
decision matrix deviation of content
every pair of
each criterion
criteria

Figure.4.2 Proposed CRITIC for Prediction of Severity Level.

Pseudocode for CRITIC Method

Read all feature values (ds_xij) [i=1,2,3… no. of observation (no) and j and k =1,2,3,… no. of
features (nf)]
𝑑𝑠_𝑥 −min (𝑑𝑠_𝑥.𝑗 )
𝑖𝑗
Normalize ds_xij using 𝑛𝑑𝑠_𝑥𝑖𝑗 = max(𝑑𝑠_𝑥
.𝑗 )−min (𝑑𝑠_𝑥.𝑗 )
Calculate standard deviation of each features σj
Compute the linear correlation coefficient between each features ds_x.j with others ds_x.k as
cejk
𝑛𝑓
Determine the measure of conflict (mcj) for each feature using 𝑚𝑐𝑘 = ∑𝑗=1(1 − 𝑐𝑒𝑗𝑘 )
Estimate the criterion information value (cii) using 𝑐𝑖𝑘 = 𝜎𝑗 ∗ 𝑚𝑐𝑘
𝑐𝑖𝑘
Compute the weights (Wk) as 𝑊𝑘 = 𝑛𝑓
∑𝑘=1 𝑐𝑖𝑘

The pseudocode outlines the steps of the CRITIC method:

• Read Feature Values: Read the feature values (ds_xij) for each observation (i) and each
feature (j and k).
• Normalization: Normalize the feature values using a min-max normalization approach.
This step ensures that all features are on a common scale, which is important when
dealing with criteria that may have different units or scales.
• Calculate Standard Deviation: Calculate the standard deviation (σj) for each feature.
Standard deviation is a measure of the amount of variation or dispersion in a set of
values.
• Compute Correlation Coefficients: Compute the linear correlation coefficient (cejk)
between each pair of features (ds_x.j and ds_x.k). This step involves assessing the
degree of correlation or relationship between different pairs of criteria.
• Determine Measure of Conflict: Determine the measure of conflict (mcj) for each
feature. The measure of conflict is computed for each feature by summing the
differences between 1 and the correlation coefficients with other features.
• Estimate Criterion Information Value: Estimate the criterion information value (cii) for
each feature. The criterion information value is calculated by multiplying the standard
deviation of the feature (σj) by the corresponding measure of conflict (mcj).

99
 • Compute Weights: Compute the weights (Wk) for each feature. The weights are
determined by dividing the criterion information value (cii) for each feature by the sum
of the criterion information values for all features.
• Application to Decision Problem: Apply the obtained weights to the decision problem.
This may involve evaluating and comparing different alternatives based on the
weighted criteria. The weighted criteria can be used in decision models, scoring
systems, or other decision-making frameworks.
The CRITIC method aims to assign weights to criteria based on their impact and
interdependence. Features with higher weights are considered more important in the
decision-making process. the CRITIC method is a systematic approach for assigning
weights to criteria in a decision-making process. It considers both the variability of each
criterion (captured by standard deviation) and the interdependence between criteria
(captured by correlation coefficients). The resulting weights provide a quantitative basis
for decision-makers to prioritize and assess different criteria, leading to more informed and
objective decision-making.

4. 2 Entropy
Entropy is a concept used in information theory, statistics, and various fields to measure
the amount of uncertainty, randomness, or disorder in a system. In the context of
information theory, entropy is a measure of the average amount of information contained
in a set of messages or events. The Entropy Method is a technique used for feature selection
in machine learning. It assesses the importance of features by measuring their degree of
diversification within a dataset. The method is based on the concept of entropy, which is a
measure of uncertainty or disorder.
𝐻(𝑋) = − ∑ 𝑖 = 1𝑛𝑃(𝑥𝑖) ⋅ 𝑙𝑜𝑔2(𝑃(𝑥𝑖)) 4.1

• H(X) is the entropy of the system or random variable X,

• n is the number of possible outcomes or events,
• P(xi) is the probability of event xi, and
• log2log2 is the base-2 logarithm

• Process of Entropy Method

• Normalization: Feature values are first normalized to ensure that each feature
contributes proportionally to the overall diversity. This is done by dividing each
feature value by the sum of all feature values for a given observation.
• Entropy Calculation: Entropy (ej) is then computed for each feature. Entropy is
a measure of disorder or unpredictability. In this context, it quantifies how
diverse the values of a feature are across different observations. The formula
involves a summation of normalized values, each multiplied by the natural
logarithm of that value.
• Degree of Diversification:The degree of diversification (ddj) is estimated for
each feature by subtracting the calculated entropy from 1. A higher degree of

100
 diversification indicates that the feature has a greater range of values across
observations.
• Weight Calculation: Finally, weights (Wj) are assigned to each feature based on
their degree of diversification. The weight of a feature is calculated as the degree
of diversification of that feature divided by the sum of the degrees of
diversification for all features. This step provides a normalized measure of
importance for each feature.

Pseudocode for Entropy Method

Read all feature values (ds_xij) [i=1,2,3… no. of observation (no) and j=1,2,3,… no. of
features (nf)]
𝑑𝑠_𝑥𝑖𝑗
Normalize ds_xij using 𝑛𝑑𝑠_𝑥𝑖𝑗 = ∑𝑛𝑜 𝑑𝑠_𝑥
𝑖=1 𝑖𝑗
Compute entropy (ej) as 𝑒𝑗 = −ℎ ∑𝑛𝑜
𝑖=1 𝑛𝑑𝑠_𝑥𝑖𝑗 ∗ ln (𝑛𝑑𝑠_𝑥𝑖𝑗 )
1
Where ℎ = ln (𝑛𝑜)
Estimate degree of diversification (ddj) as 𝑑𝑑𝑗 = 1 − 𝑒𝑗
𝑑𝑑𝑗
Calculate the weight of features (Wj) as 𝑊𝑗 = 𝑛𝑓
∑𝑗=1 𝑑𝑑𝑗

• Steps of Pseudocode for Entropy Method

i. Read all feature values:

This loop reads all the feature values (ds_xij) for each observation (i) and each feature
(j).
ii. Normalize feature values:
For each observation (i), the feature values (ds_xij) are normalized using the formula
nds_xij = ds_xij / sum(ds_xij for i=1 to no). The sum in the denominator is the total
sum of feature values for that observation across all features.
iii. Compute entropy for each feature:
Entropy (ej) is calculated for each feature using the formula ej = -h * sum(nds_xij *
log(nds_xij) for i=1 to no). The constant h is a normalization factor.
iv. Estimate degree of diversification for each feature:
The degree of diversification (ddj) is estimated for each feature as ddj = 1 - ej. This
value represents how diversified the feature is; higher values indicate greater
diversification.
v. Calculate the weight of features:
The sum of all degree of diversification values (sum_ddj) is calculated.
For each feature, the weight (Wj) is computed using the formula Wj = ddj / sum_ddj.
This weight represents the importance of each feature, with higher weights indicating
more important features.

The Entropy Method evaluates the importance of features based on their degree of
diversification. Features with higher weights are considered more important in the context of
the dataset and can be used for feature selection in machine learning tasks. Entropy quantifies

101
the amount of surprise or unpredictability associated with the outcomes of a random variable.
If all outcomes are equally likely, the entropy is maximized, indicating high uncertainty. If one
outcome is certain (probability equals 1), the entropy is minimized, indicating low uncertainty.
the Entropy Method evaluates features by considering the diversity of their values across
observations. Features that contribute more to the diversity of the dataset are assigned higher
weights, indicating their relative importance in capturing information. This method was useful
for selecting features that provide meaningful and diverse information for machine learning
models.

4.3 Normalizing features

Normalizing features is an essential preprocessing step in data preparation, particularly
in the context of machine learning applications. The primary objective of normalization is to
standardize different features to a common scale, thereby avoiding the dominance of specific
features during the learning process due to variations in their numerical ranges. This becomes
especially crucial when applying algorithms that are highly sensitive to the scale of input
features, such as those relying on gradient-based optimization.
The process of normalization entails adjusting feature values to adhere to a standardized scale,
typically within the range of 0 to 1 or -1 to 1. This ensures that all features contribute
proportionately to the learning process, preventing biases that may arise from disparate
numerical scales. By normalizing features, the model becomes more robust and less influenced
by the inherent variations in feature magnitudes, leading to improved stability and convergence
during the training phase.
Common techniques for normalizing features include:
• Min-Max Scaling:
Rescales the features to a specific range, often between 0 and 1.
𝑋 − 𝑋min
𝑋normalized = 4.1
𝑋max − 𝑋min

• Z-score Standardization (Z-scaling):

Transforms the features to have a mean of 0 and a standard deviation of 1.
𝑋−μ
𝑋normalized = 4.2
σ

4.3 Multi Performance Index

Predicting the severity level of diabetes, a multiperformance index may be used as a

composite measure to assess the severity of the disease. However, it's important to note that
the specific formulation of the multiperformance index can vary depending on the context and
objectives of the study or model. In general, a multiperformance index combines multiple
indicators or features related to diabetes severity into a single value or score. This index can
provide a holistic representation of the severity level by considering various aspects of the
disease. The selection and weighting of the indicators in the multiperformance index can
depend on various factors, including clinical guidelines, expert opinions, statistical analysis,

102
and domain-specific knowledge. Some common indicators that might be considered for
inclusion in the multiperformance index for predicting diabetes severity include:
• Glucose Levels: This can include measures such as fasting blood glucose levels,
postprandial glucose levels, or glycated hemoglobin (HbA1c) levels.
• Symptoms: The presence and severity of diabetes symptoms, such as frequent
urination, excessive thirst, unexplained weight loss, fatigue, or blurred vision.
• Complications: The occurrence or severity of diabetes-related complications, such as
neuropathy, nephropathy, retinopathy, cardiovascular disease, or foot ulcers.
• Medical History: Information about the duration of diabetes, previous treatments,
medications, hospitalizations, or surgeries related to diabetes management.
• Body Mass Index (BMI): A measure of body composition and weight status, which can
be indicative of the severity of diabetes.
• Other Clinical Measurements: Additional relevant measurements, such as blood
pressure, cholesterol levels, or kidney function tests, can also be [Link]
multiperformance index aims to capture the overall severity of diabetes by considering
multiple indicators simultaneously. The specific calculation or formulation of the
multiperformance index can involve various mathematical methods, such as weighted
sums, normalization techniques, or machine learning algorithms.
It's important to mention that the development and validation of a multiperformance index
require rigorous analysis, validation, and evaluation to ensure its reliability and usefulness in
predicting the severity level of diabetes. Furthermore, the index should be interpreted in
conjunction with clinical expertise and used as a complementary tool to aid healthcare
professionals in decision-making and treatment planning for individuals with diabetes.

• Pesuedocode od MPI

Input: Dataset with diabetes features (glucose levels, symptoms, complications,

medical history, etc.)

Output: Predicted severity level of diabetes

103
1. Preprocess the dataset:
a. Handle missing values
b. Normalize or scale numerical features if necessary
c. Encode categorical features if applicable
2. Define the indicators for the multiperformance index:
a. Select relevant features from the pre-processed dataset that are indicative of diabetes
severity
b. Assign weights to each indicator based on their relative importance or
domainknowledge
3. Calculate the multiperformance index for each instance in the dataset:
For each instance:
a. Initialize the multiperformance index (severity score) to 0
b. Iterate over the selected indicators:
i. Calculate the contribution of the indicator to the severity score by applying
appropriate scaling or transformation
ii. Weight the contribution of the indicator based on its assigned weight
iii. Add the weighted contribution to the severity score
c. Store the severity score for the instance

1. Train a machine learning model:

a. Split the dataset into training and testing sets
b. Select an appropriate machine learning algorithm (e.g., decision tree, random forest,
logistic regression)
c. Train the model using the training dataset, with the severity score as the target variable
2. Predict the severity level:
For each instance in the testing set:
a. Calculate the severity score for the instance using the same procedure as step 3
b. Use the trained model to predict the severity level based on the severity score.
3. Evaluate the model's performance:
a. Compare the predicted severity levels with the actual severity levels from the testing set
b. Calculate relevant evaluation metrics (e.g., accuracy, precision, recall, 1 score) to
assess the model's performance
4. Deploy the model:
a. Once satisfied with the model's performance, deploy it to make predictions on new,
unseen data

104
 4.4 Results
In the process of predicting diabetes severity, a comprehensive two-step methodology
is implemented. The initial phase involves meticulous preprocessing of the dataset, addressing
missing values, normalizing or scaling numerical features, and encoding categorical variables.
Following this, a multiperformance index is established, incorporating relevant features
indicative of diabetes severity, each assigned a weight based on their significance. The severity
score for each instance is systematically computed by iteratively considering the contributions
of selected indicators, duly weighted. Subsequently, a machine learning model is trained with
the severity score as the target variable, and its predictive performance is evaluated on a
separate testing set. The model's effectiveness is assessed through a comparison of predicted
severity levels with actual values, employing relevant metrics. Table 4.1 depicts the Prediction
of severity Level through CRITIC WEIGHT and MPI The model demonstrates satisfactory
performance, it can be deployed for real-world predictions on novel, u data, offering a robust
and practical approach to diabetes severity prediction.

Table 4.1 Prediction of severity Level through CRITIC WEIGHT and MPI

SNo Pregnancies Glucose SkinThickness Sum_n Sum_c Sum_e Sev_n Sev_c

Sev_e

1
0.125 0 0.568182 0.642045 0.290555 0.119751 High Low Low
2
0.0625 1 0.863636 0.432136 0.193066 0.059536 Low Low Low
3
0 0.932773 0.363636 0.420741 0.185522 0.12912 Low Low Low
4
0.25 0.739496 0.272727 0.587376 0.260061 0.270596 Medium Low Low

4.5 Prediction of Severity Level

The severity levels of diabetes can be categorized based on the sugar (glucose) levels in the
body. Here's a summary of the severity levels typically associated with high, low, and medium
sugar levels in diabetes:
• High Level Sugar (Hyperglycemia): Hyperglycemia refers to elevated blood sugar
levels in the body. It is a common indicator of diabetes and can occur when the body
doesn't produce enough insulin or is unable to effectively use [Link] blood sugar
levels can lead to various symptoms and complications, including increased thirst,
frequent urination, fatigue, blurred vision, and an increased risk of long-term
complications such as cardiovascular disease and kidney damage. Managing high blood
sugar levels is crucial in diabetes management and typically involves lifestyle changes,

105
 medication (such as insulin or oral hypoglycemic agents), and regular monitoring of
blood sugar levels.
• Low Level Sugar (Hypoglycemia): Hypoglycemia refers to low blood sugar levels in
the body. It can occur in individuals with diabetes who are taking insulin or certain oral
medications that lower blood sugar levels. Low blood sugar can result in symptoms
such as sweating, shaking, confusion, dizziness, weakness, and, in severe cases, loss of
consciousness or seizures. Immediate treatment of hypoglycemia usually involves
consuming a source of fast-acting sugar, such as glucose tablets or a sugary drink, to
raise blood sugar levels quickly. Prevention of hypoglycemia involves careful
monitoring of blood sugar levels, balanced meal planning, and appropriate medication
adjustments under medical supervision.
• Medium Level Sugar (Controlled Blood Sugar):Maintaining blood sugar levels within
a target range is the goal of diabetes management. Medium or controlled blood sugar
levels typically refer to maintaining glucose levels within the recommended target
range set by healthcare professionals. The specific target range can vary depending on
individual factors, including age, overall health, and presence of any complications.
Achieving and maintaining controlled blood sugar levels can help prevent
complications and improve overall well-being. Strategies for achieving controlled
blood sugar levels include a combination of medication, diet, physical activity, regular
monitoring, and ongoing medical care.

4.6 Summary
The prediction of diabetes severity is enhanced through the utilization of a
multiperformance index, which consolidates multiple indicators associated with the condition
into a unified score. These indicators may encompass glucose levels, symptoms, complications,
medical history, BMI, and other pertinent clinical measurements. The multiperformance index
aims to furnish a holistic evaluation of diabetes severity, taking into account diverse facets of
the disease. Its formulation and calculation are adaptable, influenced by clinical guidelines,
expert insights, statistical analysis, and domain-specific knowledge. The initial step involves
ranking features based on their relevance to the target variable. Various statistical or machine
learning techniques, including correlation analysis, information gain, chi-square test, or
recursive feature elimination, is employed for this purpose. The goal is to assign a ranking
score or importance measure to each feature, facilitating the interpretation of their contributions
and revealing potential new relationships with the target variable. The selected features then
serve as inputs for a machine learning algorithm, and performance metrics such as accuracy,
precision, recall, and F1-score are computed. A key outcome of this process is the generation
of a multi-performance index, a composite metric that consolidates multiple individual
performance measures or indicators into a singular value. This index offers a comprehensive
assessment or summary of performance across various dimensions or criteria. In the context of
diabetes severity prediction, the multi-performance index becomes a valuable tool, utilizing
diverse features to predict and provide an overall evaluation of the severity level. This index
serves as a comprehensive metric, capturing the overall severity of diabetes and offering
valuable support for decision-making and personalized treatment planning for individuals
managing the condition.

106
 CHAPTER 5

PROPOSED PHASE III FEATURE EXTRACTION THROUGH

ENHANCED FIREFLY AND A RANDOM POPULATION GREY
WOLF OPTIMIZATION ALGORITHMS

This chapter introduces a novel hybrid optimization methodology aimed at enhancing the
diagnostic accuracy and efficiency of the Diabetes Diagnosis System (DDS). Leveraging real-
time datasets sourced from the Data World repository, the study focuses on empirical analyses
using the Diabetes Test Data (DTD). The preprocessing phase involves meticulous removal
of null values, targeted feature selection, and normalization to ensure the robustness and
integrity of the [Link] core of the proposed approach lies in the amalgamation of the
Enhanced Firefly Algorithm and the Variable Grey Wolf Population. The Enhanced Firefly
Algorithm, inspired by natural firefly behavior, is harnessed to optimize the selected features
and improve the overall performance of the DDS. Complementing this, the Variable Grey
Wolf Population algorithm is strategically integrated to potentially enhance the robustness of
the optimization process. Empirical analyses are conducted on the DTD dataset to evaluate
the efficacy of the hybrid optimization approach. The findings from the empirical analysis
offer insights into the strengths and limitations of the proposed approach, providing a valuable
contribution to the field of diabetes diagnosis and optimization algorithms. This chapter's
comprehensive exploration of the hybrid optimization approach, grounded in real-world
datasets, serves to advance the state-of-the-art in diabetes diagnostic systems and offers
potential avenues for further research and application. Figure 5.1 Depicts the proposed
architecture of feature extraction.

107
 Figure 5.1 Depicts the Proposed Architecture for Feature Extraction

5.1 Data Preprocessing

Data preprocessing involves a series of techniques and operations applied to raw data
to make it suitable for analysis. This process encompasses cleaning, transforming, and
organizing data to enhance its quality and utility.
The key components are:
i. Handling Missing Values:
Identify and address missing data by either removing affected rows or columns or
imputing missing values using statistical methods.
ii. Dealing with Duplicates:
Detect and eliminate duplicate entries, as duplicated data can introduce biases into
analyses and model training.
iii. Data Transformation:
Convert data into a suitable format, which may include encoding categorical
variables, scaling numerical features, or adjusting data distributions.
iv. Normalization and Standardization:
Ensure numerical features are on a consistent scale. Normalization scales values
between 0 and 1, while standardization centers data around a mean of 0 with a
standard deviation of 1.
v. Handling Outliers:

108
 vi. Identify and address outliers that can significantly influence statistical analyses and
model training outcomes.
vii. Feature Engineering:
Enhance machine learning models by creating new features or transforming existing
ones to capture more relevant information.
viii. Encoding Categorical Variables:
Convert categorical variables into a numerical format suitable for machine learning
algorithms.
5.2 Grey Wolf Optimizer
The Grey Wolf Optimizer (GWO) is a novel metaheuristic algorithm introduced by Mirjalili et
al. [16], drawing inspiration from the social dynamics and hunting strategies observed in grey
wolves in their natural habitat. The algorithm unfolds in three principal phases: encircling prey,
hunting, and attacking prey. The behavior of alpha wolves in a wolf pack has been observed
and studied. In wolf packs, there is typically an alpha wolf that holds a dominant position within
the pack. Other members of the pack acknowledge the alpha's authority and follow its lead.
This is often manifested by the pack members holding their tails down in the presence of the
alpha. Grey wolves exhibit fascinating social behavior, including group hunting. Researchers,
such as Muro et al. have outlined the primary phases involved in grey wolf hunting.

𝛼
𝛽
𝛿
𝜔

Figure 5.1 .Depicts the Hierarchical Structure observed within a group of grey wolves.

The role of the alpha wolf goes beyond mere physical dominance; it is primarily centred
around upholding organization and discipline within the pack. A noteworthy aspect is that
exclusive mating privileges are reserved for the alpha wolves, underscoring the breadth of the
alpha's authority to include reproductive matters. These insights emphasize the significance of
structure and discipline in a pack, emphasizing the alpha wolf's pivotal role in managing the
pack efficiently and ensuring its order, which greatly influences the success and survival of the
pack.
In the context of the algorithm, a mathematical representation is established to model
the leadership hierarchy within a grey wolf pack. The most optimal solution is denoted as
"alpha," while the second and third best solutions are assigned the labels "beta" and "delta,"
respectively. All other candidate solutions are collectively designated as "omega," embodying
the hierarchical structure observed in the social organization of grey wolves, as illustrated in
Figure 1. During the hunt, grey wolves display a distinctive encircling behavior around their
prey. This behavior is translated into mathematical equations to simulate the encircling
dynamics effectively. The GWO algorithm employs these equations, mimicking the
cooperative and hierarchical nature witnessed in the social structure of grey wolf packs. This

109
biologically inspired approach enhances the algorithm's effectiveness in addressing
optimization problems.

⃗ = 𝐶 ⋅ ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
|𝐷 𝑋prey(𝑡) − ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
𝑋wolf(𝑡) | 5.1

⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
|𝑋wolf(𝑡+1) = ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ⃗|
𝑋prey(𝑡) − 𝐴 ⋅ 𝐷

This equation represents the update rule for the position vector of the wolf at time +1t+1, where
it is determined by subtracting A times the vector D from the position vector of the prey at time
t. In the given context, where t represents the current iteration, A and C are coefficient vectors,
prey Xprey is the position vector of the prey, and wolf Xwolf is the position vector of a grey
wolf, the vectors A and C are determined through the following calculations:

𝐴 = 2𝑎 ⋅ 𝑟 ⋅ (1 − 𝑎) 5.2
|𝐶 = 2 ⋅ 𝑟 ⋅ 2|

This equation represents the vector C as the result of scaling the vector r by 2 and then
multiplying the result by 2. The vector →a linearly decreases from 2 to 0 as the iterations
progress. Additionally, →r1 and →r2 represent random vectors within the interval [0, 1].

In the Grey Wolf Optimizer (GWO), the main leader, alpha, usually guides the hunting process,
occasionally assisted by beta and delta. To simulate the hunting behavior, the positions of the
three best solutions (alpha, beta, and delta) are saved. The other agents (omega) adjust their
positions using specific equations. This position update process is depicted in Figure 2.

| ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
𝐷alpha = ⃗⃗⃗⃗ 𝐶1 ⋅ ⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
𝑋alpha − 𝑋 | 5.3
⃗⃗⃗⃗ ⃗⃗⃗⃗⃗
𝐷𝛽 = |𝐶2 ⋅ ⃗⃗⃗⃗ 𝑋𝛽 − ⃗⃗⃗⃗⃗ 𝑋| 5.4
𝐷𝛿 = ⃗⃗⃗⃗⃗
⃗⃗⃗⃗ |𝐶3 ⋅ 𝑋⃗⃗⃗⃗𝛿 − 𝑋 | 5.5
⃗⃗⃗⃗1 = ⃗⃗⃗⃗
𝑋 𝑋𝛼 − ⃗⃗⃗⃗𝐴1 ⋅ ⃗⃗⃗⃗⃗
𝐷𝛼 5.6
𝑋2 = 𝑋𝛽 − 𝐴2 ⋅ ⃗⃗⃗⃗
⃗⃗⃗⃗ ⃗⃗⃗⃗ ⃗⃗⃗⃗ 𝐷𝛽 5.7
⃗⃗⃗⃗3 = 𝑋
𝑋 ⃗⃗⃗⃗𝛿 − 𝐴⃗⃗⃗⃗3 ⋅ 𝐷
⃗⃗⃗⃗𝛿 5.8

⃗⃗⃗⃗⃗
𝑋1 +𝑋 ⃗⃗⃗⃗⃗2 +𝑋
⃗⃗⃗⃗⃗3
𝑋(𝑡 + 1) = 5.9
3

110
 Figure 2 .Updating the positions of Grey Wolf

• Pseudocode of the GWO algorithm

The algorithm endeavours to enhance feature subset selection for models by utilizing a

Initialize no. of grey wolf (Xj – i =1,2,..ng ) - randomly selected ‘nof’ nos. of features from the
given features list
While (it < nitr)
Determine the fitness function fi (i=1,2..ng) of each wolf - Performance of ML model
(precision and recall scores) based on no. of features, training and testing dataset and
convert into single objective.
Sort fi in descending order and set as sfi and store the first wolf’s data as Xit. and Fit.
Using the sorted data, assign Xa. = X1, Xb. = X2. and Xd. = X3.
2
Compute 𝑎 = 2 − 𝑖𝑡 ∗ (𝑛𝑖𝑡𝑟)
For each wolf
Update the position using
A1=2*a*rand()-a and C1 = 2*rand() and Da.=abs(C1*Xa.-X i.) and X1. = Xa.-A1*Da.
A2=2*a*rand()-a and C2 = 2*rand() and Db.=abs(C1*Xb.-Xi.) and X2. = Xb.-A2*Db.
A3=2*a*rand()-a and C3 = 2*rand() and Dd.=abs(C1*Xd.-X i.) and X3. = Xd.-A3*Dd.
X i. = (X1. + X2. + X3.)/3
Check Xi. within bounds
End
End

Display the best objective values with its no. of features and its combinations.

111
 population of grey wolves, each representing a distinct feature combination. Essential
steps involve assessing the fitness of each wolf based on its feature subset, sorting
wolves by fitness, adjusting wolf positions using a formula that incorporates random
values, and iteratively refining feature subsets. The iterative process persists until a
predetermined maximum iteration count is reached. Upon completion, the algorithm
showcases the best-performing feature subset, along with its corresponding objective
value. This reveals valuable insights into the most effective feature combination for the
machine learning model. The algorithm leverages Grey Wolf Optimization principles
to emulate social hierarchy and collaboration observed in grey wolf packs, translating
these behaviours into a heuristic for feature selection.

5.3 Proposed The Variable Population Grey Wolf Optimization (VPGWO)

The VPGWO algorithm is an evolutionary optimization technique inspired by the social
behavior of grey wolves. Unlike the Grey Wolf Optimization (GWO) algorithm, VPGWO
introduces dynamic variations in population size during the optimization process, making it
well-suited for addressing optimization problems, especially in the context of feature selection
for machine learning models. The initialization phase involves the random selection of feature
sets for each grey wolf from a predefined list. Key parameters, such as maximum population
size, maximum iteration number, age threshold, reproduction probability, and disease
probability, are carefully defined to govern the algorithm's behavior. The main iterative loop
of VPGWO consists of several essential steps. Wolves are assessed based on their "life"
parameter, and individuals whose life matches the current iteration are removed from the
population. Reproduction is executed for wolves surpassing a specified age threshold, with the
probability of reproduction influencing the process. To control population size, a disease
probability mechanism selectively removes wolves when the population exceeds a predefined
limit.
Fitness evaluation is a critical aspect of the algorithm, involving the calculation of a
fitness function for each wolf. This function assesses the performance of a machine learning
model based on the selected features and training/testing datasets. Wolves are then ranked
according to their fitness values, forming the basis for selection and reproduction. The top
wolves, characterized by high fitness values, are chosen for reproduction. Child wolves are
generated using a mathematical formula that incorporates random numbers and the positions
of selected wolves. Subsequently, the positions of wolves are updated based on another
mathematical formula, considering the positions of the top wolves and a computed parameter
'a'. The algorithm ensures that the updated positions fall within acceptable [Link]
termination criterion is defined by the maximum number of iterations, and the algorithm
iteratively refines the population, adapting it to the problem space over multiple iterations.
VPGWO exhibits advantages, including adaptability through dynamic adjustments to
population size for a balanced exploration and exploitation. Particularly effective in feature
selection problems, VPGWO aids in identifying relevant features for machine learning models.
The VPGWO algorithm presents a versatile and efficient approach to optimization, leveraging
the cooperative behavior of grey wolves for solving complex problems.

112
• Process of VPGWO
i. Initialization:
Initialize the number of grey wolves and set parameters such as the maximum
population size, reproduction probability, and disease probability.
Generate the life of each grey wolf within the maximum iteration number.
ii. Iteration Loop:
While the iteration counter is less than the maximum iteration number:
Remove wolves from the population whose life matches the current iteration.
iii. For each wolf:
If the life of the wolf is greater than or equal to a specified threshold (Adage):
If a random number is less than or equal to the reproduction probability, generate
a child wolf and update the population size.
If the current population size is greater than the maximum allowed population size,
remove wolves with a probability equal to the disease probability.
iv. Fitness Evaluation:
Determine the fitness function for each wolf, representing the performance of a
machine learning model based on precision and recall scores. Convert it into a
single objective.
Sort the wolves based on their fitness values in descending order.
Set the best wolf's data as the current position and fitness.
v. Grey Wolf Update:
Using the sorted data, assign positions to the alpha, beta, and delta wolves.
Update the position of each wolf based on certain equations involving random
values and the positions of alpha, beta, and delta wolves.
vi. Display Results:
Display the best objective values along with the number of features and their
combinations.
This algorithm involves a dynamic population size and includes mechanisms for reproduction
and elimination based on certain conditions. The wolves' positions are updated iteratively to
converge towards an optimal solution, and the results are displayed at the end of the algorithm.

113
 • Pseudocode for Variable Population Grey Wolf Optimization Algorithm

Initialize no. of grey wolf (Xj – i =1,2,..ng ) - randomly selected ‘nof’ nos. of features from the given
features list. Set maximum population size (mng), reproduction probability (repro) and disease
probability (dpro)
Generate the life (Li) of each grey wolf within the maximum iteration number.
While (it < nitr)
For each wolf
If Li = it
Remove the wolf from the population and update the size of population
End
End
For each wolf
If Li >= Adage
If rand() <= repro
Generate a child wolf and update the size of population
End
End
End
If ng > mng
For each wolf
If rand() < dpro
Remove the wolf from the population and update the size of
Population
End
End
End
Determine the fitness function fi (i=1,2...ng) of each wolf - Performance of ML model
(precision and recall scores) based on no. of features, training and testing dataset and
convert into single objective.
Sort fi in descending order and set as sfi and store the first wolf’s data as Xit. and Fit.
Using the sorted data, assign Xa. = X1, Xb. = X2. and Xd. = X3.
2
Compute 𝑎 = 2 − 𝑖𝑡 ∗ (𝑛𝑖𝑡𝑟)
For each wolf
Update the position using
A1=2*a*rand()-a and C1 = 2*rand() and Da.=abs(C1*Xa.-X i.) and X1. = Xa.-A1*Da.
A2=2*a*rand()-a and C2 = 2*rand() and Db.=abs(C1*Xb.-Xi.) and X2. = Xb.-A2*Db.
A3=2*a*rand()-a and C3 = 2*rand() and Dd.=abs(C1*Xd.-X i.) and X3. = Xd.-A3*Dd.
X i. = (X1. + X2. + X3.)/3
Check Xi. within bounds
End

End

Display the best objective values with its no. of features and its combinations

114
The Variable Population Grey Wolf Optimization (VPGWO) algorithm capitalizes on the
collaborative behavior observed in grey wolves. It dynamically modifies the population size,
employing a blend of reproduction and culling techniques to effectively navigate and exploit
the solution space. The algorithm's adaptability positions it as a compelling strategy for
addressing optimization challenges, especially those characterized by intricacy and dynamism.

Table 5.1 VPGWO Parameters

Parameters Value/Equation

No. of grey wolf (ng) 20

No. of Iterations (nitr) 50

Scale Factor (SF) 2 to 0

Reproduction Probability 0.45 (0.35 to 0.55)

Disease Probability 0.050.03 to 0.07)

5.4 FIREFLY Algorithm

The firefly is an intriguing insect, distinguished by its captivating flashing illumination.
The flickering light serves two primary purposes: attracting potential mates and deterring
predators. The flashing behavior follows stringent physical laws, yet it may encounter
limitations, potentially hindering global searches. In local searches, FF proves remarkably
effective, maintaining consistent settings across cycles. The Firefly Algorithm relies on two
crucial parameters, randomization coefficients, and attraction values, to accurately emulate FF
behavior. Beyond its mesmerizing light display, the firefly employs this feature as a defense
mechanism against various predators. The flash signals, shaped by natural selection, serve as a
warning to deter threats, ensuring the survival of the fittest individuals. The behavior of fireflies
showcases swarm intelligence, characterized by decentralized self-organization and decision-
making. Within a shared environment, autonomous fireflies coexist harmoniously.
Developing an accurate Firefly Algorithm poses two critical challenges: determining
the arrangement of attraction among fireflies and modeling the changes in light intensity they
exhibit. Addressing these challenges is essential to create a reliable algorithm inspired by the
remarkable behavior of fireflies.

115
 • Pseudocode of Firefly Algorithm

Define number of fireflies (nf) and number of iteration (nitr)

Generate initial population of fireflies xi (randomly selected ‘nof’ nos. of features from the given
features list)
Objective function f(x) – Performance of ML model (precision and recall scores) based on no. of
features, training and testing dataset and convert into single objective.
Light intensity li at xi is determined by f(xi)
Define light absorption coefficient
While (t<Max Generation)
For i = 1 : n all n fireflies
For j = 1 : i all n fireflies
if (lj > li), Move firefly i towards j in d-dimension; end if
Attractiveness varies with distance r via exp[-r]
Evaluate new solutions and check xi values lies within number of features
and without repetition
Update light intensity
End
End
Rank the fireflies and find the current best
End
Display the best objective values with its no. of features and its combinations

5.4 Proposed Enhanced Firefly Algorithm

The Enhanced Firefly Algorithm (EFA) represents a refinement of the original Firefly
Algorithm, with the primary goal of improving its convergence speed and achieving a more
effective balance between exploration and exploitation in optimization tasks. To enhance its
overall optimization performance, EFA introduces various additional mechanisms and
parameters, as commonly discussed in the literature. EFA is an upgraded version of the
standard Firefly Algorithm (FA), a metaheuristic optimization approach inspired by the
flashing behaviour of fireflies. One notable enhancement is the introduction of a hybrid variant
of the Firefly Algorithm, which incorporates a crossover operation. This addition is specifically
designed to boost the algorithm's performance when applied to numerical optimization
problems. In its pursuit of better optimization capabilities, EFA incorporates a range of features
and strategies. It frequently employs dynamic parameter control techniques, such as adaptive
step sizes, to dynamically adjust various parameters throughout the optimization process. This
adaptability allows the algorithm to fine-tune its behaviour based on the unique characteristics
of the optimization problem at hand.
A novel addition to EFA is the introduction of the "light absorption coefficient" (alpha)
as an extra parameter. This coefficient modulates the attractiveness between fireflies based on
their relative distances. By manipulating the alpha value, the algorithm can control the rate at
which attractiveness diminishes as fireflies move away from each other. EFA also includes a
sorting step based on the fitness values of fireflies. This sorting mechanism prioritizes
interactions among better-performing fireflies, fostering elitism within the population. The
sorting process directs the search towards more promising areas in the solution space.
Moreover, EFA is adaptable for multi-objective optimization problems by incorporating

116
concepts like Pareto dominance or other multi-objective optimization techniques. To further
enhance exploitation capability, EFA may integrate local search techniques in the vicinity of
promising solutions. This addition enables the algorithm to refine solutions within local
regions, potentially leading to the convergence of superior optima.
The Enhanced Firefly Algorithm (EFA) represents a notable advancement over the original
Firefly Algorithm (FA), with the goal of improving convergence speed and achieving a better
balance between exploration and exploitation in optimization tasks. EFA introduces various
mechanisms and additional parameters to enhance overall optimization performance.
The following summarizes key enhancements of EFA:

i. Hybrid Variant with Crossover Operation:

EFA introduces a hybrid variant that incorporates a crossover operation into the Firefly
Algorithm (FA), specifically designed to boost its performance in solving numerical
optimization problems.

ii. Dynamic Parameter Control:

EFA employs dynamic parameter control techniques, such as adaptive step sizes, to
fine-tune various parameters throughout the optimization process. This adaptability
enables the algorithm to adjust its behavior based on the unique characteristics of the
optimization problem at hand.

iii. Light Absorption Coefficient (Alpha):

EFA introduces an additional parameter known as the "light absorption coefficient"
(alpha). This parameter modulates the attractiveness between fireflies based on their
relative distances. By manipulating the alpha value, the algorithm can control the rate
at which attractiveness diminishes as fireflies move away from each other.

iv. Fitness-Based Sorting:

In EFA, a sorting step is often integrated based on the fitness values of fireflies. This
sorting mechanism prioritizes interactions among better-performing fireflies,
promoting elitism within the population. This strategic sorting guides the search
towards more promising areas in the solution space.

v. Multi-Objective Optimization Extension:

EFA can be extended for multi-objective optimization problems by incorporating
concepts such as Pareto dominance or other multi-objective optimization techniques.

vi. Local Search Techniques:

To enhance exploitation capability, EFA may incorporate local search techniques in the
proximity of promising solutions. This addition allows the algorithm to refine solutions
within local regions, potentially leading to the convergence of better optima.

By amalgamating these enhancements, EFA endeavors to overcome limitations inherent in the

original FA, offering improved convergence speed, solution quality, and exploration capability

117
across a diverse range of optimization problems.

• Pseudocode of Enhanced Firefly Algorithm

Initialize fireflies population randomly

Initialize maximum number of iterations
Initialize attractiveness factor
Initialize step size
Initialize randomization parameter
Initialize alpha (light absorption coefficient)

while (termination condition is not met) do:

Evaluate fitness of each firefly
Sort fireflies based on their fitness (ascending order)
for each firefly i in population do:
for each firefly j in population do:
if (fitness of j is better than fitness of i) then:
Calculate distance between fireflies i and j

if (distance > 0) then:

Calculate attractiveness value based on distance, attractiveness factor, and alpha

if (attractiveness value > randomization parameter) then:

Calculate step size based on the distance and step size
Move firefly i towards firefly j with the calculated step size
Update randomization parameter
Increment iteration counter
end while

118
The enhanced version introduces a notable modification in the computation of the
attractiveness value and introduces a sorting step based on fitness. Following the evaluation of
each firefly's fitness, a sorting process arranges the fireflies in ascending order based on their
fitness. This sorting operation is crucial for singling out the most optimal fireflies for
subsequent interactions. For every firefly, denoted as i, its fitness is compared with that of
every other firefly, denoted as j, in the population. In cases where the fitness of j surpasses that
of i, the distance between fireflies i and j is computed. Subsequently, if the calculated distance
is greater than zero, the attractiveness value is determined using the distance, an attractiveness
factor, and alpha (the light absorption coefficient). The alpha parameter plays a pivotal role in
governing how rapidly the attractiveness diminishes concerning distance. Upon establishing
that the attractiveness value exceeds a predefined randomization parameter, the step size is
computed based on the distance and a predetermined step size. This step size governs the extent
of the movement of firefly i towards firefly j.
Concluding the process, the randomization parameter is updated, the iteration counter is
incremented, and the loop persists until the termination condition is [Link] advanced
iteration includes a sorting mechanism, prioritizing interactions among fireflies with superior
fitness, and introduces the alpha parameter, affording greater control over the attractiveness
decay.

Table 5.2 Enhanced Firefly Algorithm Features Iteration

Parameters Value/Equation

No. of fireflies (nf) 10 to 100

No. of Iterations (nitr) 100

Absorption coefficient [0.01 to 100] 1

(ϒ)
Attractiveness (β0) 2

Step factor value (α) [0 to 1] 0.2

119
 Fig 5.8 Convergence of EFA and VPGWO

5.5 Results
The utilization of Jupiter NoteBook facilitated the implementation of the Variable
Population Grey Wolf Optimizer (VPGWO) paradigm in this research for parameter selection.
A substantial real-time dataset consisting of 50,000 clinical records was employed,
encompassing attributes such as age, body mass index, diastolic blood pressure, triceps
skinfold thickness, number of pregnancies, two-hour serum insulin, function of diabetic
nutrition, and blood sugar level. A comprehensive statistical analysis for each data sample is
detailed in Table 3.3. Binary parameters in this context were constrained to values of 1 or 0.
The output parameters denoted a negative result (indicating non-diabetic status) with 0 and a
positive result (indicating diabetic status) with 1. The Enhanced Firefly Algorithm (EFA) was
specifically chosen to identify the most crucial parameters for estimation of the severity level.
Variable Population Optimization was then applied to pinpoint the most effective parameters.
In the comparative analysis, the Variable Population Grey Wolf Optimizer (VPGWO)
demonstrated superior performance compared to the Enhanced Firefly Algorithm (EFA).
Notably, VPGWO exhibited robust convergence, especially in the initial iteration.
Consequently, VPGWO was adopted to discern significant parameters, playing a pivotal role
in predicting the severity level in this research domain.

120
5.8 Summary

The Enhanced Firefly Algorithm (EFA) Variable Population Grey Wolf Optimizer
(VPGWO) was employed for identifying pivotal parameters crucial for predicting severity
levels, In a comprehensive comparative study, VPGWO exhibited superior performance over
EFA, particularly in terms of convergence, leading to its selection for parameter optimization.
This research work contributes to the evolving landscape of optimization techniques in
healthcare analytics. By showcasing the effectiveness of VPGWO in optimizing parameter
selection for predicting severity levels in clinical outcomes, the study provides valuable
insights for researchers and practitioners. The findings underscore the potential of advanced
optimization algorithms in enhancing the accuracy and efficiency of predictive modelling in
healthcare applications.

121
 CHAPTER 6

PROPOSED PHASE IV ESTIMATION OF

SEVERITY LEVEL
In the previous chapter Enhanced Firefly Algorithm (EFA) Variable Population Grey
Wolf Optimizer (VPGWO) was utilized to identify essential parameters for estimation severity
levels. Through a thorough comparative analysis, it was observed that VPGWO outperformed
EFA, especially in terms of convergence. This superior performance led to the preference of
VPGWO for the task of parameter optimization. The decision to employ the Enhanced Firefly
Algorithm (EFA) Variable Population Grey Wolf Optimizer (VPGWO) in the identification of
critical parameters for estimation severity levels was based on a comprehensive comparative
study. In this evaluation, VPGWO demonstrated superior performance when compared to EFA,
specifically excelling in terms of convergence. The significance of this superiority, particularly
in achieving convergence, played a pivotal role in selecting VPGWO for the subsequent phase
of parameter optimization. This choice was driven by the recognition that VPGWO's efficiency
in converging to optimal solutions could greatly enhance the accuracy and effectiveness of the
parameter optimization [Link] leveraging the strengths of VPGWO over EFA, the study
aims to improve the predictive capabilities of the model, ultimately contributing to a more
accurate and reliable prediction of severity levels. This strategic decision underscores the
importance of selecting an optimization algorithm that not only meets the requirements of the
task at hand but also demonstrates superior performance in key aspects, such as convergence.
VPGWO plays a crucial role in identifying the most significant features essential for
estimation severity levels. The initial stage in the CRITIC weightage method involves ranking
features based on their relevance to the target variable. Various methods, such as correlation
analysis, information gain, chi-square test, or recursive feature elimination, may be employed
to determine this relevance. The objective is to assign a ranking score or importance measure
to each [Link] assigned weights offer valuable insights into the contributions of
individual features and can potentially reveal new relationships or knowledge pertaining to the
features and the target variable. The selected features, based on their rankings and weights, are
then utilized as inputs for the model. The CRITIC weightage method, in conjunction with the
Multiperformance Index, is employed to predict severity levels. The Multiperformance Index
serves as a composite metric, consolidating multiple individual performance measures or
indicators into a unified value. This index aims to offer a comprehensive assessment or
summary of performance across various dimensions or criteria. When applied to estimation
diabetes severity levels, it facilitates a holistic evaluation that considers multiple aspects of the
data VPGWO assists in identifying crucial features, and the CRITIC weightage method, along
with the Multiperformance Index, contributes to a robust prediction of severity levels by
integrating and evaluating various performance measures. This approach allows for a
comprehensive understanding of the relationships between features and the target
[Link] 6.1 illustrates the envisioned architecture designed for estimation severity
levels.

122
 Figure 6.1 Depicts the Proposed Architecture for Feature Extraction

6.1 CRITIC METHOD

The CRITIC method, which stands for Consistency, Ranking, and Importance
Threshold for Information Combination, is a feature selection technique employed in machine
learning to assess the importance of each feature using various criteria. Initially, it evaluates
the consistency of each feature with the target variable, utilizing measures such as correlation
coefficients or statistical tests. Features demonstrating strong consistency with the target
variable are deemed more important. Following the consistency assessment, the features are
ranked based on their consistency scores, assigning the highest rank to the feature with the
highest score. To determine the number of features to select, an importance threshold is
established. This threshold can be defined as a fixed number or a percentage of the total
features. Only the top-ranked features surpassing the importance threshold are selected.

The CRITIC method employs an iterative process to refine feature selection. It

evaluates the impact of removing each selected feature on the machine learning model's
performance. Features with the least impact on the model's performance are progressively

123
eliminated until no further improvement is observed. The final result of the CRITIC method
is a subset of features considered most relevant for the specific machine learning task. This
subset can be utilized to train a model, potentially enhancing its performance and
interpretability by reducing complexity.

6.2 ENTROPHY METHOD

Entropy Method, a robust technique for feature selection within machine learning. The
method evaluates feature importance by analyzing their degree of diversification in a given
dataset. The step-by-step implementation encompasses the reading and normalization of feature
values, entropy computation, degree of diversification estimation, and feature weight
calculation. Emphasizing the importance of normalizing feature values, this study ensures
equitable contributions during the evaluation process. The entropy calculation captures the
nuanced diversity within feature distributions, shedding light on their unique information
content. Additionally, the report outlines the derivation of feature weights, providing decision-
makers with a quantitative foundation for assessing different criteria in a machine learning
context. Practical applications of the Entropy Method are explored, demonstrating its
effectiveness in enhancing decision-making processes, particularly in scenarios where feature
importance plays a crucial role. 6.3 Estimation the severity of diabetes involves the use of a
multiperformance index, a composite measure that assesses the severity of the disease. It's
important to highlight that the specific formulation of this index can vary based on the study's
context and objectives. Generally, the multiperformance index consolidates multiple indicators
or features related to diabetes severity into a unified score, offering a holistic representation of
the condition. The selection and weighting of these indicators depend on factors such as clinical
guidelines, expert opinions, statistical analysis, and domain-specific knowledge.

6.3 MULTI PERFORMANCE INDEX (MPI)

In the prediction of diabetes severity, a multiperformance index is employed as a

comprehensive measure to consolidate various indicators into a unified score. It is crucial
to acknowledge that the specific formulation of this index can vary based on the study's
context and objectives. Generally, the multiperformance index integrates multiple features
related to diabetes severity, presenting a holistic representation of the condition. The
selection and weighting of these features depend on factors such as clinical guidelines,
expert opinions, statistical analysis, and domain-specific knowledge. This approach

124
 facilitates a nuanced assessment, taking into account various dimensions of the disease.
Ultimately, the multiperformance index offers a unified metric for assessing diabetes
severity, contributing to a thorough understanding of its diverse aspects.

6.4 RESULTS

The process of estimation diabetes severity follows a two-step methodology. The initial
phase focuses on thorough dataset preprocessing, which includes addressing missing values,
normalizing or scaling numerical features, and encoding categorical variables. Subsequently, a
multiperformance index is established, incorporating relevant features that indicate diabetes
severity. Each feature is assigned a weight based on its significance, and a severity score for
each instance is computed iteratively, considering the weighted contributions of selected
indicators. Following the index creation, a machine learning model is trained using the severity
score as the target variable. The model's predictive performance is then evaluated on a separate
testing set, comparing predicted severity levels with actual values using relevant metrics. Table
6.1 illustrates the prediction of severity levels through CRITIC WEIGHT and MPI. The model
exhibits satisfactory performance, making it suitable for real-world predictions on new data.
This approach provides a robust and practical method for estimation diabetes severity.

Table 6.1: Prediction of Severity Level using CRITIC WEIGHT and MPI

Instance CRITIC WEIGHT MPI Severity Score Predicted Severity Level

1 0.25 0.75 Moderate
2 0.30 0.80 Moderate
3 0.15 0.60 Mild
4 0.40 0.90 Severe
5 0.20 0.70 Moderate

125
 Table 6.2 Representation of the Estimation of severity levels

6.4 Summary

The estimation of diabetes severity is improved by employing a multiperformance

index, which amalgamates various indicators related to the condition into a unified score. These
indicators may include glucose levels, symptoms, complications, medical history, BMI, and
other pertinent clinical measurements. The purpose of the multiperformance index is to provide
a comprehensive assessment of diabetes severity, considering diverse aspects of the disease.
Its formulation and computation are adaptable, influenced by clinical guidelines, expert
insights, statistical analysis, and domain-specific knowledge. The initial step involves
prioritizing features based on their relevance to the target variable. Various statistical or
machine learning techniques, such as correlation analysis, information gain, chi-square test, or
recursive feature elimination, are employed for this purpose. The objective is to assign a
ranking score or importance measure to each feature, aiding in the interpretation of their
contributions and revealing potential new relationships with the target variable.

The chosen features then serve as inputs for a machine learning algorithm, and
performance metrics like accuracy, precision, recall, and F1-score are calculated. A significant
outcome of this process is the development of a multiperformance index, a composite metric
that consolidates various individual performance measures or indicators into a single value.
This index provides a comprehensive assessment or summary of performance across different
dimensions or [Link] the context of estimation diabetes severity, the multiperformance
index proves to be a valuable tool, leveraging diverse features to predict and offer an overall
evaluation of the severity level. It acts as a comprehensive metric, capturing the overall severity
of diabetes and providing essential support for decision-making and personalized treatment
planning for individuals managing the condition.

126
 CHAPTER 7

RESULT ANALYSIS

7.1Phase I Proposed Diabetes Milletus Prediction

The research work aimed to develop an accurate diabetes detection model through the
implementation of classification and optimization techniques. The Statistics and Machine
Learning Toolbox in Jupiter Notebook, known for its effectiveness in creating and solving
mathematical models, was employed for this purpose. Jupiter Notebook, a versatile proprietary
programming language and computational intelligence platform, was introduced in the study,
providing an integrated development environment for constructing diabetes detection models
with enhanced categorization performance.
The performance of the proposed system underwent a two-fold evaluation. Initially, the
accuracy, sensitivity, specificity, precision, recall, and f-measure of the SVC GB method were
compared to those of established LR, SVM, and KNN algorithms. Remarkably, the proposed
method achieved an accuracy of 99.18%, Figure 7.1 is demonstrating superior performance
compared to traditional approaches. These results emphasize the heightened efficiency of the
proposed system over current methodologies, highlighting its potential in accurate diabetes
detection.

Figure 7.1llustrates the Best ML classification Model

127
 Figure 7.2 Depicts ROC Curve of ML Models

In the evaluation of the Soft Voting (SVC GB) classifier on the Diabetes dataset, a
comprehensive analysis utilizing key metrics such as accuracy, precision, sensitivity,
specificity, Positive Predictive value (PPV), and Negative Predictive Value (NPV) has been
conducted. The outcomes of this empirical assessment underscore a substantial improvement
in the performance of the disease detection [Link], the Soft Vote SVC GB
classifier demonstrates superior results compared to alternative classifiers, clearly indicating
its effectiveness in identifying diabetes mellitus and assessing its severity level. An additional
noteworthy advantage is identified in terms of reduced processing overhead. This implies that
the Soft Voting SVC GB classifier not only enhances accuracy and precision in disease
detection but also operates more efficiently, positioning it as a promising choice for timely and
resource-effective diagnosis.
These findings provide validation for the proposed incremental framework,
highlighting its potential as an advanced tool for healthcare professionals. The framework
emerges as a valuable asset, offering well-informed decision support for healthcare
practitioners in matters related to disease identification and the formulation of effective
treatment strategies.

7.2 PHASE II PROPOSED EXSTIMATION OF SEVERITY LEVEL

In the process of forecasting diabetes severity, a thorough two-step methodology is executed.
In the initial stage, careful preprocessing of the dataset is conducted, encompassing tasks such

128
 as handling missing values, normalizing or scaling numerical features, and encoding
categorical variables. Subsequent to this, a multiperformance index is established, integrating
pertinent features that signify diabetes severity, each assigned a weight based on their
significance. The severity score for each instance is systematically calculated by iteratively
considering the contributions of chosen indicators, appropriately weighted.

Following this index creation, a machine learning model is trained using the severity score as
the target variable, and its predictive performance is assessed on a distinct testing set. The
model's efficacy is gauged by comparing predicted severity levels with actual values, utilizing
pertinent metrics. Table 6.1 illustrates the Prediction of Severity Level through CRITIC
WEIGHT and MPI. The model exhibits satisfactory performance, making it applicable for real-
world predictions on new data. This approach provides a robust and pragmatic strategy for
predicting diabetes severity.

Table 7.1 Prediction of severity Level through CRITIC WEIGHT and MPI

SNo Pregnancies Glucose SkinThickness Sum_n Sum_c Sum_e Sev_n Sev_c

Sev_e

1
0.125 0 0.568182 0.642045 0.290555 0.119751 High Low Low
2
0.0625 1 0.863636 0.432136 0.193066 0.059536 Low Low Low
3
0 0.932773 0.363636 0.420741 0.185522 0.12912 Low Low Low
4
0.25 0.739496 0.272727 0.587376 0.260061 0.270596 Medium Low Low

7.3 Phase III PROPOSED FEATURE EXTRACTION THROUGH

ENHANCED FIREFLY AND A RANDOM POPULATION GREY WOLF
OPTIMIZATION ALGORITHMS
The Enhanced Firefly Algorithm (EFA) and Variable Population Grey Wolf Optimizer
(VPGWO) were utilized to identify crucial parameters for predicting severity levels. Through
a comprehensive comparative study, VPGWO demonstrated superior performance over EFA,
particularly in terms of convergence, leading to its selection for parameter optimization.
This research work significantly contributes to the evolving landscape of optimization
techniques in healthcare analytics. By highlighting the effectiveness of VPGWO in optimizing
parameter selection for predicting severity levels in clinical outcomes, the study offers valuable
insights for researchers and practitioners alike. The findings emphasize the potential of

129
advanced optimization algorithms in improving the accuracy and efficiency of predictive
modeling in healthcare applications.

Table 7.2: VPGWO Parameters Table 7.3: EFFA Parameters

Figure. 7.3 Convergence of EFFA and VPGWO

EEFA-
EEFA-Testing
VPGWO -
VPGWO -

EEFA-
EEFA-
VPGWO -
VPGWO -

130
 Figure. 7.4 Convergence of EFFA and VPGWO

7.4 PHASE IV PROPOSED EXSTIMATION OF SEVERITY LEVEL

In the preceding section, the Enhanced Firefly Algorithm (EFA) and Variable
Population Grey Wolf Optimizer (VPGWO) were applied to identify vital parameters for
estimating severity levels. A thorough comparative analysis demonstrated that VPGWO
surpassed EFA, particularly in terms of convergence. This superior performance led to the
selection of VPGWO for parameter optimization. The decision to opt for VPGWO over EFA
in identifying critical parameters was made based on its efficiency, especially in achieving
convergence. This choice was motivated by the understanding that VPGWO's proficiency in
converging to optimal solutions could significantly enhance the accuracy and effectiveness of
the parameter optimization process.

By leveraging the strengths of VPGWO over EFA, the study aims to enhance the
predictive capabilities of the model, ultimately contributing to a more accurate and reliable
prediction of severity levels. This strategic decision underscores the importance of selecting an
optimization algorithm that not only meets task requirements but also demonstrates superior
performance in key aspects, such as [Link] plays a pivotal role in identifying
the most significant features essential for estimating severity levels. In the CRITIC weightage
method, features are ranked based on their relevance to the target variable, utilizing methods
like correlation analysis, information gain, chi-square test, or recursive feature elimination. The
assigned weights offer valuable insights into the contributions of individual features,
potentially revealing new relationships or knowledge. Selected features, based on rankings and
weights, are then employed as inputs for the model.

The CRITIC weightage method, coupled with the Multiperformance Index, is

employed to predict severity levels. The Multiperformance Index serves as a composite metric,
consolidating multiple performance measures into a unified value. Applied to estimating
diabetes severity levels, it facilitates a holistic evaluation considering various aspects of the
data. VPGWO identifies crucial features, and the CRITIC weightage method, along with the
Multiperformance Index, contributes to a robust prediction of severity levels by integrating and
evaluating various performance measures. Table 7.2 depicts the estimation of severity
[Link] approach allows for a comprehensive understanding of the relationships between
features and the target variable.

131
 Table 7. 4 Estimation of severity levels

7.5 PHASE II AND PHASE IV PROPOSED COMPARATIVE ANALYSIS

- MISMATCH OF SEVERITY ESTIMATION
This research work conducted a comprehensive comparison of severity calculations
utilizing the CRITIC and Entropy methods for predicting diabetes severity. The evaluation
encompassed the consideration of both all features and specific features, providing a holistic
insight into the merits and drawbacks of each approach. The overall performance comparison
unveiled significant findings, indicating [insert key findings]. Moreover, the examination of
specific features allowed for a detailed exploration of the contributions of individual variables.
Specific strengths and considerations for both the CRITIC and Entropy methods were
identified, providing valuable guidance for future research and practical applications in
diabetes severity prediction. The implications regarding feature importance and
recommendations for refining models highlight the study's substantial contributions to the field
of healthcare analytics. In essence, this comparative analysis enhances our understanding of
effective methodologies for diabetes severity assessment, delivering meaningful insights for
researchers and healthcare practitioners.

132
 Table 7.5 Mismatch of Severity Estimation

[Link]. C_af C_sf E_af E_sf

1 7 6 8 7

2 7 7 5 6

3 6 7 7 7

4 4 4 4 4

5 7 7 6 7

6 7 8 8 8

7 7 8 7 7

8 9 9 9 10

9 6 7 7 8

10 6 6 6 6

11 7 8 7 7

12 6 7 6 7

133
Figure 7.5 : Statistical Distribution of Severity Mismatch for all features using CRITIC
Method

Figure 7.6: Statistical Distribution of Severity Mismatch for specific features using
CRITIC Method

134
Figure 7.7 : Statistical Distribution of Severity Mismatch for all features using Entropy
Method

Figure 7.8: Statistical Distribution of Severity Mismatch for specific features using
Entropy Method

135
 The p-values in the figures, all above 0.005, indicate that the differences in severity
calculations between CRITIC and Entropy methods for the sample data size of 100 in 12 runs
are acceptable. This means that the observed variations are within an expected range, and there
is no statistically significant difference between the two methods. The results suggest
consistency and reliability in the severity calculations obtained from CRITIC and Entropy
methods for this specific dataset and number of runs.

Table 7.6 Friedman’s Test for equal weight and CRITIC method by considering all and specific
features

Table 7.7 Friedman’s Test for equal weight and Entropy method by considering all and specific
features

The probability values displayed in the figures above, all exceeding 0.05, indicate that
there is no substantial impact of feature selection on the mismatch of severity calculated based
on all features and selected features in both the CRITIC and Entropy methods. In statistical
terms, a p-value greater than 0.05 suggests a lack of sufficient evidence to reject the null
hypothesis, signaling that feature selection does not exert a statistically significant influence
on the observed mismatch in severity calculations. This outcome implies that whether certain
features are included or excluded, utilizing either the CRITIC or Entropy methods, does not
significantly alter the results. This underscores the stability and reliability of the severity
calculation process in both scenarios.

136
 12

Number of Mismatch of Severity

C_af C_sf E_af E_sf
10

0
1 2 3 4 5 6 7 8 9 10 11 12
Run Number

Figure 7.9 Number of Mismatches of Severity

The evaluation of severity calculation mismatches based on CRITIC and Entropy

methods, considering both all features and specific features, offers valuable insights into the
efficacy of these approaches for predicting diabetes severity. This analysis investigates the
alignment of severity calculations when utilizing all available features compared to scenarios
where specific features are chosen. The primary objective is to comprehend the influence of
feature selection on the accuracy of severity predictions using CRITIC and Entropy methods.
The outcomes serve as valuable guidance for practitioners and researchers, aiding in the
understanding of the reliability and effectiveness of each method in diverse situations. This
information assists in making informed decisions regarding feature selection for enhanced
accuracy in diabetes severity assessments.

SUMMARY

Throughout the investigation, a comprehensive examination of methodologies for

predicting diabetes severity was undertaken, placing particular emphasis on the CRITIC and
Entropy methods. These techniques involve the normalization of feature values, computation
of entropy, estimation of diversification degree, and calculation of feature weights. Notably,
the CRITIC method introduces a systematic approach, incorporating standard deviation and
correlation coefficients for weight assignment in decision-making. The analysis of severity
calculation mismatches between CRITIC and Entropy methods, considering both all and
specific features, consistently produced p-values surpassing 0.05. This suggests that the
observed mismatches fall within an acceptable range, indicating an absence of statistically
significant differences in severity calculations between the two methods. Furthermore,
probability values exceeding 0.05 affirm that feature selection does not exert a notable impact
on severity mismatches in either method. In the broader context of predicting diabetes severity,
the utilization of a multiperformance index as a composite measure was explored. This index
amalgamates various indicators into a unified score, with its formulation adaptable based on
the research context and objectives. It provides a holistic representation of diabetes severity by
considering diverse features, with the selection and weighting influenced by clinical guidelines,
expert opinions, statistical analysis, and domain-specific knowledge.

137
In summary, this research work contributes original insights into the reliability of CRITIC and
Entropy methods for predicting diabetes severity. The comparison outcomes offer practical
guidance for practitioners and researchers regarding feature selection, and the exploration of a
multiperformance index underscores the significance of adopting a comprehensive approach
to assess diabetes severity.

138
 CHAPTER 8

CONCLUSION
This chapter wraps up the proposed work for diabetes detection and discusses the
algorithms that have been investigated for categorizing and predicting diabetes mellitus and
estimation of severity levels and the future direction of the research.

CONCLUSION
CDSS data classification aids physicians in discovering hidden data trends through the
use of big, real-world databases for ML-based classifier training. When it comes to health, DM
is a major threat on a global scale. It is necessary for the diagnosis, analysis, and management
of diabetes and its complications. Recent years have seen a dramatic increase in the use of ML
methods in the healthcare business, making this an exciting time for researchers interested in
applying these methods to the task of illness diagnosis. Understanding prognoses, providing
appropriate treatment, and avoiding complications all depend on early detection and
classification of diabetes. Due to the difficulties in distinguishing healthy cells from infected
ones, early diagnosis of DM presents a significant challenge in the field of disease management
systems. This work has made a substantial contribution in a number of ways because it takes
these into account and addresses them.
Medical care must be provided on an ongoing basis to the DM managing applications
in order to reduce the potential for serious, even fatal, complications. Despite extensive prior
work in DM classification using machine learning methods, accuracy remains a concern.
Furthermore, typical machine learning-based diabetes classification methods have significant
limitations in areas including privacy, missing values in input datasets, and incremental
categorization. The primary objective of this study is to create a robust DM classification model
that can accurately label patient data as either normal or diabetic and estimate the severity
level..
To provide an efficient diagnostics system, a hybrid optimizer based on SVC GB is
presented for the first diabetes categorization system. For maximum GB utilization, this system
incorporates an EFF and VPGWO within the DDS. Careful evaluation of the proposed
classifier's efficacy on a real-world dataset consisting of PIDD and real-time data is conducted.
A number of state-of-the-art GB- based methods are compared to the proposed classifier's
implementation in order to evaluate its performance. These metrics include classification
accuracy, precision, IoU, specificity, sensitivity, and AUC. Based on the actual evidence
provided, EFF-VPGWO emerges as the superior optimize r. To determine whether or not the
proposed classifier significantly improves assessment measures, the ANNOVA test is applied.
The majority of the p-values in this test are less than 0.05, indicating that the improvement
brought about by the EFF-VPGWO is indeed statistically significant.

To overcome classification issues, traditional machine learning-based DDS must first be

trained on predefined datasets in an offline (non-incremental) setting. In addition, they need to

139
collect the entire database once more to classify fresh samples because traditional supervised
learning is not always able to be done incrementally. Because of the ongoing nature of data
evaluation in clinical datasets, incremental updates to the trained models are necessary to cut
down on data classification processing time and storage needs. This paper proposes the model,
which streamlines computations by combining an SVC GB for classification with an EFF-
RVGWO approach to missing value imputation, data clustering, and feature extraction. On
PIDD, the HOMED's efficiency is measured in a variety of ways. The experimental results
demonstrate that compared to other offline methods, the computational complexity of SVC GB
is reduced while the predictive accuracy is increased.

Using healthcare information technology platforms in the form of EHR, sensitive user data
should be managed and shared to support care delivery and disease management. Patients are
hesitant to disclose personal information for reasons unrelated to treatment, and they desire to
be informed about data sharing. Therefore, it is essential to protect patient information if they
are to continue exchanging health-related information. Moreover, a disease diagnosis involves
numerous variables, making it difficult to arrive at a precise and timely diagnosis as well as
treatment plans. The EHR system requires a combined clinical decision support system and
ontology to construct knowledge bases that are efficient, dependable, re-usable, extensible, and
semantically intelligent.

This work aims to construct a safe diabetes classification model that protects the security of
sensitive user data by taking into account the above mentioned challenges. The proposed model
protects sensitive information using Kronecker product-based SVC GB, which optimizes the
parameter of Secrecy Usage (SU). Then, the domain ontology is developed to detect diabetes
by determining clinical similarities between the ontological rules with the trained dataset.
Hence, this work implements the SVC GB classification model by integrating privacy
protection, ontology, and GB-based classification algorithm. The PIDD findings illustrate that
the proposed SVCGB outperforms other similar techniques in terms of similar performance
measures.

This proposed Diabetic Mellitus Classification model can be utilized by healthcare

practitioners as a decision support system to provide secure and accurate online categorization
of diabetes mellitus with enhanced accuracy and other performance metrics. This system may
benefit many diabetic patients, particularly those who are oblivious of their disease type or how
to manage it.

140
 CHAPTER 9

FUTURE DIRECTION

Based on the conclusion and contribution made in this research, there is a list of
recommendations for those involved in the classification of brain malignancies. The
development of this endeavor in the future can be summarized as follows:
• The proposed DM classification method is easily adaptable to the categorization of
large datasets utilizing deep learning techniques. There will be several challenges in the areas
of data security, privacy, and governance brought on by the growth of medical big data. Finding
solutions to these problems is the first step toward a safe and effective data application in
healthcare. Important new understandings will emerge in areas such as drug management, the
delivery, intensification, and financing of basic healthcare services, all of which will contribute
to an overall quality improvement of treatment provided to patients.
• Text mining algorithms can be used to further categorize DM subtypes present in
medical records, including T1DM, T2DM, prediabetes, and GDM.
• Reducing the computational and time complexity is also important, since regular
applications for medical analytics shouldn't take more than a few minutes. So, it has to be
optimized so that it can be used quickly on high-performance computer platforms that run in
parallel mode.

141